A couple of weeks ago, I was speaking with Chris Granger, COO of Empire Today, about how Empire approaches customer service at their contact centers. He brought up an image of a duck: gliding above the surface while paddling furiously below. Chris said the customer at the other end of the line should never see all of the work being done behind the scenes. It should look easy.

That analogy stuck with me throughout the next few days. It’s exactly how we approach our product, too! We don’t need our partners to understand complex orchestration or the model decisions we make (which isn’t to say our engineers wouldn’t be happy to break it down for whoever is interested.) At risk of being cliche, we want to make AI feel simple. Easy.

A great agent is just like a duck: the easier the call feels, the more work is happening out of sight. So what’s going on down there?

The customer’s experience

Susan calls in to a customer service center to make some changes to her order.

An agent on the line greets her by name, and asks if she’s calling about last week’s gift for her mom. She says yes! While Susan hunts down the confirmation number, the agent waits rather than jumping in on the first pause. Halfway through a thought, Susan remembers she needs to change the delivery address. The agent follows the correction, confirms the new address, and reassures her that the gift will still arrive in time. Before hanging up, Susan adds a card to the order. The call takes a few minutes, and Susan hangs up feeling taken care of.

The work behind it

A lot has to happen behind the scenes to make that call go as smoothly as it did. Here’s the paddling:

1. A natural, seamless conversation. This is the most important part of any voice AI conversation, and takes a lot of parts working together. Tango, our industry-leading conversational awareness model, accurately identifies interruptions and end-of-turn. Our voices sound like actual people, and answer fast enough that the line never goes awkwardly quiet. Tango also reads sentiment, which is how our agents pick up on how customers are feeling.
2. Knowing your customer. Our agents remember every customer: their history, their preferences, and what they ordered last. This same memory carries across phone, text, email, and chat. So every conversation picks up where it left off, and your customer always feels like a regular.
3. Knowing your business. Customers often call in hunting for answers. Does this product come in another size? Is it still in stock? What’s the current discount on a two-pack? Connecting an agent to a constantly updating knowledge base, while ensuring it never hallucinates, is a complex task.
4. Taking action. More often than not, customers need your agent’s assistance: placing an order, changing an address, filling out an application. Our agents do so inside your systems accurately and consistently, so the change is real by the time the call ends. Smart AI isn’t a stopgap or an attempt to route callers away from your human agents. It’s a solution.
5. Following up. In many instances, customers reach out to your call center over different touch points throughout their journey. Our agents handle their whole relationship. For Renewal by Andersen, that means pipelines that include lead qualification, appointment booking, reminders, and follow-ups. For JG Wentworth, it’s walking a customer through an application process, even when it takes a few tries.

On working hard, and why it pays off

Like Chris, we at Simple are obsessed with quality. We hire the best of the Silicon Valley, and put in the hours to ensure that our product is the best on the market. It’s a core part of the Simple philosophy, cemented by my years at MIT, Meta, and Y Combinator: we’re not afraid to put in the work. Actually, we’re motivated by it!

Also like Chris, we hope to make what we do look easy. You shouldn’t have to think about orchestration or models, and your customers shouldn’t either (most of our work goes into making sure they don’t). It’s a call center that just works. We’ll keep paddling away.

Of all the numbers on a contact center dashboard, call abandonment rate is one of the quietest. When a customer gives up in the queue, there’s no complaint, angry survey, or escalation. On a support line, this means a problem left unsolved. On a sales line, it’s a customer who was ready to buy, and decided the wait wasn’t worth it.

Most contact centers track abandonment, but not everyone knows what a good number looks like, whether their definition matches anyone else’s, or how much revenue they’re losing as a result. This guide covers it all: what call abandonment rate measures, how to calculate it accurately, and what counts as normal. As a bonus, we’ll share how Omaha Steaks brought theirs from 16% down to 2%. In December alone, every percentage point for them signals $1 million of revenue lost.

What is call abandonment rate?

Call abandonment rate is the percentage of inbound calls where the caller hangs up before reaching someone who can help them. It’s one of the clearest signals of whether a contact center has enough capacity to meet demand. When the rate climbs, callers are telling you that the wait cost more than their problem was worth.

A call counts as abandoned whether the caller gives up in the queue, on hold, or partway through an IVR menu. How teams count each of those is the biggest discrepancy in how contact centers calculate abandonment. Two contact centers with the same performance often report very different numbers.

How to calculate call abandonment rate

The formula is simple:

Call abandonment rate = (abandoned calls ÷ total inbound calls) × 100

Say your line takes 2,000 calls on a Monday and 120 callers hang up before being answered. Your abandonment rate is 120 ÷ 2,000 × 100 = 6%.

As you calculate, take into account these three questions:

1. Are you excluding short abandons? Most teams ignore hang-ups in the first 5 to 10 seconds, since those are usually misdials. SQM Group estimates these false abandons can make up as much as 2% of call volume, which is enough to push a healthy line into the warning zone on paper.
2. Does IVR abandonment count? A caller who gives up three menus deep is often just as frustrated, if not more, as one who gives up in the queue. We recommend tracking IVR abandonments, even if you report them separately.
3. What’s your denominator? We recommend using calls offered, meaning every call that reached your system, as opposed to only the calls that made it to an agent queue. Otherwise, you’re hiding callers your routing lost before then.

What is a normal call abandonment rate?

Most industry benchmarks put the average between 5% and 8%. SQM Group puts the industry benchmark at 5% and finds that contact centers with the strongest customer satisfaction tend to run at 3% or less.

Here’s how to read your number:

Take into account that these benchmarks often also shift by industry. Healthcare lines, for example, average around 7% according to Sprinklr, driven by complex questions and longer calls.

One note on outbound: abandonment means something different when you’re the one dialing. For outbound campaigns, an abandoned call is one where the customer picks up and no agent is there. US telemarketing rules cap that at 3%, which makes it a compliance metric rather than a customer experience one. In this post, we’re focusing solely on inbound.

Why the average hides the real problem

A 6% monthly average sounds fine, but it could break down to 2% one day, and 20% another. Averages flatten peaks, but peaks are the signals that something’s gone wrong.

Two cuts tell you far more than the headline number:

1. By time: Break abandonment out by hour, day, and season. If the spikes line up with your busiest weeks, your problem probably isn’t service quality: it’s concurrency.
2. By intent: Abandonment on a support line costs you a callback, while abandonment on a sales line costs you the sale, and probably the customer.

What causes high call abandonment?

Most abandonment comes down to a handful of things:

1. Long wait times: This is often the single biggest driver, with every extra minute in the queue leading to more calls abandoned.
2. Understaffing during peaks: Forecasts miss, volume surges, and your team is in trouble.
3. IVR friction: Deep menus and “press 1 for” mazes lose callers before they ever reach your queue.
4. No visibility: Callers who don’t know how long they’ll be waiting tend to assume the worst.
5. No callback option: A callback gives callers a way out of a long hold while still getting help. Without one, hanging up is their only exit.

How to reduce call abandonment rate

The standard playbook is well known: forecast and staff more precisely, offer callbacks, announce wait times, streamline IVRs, and improve first call resolution. All of it helps, but none of it solves the underlying problem: if you’re relying on human agents only at your call center, your workflow is set by limited resources.

When you have more callers than agents, potential customers have to wait. Some of them, eventually, hang up.

This is one of the clearest benefits of voice AI. AI agents answer on the first ring, and are capable of handling a large share of calls end to end. The queue that used to form in front of your agents either gets shorter, or never forms at all.

How Omaha Steaks went from 16% to 2%

Two years ago, 16% of callers to Omaha Steaks hung up before reaching their agents. Last year, the number was 9%. Today? It’s 2%, measured across a line that handles 12 times its normal volume in December. That’s top-tier performance by any benchmark, on one of the hardest seasonal curves in retail. Read the full Omaha Steaks story.

Three things changed when they started working with Simple:

1. Every call gets answered immediately. Our AI agent picks up on the first ring, no matter the season. December’s surge no longer turns into a December queue.
2. Most callers never need the queue. With 72% of calls fully handled by Simone, the customers who do need a live agent join a much shorter line.
3. The callers who wait, happily wait longer. This last one surprised even us. Callers who talked to an AI agent before holding for a human stayed on the line about 30 seconds longer than they had prior. Being greeted, understood, and in the know made customers more patient overall.

What does that look like on a sales line? If yours takes 10,000 calls a week at peak, 16% abandonment means 1,600 callers hang up. At 2%, it’s 200. That’s 1,400 more customers per week, reaching someone ready to take their order.

Every abandoned call is a customer who wanted something

Maybe they wanted to place an order, or maybe they had a problem you could fix. Regardless, they gave up because nobody was there.

We think that’s a solvable problem. Nobody should have to wait in line to be helped, and with the right voice AI solution, nobody has to. With Simple, service starts on the very first ring.

Every customer conversation carries two kinds of information: what the customer says, and how they say it. People pick up on the latter without trying, but AI often doesn’t. We can hear the hesitation in a “sure,” the strain in a “that’s fine,” and the relief when a problem finally gets solved. We’re naturally attuned to it. In contrast, AI agents typically work off of transcripts, which don’t include any of these nuances to begin with. They can answer every question correctly and still feel completely tone-deaf if they don’t notice that a customer is confused, frustrated, or ready to move forward.

Our customers have been asking for a way to know how their callers feel, and we wanted our agents to know, too. So we built it into our most recent model launch.

Today, we’re introducing sentiment in Tango, Simple AI’s conversational awareness model. Tango tracks frustration, confusion, and happiness in every call, relying on audio-first processing for nuance and accuracy.

Why a Transcript Can’t Capture Sentiment

When we introduced Tango, we explained how most voice AI works. Models stitch together to transcribe what the customer says, generate a response from that text, then convert it back to speech. Tango moved the decisions that matter most, like end-of-turn and interruption, out of that pipeline by reading them straight from the audio. Sentiment is the clearest example of why that matters.

Much of how we read emotion lives in the parts of speech that don’t make it into a transcript to begin with, like pitch, pacing, volume, and emphasis. Text-based sentiment tools have to infer feeling from word choice alone, which works when the words themselves are emotional, like “this is ridiculous,” but breaks down when the words are neutral and the tone carries the meaning. This is common in customer conversations, where people tend to stay polite even when they aren’t happy. “Thank you” can be gracious, earnest, or deflated, and a transcript records them all the same way. Tango hears the difference.

The same “thank you,” in four tones

A transcript records all four the same way. Tango hears the difference.

Why Sentiment Is a Hard Problem

For people, reading emotion is second nature. For machines, it has been one of the harder problems in speech research.

The first reason for this is that there’s no agreed-upon way to divide up emotion. Some research datasets label frustration and confusion, some measure intensity, and some split happiness into separate categories: joy, delight, and satisfaction. Each dataset carves up the emotional space differently, so results from one model rarely translate to another.

The second is data. Labeling emotion from audio is slow and subjective, so many datasets rely on actors reading scripted lines, which keeps them small and doesn’t always sound like a real phone call. Even trained labelers don’t always agree on what they’re hearing.

The third is that the field is still young. Recent progress has come from feeding audio directly into language models and classifiers, which is the same family of approaches that Tango is built on.

How Tango Scores Sentiment

Three signals, each scored from 0 to 3

Rather than try to label every possible emotion, we focused Tango on three signals that shape a customer conversation and that an agent can actually act on. Confusion tells an agent to slow down and clarify. Frustration tells it to change its approach. Happiness tells it the conversation is going well.

Each signal is scored on a scale from 0 to 3. 0 means the feeling isn’t present, while 3 is the strongest expression of it: for frustration, think of a caller who has significantly raised their voice. The levels in between capture how strongly the feeling comes through. Each signal has its own dedicated output in the model, so the three scores move separately, and a caller can register as both confused and frustrated at the same time.

Scores that follow the conversation

Tango updates its sentiment scores continuously as the caller speaks, updating every 80ms. The scores describe the caller’s state over time, rather than in isolated outbursts.

Take a caller who was unhappy with their order. As they explain the problem, frustration rises to a 3, and stays there for as long they continue talking. When the agent finds the order and offers a resolution, you can hear the relief in the caller’s voice, and the score comes back down.

The first thank-you

A caller phones about a free pot roast that arrived in bad shape. Frustration registers at 3 from her first words.

She asks to file a complaint. The score holds at 3 the whole time she is speaking.

The agent finds the order and offers to file the claim.

You can hear the relief. Happiness registers for the first time as she thanks the agent.

Tango reacts to the conversation as it happens, and doesn’t try to decide whether a later rise is the same confusion or a new one. It reports how the caller sounds at every point in the call, which is exactly what an agent needs to decide what to do next, and what a reviewer needs to see to identify where a call turned.

One model, one pass over the audio

Sentiment isn’t a separate system added on top of Tango. The same model that detects end-of-turn and interruptions produces the sentiment scores, in the same pass over the same audio. There’s no second pipeline to run, no audio to reprocess, and no added latency for the agent. Because Tango already computes sentiment on every call, turning it on is simply a matter of deciding what to do with it.

In development, we found that training these skills together also made Tango better at each of them. Understanding when a caller is frustrated helps the model understand their pacing, and understanding pacing helps it read frustration.

And because Tango is built for real phone audio, sentiment works on the compressed 8kHz telephony that most customer calls still run on.

What This Looks (and Sounds) Like

The best way to understand sentiment is to hear it. In each example below, listen to the audio, then watch how Tango’s scores move.

Worry, then warmth

A caller is ordering a birthday gift. She starts out upbeat.

She thinks she can’t pay from her landline, and frustration builds to 2.

The agent reassures her and walks her through how to pay.

Her mood lifts. Happiness jumps to 3 and frustration drops to 0.

Happiness is as useful a signal as frustration. Here, Tango hears a caller’s mood lift, and reacts accordingly.

Reads fine on paper

A caller is checking that the wheel bearings he received are the right parts for his car.

One part number doesn’t match. Confusion registers at 1 and holds while he reads numbers back.

Confusion climbs to 2 while he keeps saying “okay.” The agent is patient and confirms both parts.

With the confirmation, confusion eases back to 1.

Confusion rises as a customer tries to identify that the parts he received were the right ones for his car. The agent is patient, communicates what he’s looking for, and provides the confirmation.

Frustrated from the first sentence

A second call to place the same order.

Frustration reaches 2 within two seconds of the caller speaking.

The agent slows down to acknowledge it and apologizes. Frustration holds at 2.

A preview of what’s coming: Tango hears frustration building, and the agent slows down to acknowledge it before moving on.

What Sentiment Makes Possible

A sentiment score on its own is information. The value comes from what agents and teams do with it, and that falls into two broad categories: changing how an agent behaves during a conversation, and understanding customers better across many conversations.

During the call: agents that adapt

A good human agent adjusts constantly based on how the customer sounds. Sentiment gives an AI agent the same input, so its next move can depend on how the customer feels, not just what they said.

• Clarifying when confusion rises: Instead of moving on to the next step, the agent can slow down, rephrase, or check that the customer is following.
• Changing course when frustration builds: The agent can acknowledge the frustration, offer a different path, or bring in someone from your team when a person is the right answer.
• Matching the customer’s tone: A brisk, cheerful caller and a hesitant one need different pacing. Sentiment lets the agent tell them apart.
• Recognizing the right moment: When a customer is happy with how the conversation is going, that’s often the natural time to suggest what comes next, whether that’s an upgrade, an add-on, or scheduling the next appointment.

After the call: insight across every conversation

With Tango, every call becomes a record of how the customer felt from start to finish. That means anyone reviewing it can see exactly where the conversation turned, for better or worse, without listening to the whole thing.

Across many calls, the patterns get more useful. If frustration rises every time a particular policy, product, or step in the process comes up, that’s a signal about the customer experience itself, showing the business exactly where to improve.

We use the same signal to improve our own agents, finding the moments where callers get stuck and fixing them together with each customer.

What This Means for Your Business

Sentiment connects directly to the numbers your team is already measured on.

• Conversion: A customer who’s confused about price, timing, or next steps is less likely to commit. An agent that hears confusion can clear it up while there’s still a sale to make.
• Customer satisfaction: Surveys reach a small share of customers, and only after the call is over. Sentiment gives you a read on every call, as it happens.
• Loyalty and lifetime value: Relationships are built over many conversations. Knowing how each one actually felt to the customer, not just whether the issue was resolved, shows you which relationships are healthy and which need attention.

What’s Next for Sentiment

Think about the best sales associate you’ve ever had. They knew your name, remembered what you bought last time, and could tell from your voice whether today was a good day. That kind of attention used to be reserved for luxury retail, because it was too costly to give every customer. Sentiment brings us one step closer to offering it on every call.

Want to hear it for yourself? Schedule a demo with our team.

Earlier this month, our CEO Catheryn Li took to the CCW Nashville stage in a panel moderated by Jordan Zivoder (CMP), alongside Rob Bradshaw (Omaha Steaks) and Phil Engle (Renewal by Andersen). The topic? How the two legacy brands used Simple to cut abandonment, improve containment, and book $400K in new revenue in a single week. For the full conversation, watch the video below.

The problem before AI

Omaha Steaks: the December crunch

Omaha Steaks, which has sold premium steaks and seafood for over a century, does its biggest business in December. Every year, Omaha’s call center traffic increases 12x, far more than its 200 agents can manage. Omaha historically hired 5,000 seasonal employees every October, and with experienced agents moved to other seasonal jobs, the newest hires took the most important calls. Abandonment hit 16% at peak, with each percentage point costing about $1M in revenue.

Renewal by Andersen: overflow and unused leads

For Renewal by Andersen, the problem lasted all year. The company, one of the largest home improvement brands in the US, was missing out on opportunities due to a lack of resources. RbA spent up to $10K/month on a third-party overnight call center that set poor-quality leads, misrouted calls and ignored their rules. Their highly skilled reps spent most of their time fielding inbound calls, rather than pursuing high-value outbound work. Lists of potential leads went unused, because no one had the bandwidth to call them.

Finding the right vendor

Omaha Steaks first tried building its own IVA. In Five9 Studio, they built one narrow use case and reached 20% containment after 6 months. They then evaluated five vendors, and Simple won by replacing that use case in 8 days and reaching 61% containment within 3 months. As Rob said, “I took off my Superman, build-your-own-IVA cape. I threw it in the trash and I’ve never worn it again since.”

Renewal by Andersen ran a head-to-head POC for overnight rollovers. In a 4-month trial, the competitor never got as far as Simple did in month 1. But that wasn’t the only reason RbA chose Simple. Simple took DNC compliance seriously where the other vendor didn’t, and, most important to Phil, shared ownership of the results. Phil said, “My problems were [Cat’s] problems equally. And I really appreciated that.”

The results speak for themselves

At Omaha Steaks, Simple started with one sales use case, expanded to all sales, then added support, chat, and QA. Next they’re adding SMS and email, and pushing containment even higher.

At Renewal by Andersen, Simple took over overnight calls and started following up on missed leads. Increased bandwidth allows reps to convert leads and handle objections, rather than just answering inbound.

What makes voice AI work in production

Cat’s answer to this question came down to two things: the technology itself, and customer care. The panel kept coming back to a third: personalization.

On technical infrastructure

Off-the-shelf models are either too slow or not accurate enough for a live call, so we fine-tune and host our own. A harder problem is knowing when the caller is done talking: cut in too early or wait too long, and the conversational quality is lost. That’s what Tango, our newest model, solves. Tango scores 96% on LiveKit’s public end-of-turn benchmark, making it the best in the industry. Rob said, “I didn’t even have to be told when it was put in place. I knew because I saw a great big uptick in our containment.”

On customer care

Even the best technology needs a team that cares about getting things right. As Cat put it, you can’t drop in a prompt and expect it to work. Our customers know their businesses better than anyone, and we rely on that expertise. Every agent is built through constant back-and-forth: customers test it, tell us what’s working, and we iterate from there.

At Omaha Steaks, that meant our developers working on-site for the launch, and coming back every quarter since. At RbA, Phil noted that he never had to remind the Simple team where they’d left off. As he said, “It’s not just looking at what the bot does. It’s who’s making the bot.”

On personalization

If a customer places an order and calls back a week later, imagine their delight in hearing, “I see you ordered a package last week, are you calling about that?” Simple’s agents have persistent memory, which means they remember customers’ preferences, addresses, order history, and life events. According to CMP’s research, customers that are satisfied with their level of personalization are four times more likely to be satisfied with their experience. Whether it be over voice, chat, email, or SMS, that memory carries through.

Three lessons for contact centers just starting with AI

1. Don’t be afraid to start with revenue. AI can outsell seasonal reps. For Omaha, “The AI agent actually beat the humans at the upsells and the add-on items. It was not even really close.”
2. Be creative. Test voice, accent, phrasing and offers faster than you ever could with human reps.
3. Change your mindset. When AI learns from your best reps, the new job is making your best even better.

The takeaway

Omaha Steaks and Renewal by Andersen came to Simple with very different problems. One needed to survive a 12x holiday surge, while the other needed to stop leaving leads on the table. Both ended up in the same place: with a contact center that drives revenue.

Each started with one narrow use case, proved it worked, and built from there. Today, Omaha’s AI agent beats its human reps on upsells, and RbA turned a forgotten lead list into $400K in a single week.

If your contact center is still seen as a cost center, start with one solvable problem. Book a demo to talk to our team about yours.

Most pieces of a voice agent can be bought off the shelf today. Speech-to-text, large language models and text-to-speech are each a few API calls away and can be wired to a phone line without much trouble. So when a consumer brand starts thinking about putting AI on its sales line, building in-house can look like the obvious move. Why pay a partner for something your own engineers could put together?

At Simple AI, we build voice AI agents that sell in the contact centers of consumer brands like Omaha Steaks, Speedway Motors, and Renewal by Andersen. Getting an AI to speak out loud might feel simple. Getting it to reply as fast as a person would, on a noisy phone line, without talking over your customer, is one of the hardest engineering problems in voice AI. We’ve hired the best engineers to solve it, and that’s why our voices are considered some of the best in the industry. But voice quality is only where the work starts. The real revenue impact of voice AI comes from what the agent does once the conversation is going: recognizing a returning customer and what they bought last time, changing an order inside your systems while they wait, and texting them a week later when they’re ready to buy. That’s where the gap between building and buying gets even wider.

Should you build or buy voice AI?

In the simplest terms: if you need voice AI to drive revenue, buy it. Selling means looking up orders, writing to your CRM, taking payments and following up across phone, text and email until the sale closes. Each of those steps has the agent working in the systems your business runs on, with a customer counting on it to get things right.

Even a use case that sounds simple, like answering common questions or routing callers to the right team, takes more than it seems. The agent still has to sound human and get every answer right, and your team still needs QA on every call, structured data pulled from every conversation, and reporting it can act on. Each of those is hard to build in-house. Many of our customers start with a single use case like this and expand as their teams see what else Simple can do.

For Epiq, Simple handles calls about legal settlements and class actions: checking whether someone is eligible, verifying their details, and helping when they’ve forgotten a PIN. On paper, it’s an FAQ line. Epiq still ran a bake-off against multiple vendors and chose Simple for the best voice quality and response times, and the most accurate answers. Since going live, their team has used Simple to test every change in simulation, get a daily report on every escalated call, and keep an audit log of every change to their agents. None of that comes with a voice model and a phone line.

In practice, build vs. buy covers three options. The leanest connects a voice model straight to your phone system. Next are self-serve platforms, where your team configures and runs a capable agent on its own. The most complete is a partner that builds the agent with you: integrating with your contact center platform (CCaaS) and CRM, giving your team tools to run it, and staying hands-on through deployment. We built Simple to be the third option, because the closer an agent gets to a sale, the more integration, testing and upkeep it needs behind it.

What does it take to build voice AI that drives revenue?

Building voice AI that drives revenue takes far more than a speech model and a phone line. Your team would need to build real-time conversation timing, integrations with your contact center and CRM, customer memory, security redaction, simulation testing, QA and reporting, and then maintain all of it as models change. If you decide to build, you might think you just need to find a text-to-speech model and link it to your phone line. The real to-do list looks a little more like this:

• Choose speech-to-text, language and text-to-speech models, and benchmark each one on real phone audio
• Chain them together fast enough that callers never notice a gap
• Teach the agent to tell a finished thought from a pause, and a “mm-hmm” from an interruption
• Tune every model for speakerphones, background noise and bad connections
• Connect to your CCaaS, CRM, order management and payment systems
• Give the agent a memory of every customer interaction across phone, text and email
• Set rules for when the agent follows up, and on which channel
• Mask and redact card numbers and personal data on every call, then prove it to a security review
• Build a tool so people outside engineering can update offers, scripts and policies
• Build simulations that test every change before it reaches a customer
• Build QA that tags calls, tracks outcomes and compares AI agents against human ones
• Pull structured data out of every conversation and send it where your team needs it
• Monitor live calls and catch failures before your customers do
• Re-test everything each time a model provider ships a new version

And that’s just to launch! Every month your team spends working through that list is a month of sales your agent isn’t closing. None of it is something your engineers can put together on the side.

What your voice AI needs to drive sales

A voice AI agent that sells needs six things: conversation that moves at human speed, a memory of every customer, tools your team can use without engineers, built-in measurement, security for card and personal data, and a team working on it full time. Each is covered below.

1. A conversation that moves at human speed

In a study of conversations across 10 languages, researchers found that people typically leave a gap of around 200 milliseconds between turns. To come close, a voice agent has to run three models back to back: one to transcribe what the customer said, one to decide how to respond, and one to say it out loud. That reply then has to travel across a phone network. Every handoff adds a delay that then gets passed on to your customer. In development, we found that the smartest models are often the slowest, while the fastest are the hardest to keep accurate. An agent that’s both quick and right takes a pipeline tuned as a whole, not the best model for each step.

On a sales call, timing does a lot of quiet work. Customers interrupt, pause mid-thought to check a measurement, and nod along to show they’re still listening. An AI agent has to make those same decisions continuously, working out whether the customer has finished speaking or is just catching their breath. Get it wrong in one direction and the agent talks over someone who was about to say yes. On the other? The line goes quiet long enough to make the whole experience feel stilted and robotic.

We’ve built a pipeline around the calls customers actually make, whether on speakerphone, in the car, or with a TV going in the background, and we trained our own end-of-turn model, Tango, to tell a finished thought from a pause. A customer halfway through an order shouldn’t have to repeat themselves.

2. A memory of every customer

The best sales associates remember their customers. They know your size, which piece you’ve had your eye on, and call the day it comes in. That level of attention has mostly been found in luxury stores, where every customer gets a person’s time. With AI, larger companies can now give it to them.

To do that, an agent has to remember each customer’s history, contact them at the right moment on the channel they prefer, and keep the conversation going across phone, text and email for as long as the journey takes. That means connecting it to every system that holds your customer data, and deciding when it should reach out, not only how it should respond.

For JG Wentworth, Simple guides applicants through a multi-step application by phone and text, and keeps following up for days until it’s complete. For Empire Today, Simple stays with each customer from before the purchase to long after the floors go in: answering questions up front, confirming the consultation and the installation, and following up months later about the next room. In both cases, the sale takes more than one conversation, and the agent remembers each one.

3. Tools for the people who run the agent

Say your team decides to run a promotion with your voice AI agent. It needs to know the offer, mention it to the right customers, and answer questions about it correctly from the first call. If each of those changes goes through an engineering ticket, the agent ends up a step behind the business it’s supposed to be growing.

Building your own agent means building a second product as well: a way for people outside engineering to update the agent safely, and simulations that test every change before it reaches a customer. Simple comes with both, so the people who know your customers best can make those changes themselves.

4. Measurement that fits how you already work

Putting AI on the phones doesn’t change what you need to know. What share of calls end in a purchase? Where do customers drop off? Is the new script converting better than the old one? You need those answers for your AI agents the same way you need them for your people.

Getting them takes a measurement layer on top of the agent, and in Simple, that layer comes standard. Teams can tag calls, track outcomes, and compare performance across AI agents, AI versions and human agents in one place. Some of our customers now route their human agents’ calls into Simple just to use our analysis tools, which tells us the measurement is worth having on its own.

5. Security you can stand behind

Sales calls carry card numbers, account details and personal information. Masking and redacting that data on every call, without exception, and proving it to a security review takes specialized work, and the requirements keep changing. We handle it across every deployment, so your team doesn’t have to take it on.

6. A team that works on it full time

None of the work above is ever truly finished. Customers find new ways to ask for things, your business keeps changing what it needs the agent to do, and the models underneath change constantly. Model providers ship new versions every week, and each one has to be benchmarked on everything your agent does, from understanding a caller to analyzing a call to striking the right tone, before it’s safe to put in front of a customer. Inside a consumer brand, all of that competes for engineering time with the website, app and everything else on the roadmap.

Our team works on voice AI agents 100% of the time. That’s far more bandwidth than most companies could set aside for a single internal project, and each improvement we make reaches all of our customers. We test every new release against our own benchmarks and run whichever performs best, whether it comes from a model provider or from the self-hosted and self-trained models we run ourselves, like Tango. Your agent gets the best performance money can buy, with no overhead for your engineers.

What do you get when you buy?

When you buy from a partner that builds with you, you get to revenue sooner, start with a pipeline already proven on live sales calls, work with voice AI experts, and pay with pricing tied to outcomes. Here is what that means in practice.

1. Revenue sooner. Our dedicated deployment team can stand up a new use case in days.
2. A head start. The pipeline behind every Simple agent reflects everything we’ve learned on live sales calls, and we upgrade it every time a better model ships. You start from there, instead of from zero.
3. Experts who build it with you. Simple isn’t a do-it-yourself platform. We invest heavily in discovery, integrate with the CCaaS and CRM systems you already use, and configure every agent around your workflows and the numbers your team is measured on. You bring the knowledge of your business and we bring the AI expertise. No one has to learn the other’s job.
4. Incentives that match yours. Our pricing is tied to outcomes, so we only do well when your agent does.

For Omaha Steaks, that meant far fewer customers hanging up before anyone could help them. After they deployed Simple, their call abandonment rate dropped from 16% to 2%.

When does building your own voice AI make sense?

Building makes sense when voice AI is itself your product, or when you have a dedicated voice engineering team and a use case that no partner can support. In those cases, owning the stack can be worth the cost, but for a brand that wants voice AI to grow sales, it rarely is.

Making the call

Any AI investment has to answer the same three questions: will it grow revenue, how soon will it do so, and how much will it ask of your team? Building your own voice AI puts all three at risk. You’d be betting revenue on a product your team has never built, on a timeline no one can promise, with engineers pulled away from everything else on the roadmap. A partner whose whole team works on this can answer all three now.

If revenue, performance and customer experience matter to your business, the choice is clear. Buy.

We built Simple so consumer brands can give every customer the kind of attention that used to belong to luxury stores, without having to become AI companies to do it. We take on the hard parts, from timing a reply to following up months later, so your team can spend its time on your customers.

Schedule a demo with our team to see what that looks like for your business.

Voice AI is software that has real, spoken conversations with customers over the phone: understanding requests, responding naturally, and taking action without a script tree. Today’s AI agents are pretty extraordinary, able to handle complex flows like scheduling, sales, and customer support both accurately and reliably. On the backend, platforms combine speech recognition, language understanding, and natural-sounding speech to hold conversations that feel inexplicably human.

Why this guide exists

Voice AI often sounds a lot more complicated than it is, and can seem intimidating when you start looking for a solution for your call center. The engineering is certainly complex, but the ideas themselves aren’t hard to explain, and CX leaders shouldn’t need a computer science degree to make the right decision. This guide breaks down how voice AI actually works, and defines the terms you’ll run into while evaluating it. Plain English only.

How voice AI actually works, in one pass

A customer calls into your call center, and a voice agent picks up. A few things happen next, usually in under a second:

1. Speech-to-text (transcription): the caller’s spoken words get converted into text the system can parse.

2. Understanding intent: the system identifies what the caller is calling for.

3. Generating a response: the system decides what to say or do next, whether that’s answering a question, pulling up an order, or routing to a live agent.

4. Text-to-speech: the response gets converted back into natural-sounding audio.

There’s a subtler piece stitched into this same pipeline: knowing when a caller is actually done talking. That’s called end-of-turn detection, and you’ll find more about it below.

The core vocabulary

What is containment rate?

Containment rate is the percentage of calls a voice AI resolves on its own, without transferring to a human agent. It’s the industry’s most common success metric, but it’s also inconsistently defined between vendors. Does a call that gets a wait-time message before resolving still count, and does chat volume get folded in with calls? Always confirm the exact definitions when comparing containment.

What counts as call abandonment?

Call abandonment rate is the percentage of callers who hang up before reaching a resolution, typically while waiting on hold. It’s one of the clearest signals of a broken phone experience, since every abandoned call is a customer who gave up on being helped.

What is opt-out rate?

Opt-out rate is the percentage of callers who ask for a human agent, even when the AI could have continued to help. It’s a useful companion to containment rate rather than a rival to it: containment tells you how many calls the AI handled end-to-end, while opt-out tells you how many people didn’t want it to after it had already engaged them.

What is an IVR, and why is it being replaced?

An IVR (interactive voice response) is the ‘press 1 for sales, press 2 for support’ phone-tree system most contact centers still run today. IVRs route calls based on rigid menu choices rather than understanding what a caller is actually asking for, which is why they’re increasingly being replaced by systems that let callers simply state what they need.

What is CCaaS?

CCaaS (Contact Center as a Service) is the cloud platform your team likely already uses to route calls and run day-to-day contact center operations, such as Five9, NiCE, or Genesys. Voice AI doesn’t need to replace your CCaaS to work. Most platforms plug in as an added layer that handles a slice of your call volume, while your existing system keeps doing everything else.

What’s the difference between an IVA and voice AI?

IVA (intelligent virtual agent) is an older term for the same category that voice AI now falls under, and the two are often used interchangeably. If a vendor or an RFP brings up IVA, they’re almost always describing the same kind of system this guide covers.

What makes an AI agent ‘agentic’ vs. scripted?

A scripted bot follows a fixed decision tree, while an agentic AI system can reason through a conversation, pull in outside information, and take multi-step actions. The practical difference between the two shows up when a caller goes off-script: a flowchart bot breaks, while an agentic system can still follow the conversation.

What is NLU?

NLU (natural language understanding) is the technology that lets a system figure out what a caller actually means, beyond the words they used. It’s what helps your agent realize that ‘where’s my order’ and ‘I never got my package’ have the same intent, even though the sentences barely overlap.

What’s a warm transfer?

A warm transfer hands a call from AI to a human agent along with full context, so the caller never has to repeat themselves. A cold transfer, with no context passed along, is one of the most common reasons callers get frustrated with automated systems. Warm transfers help customers feel taken care of, and make the experience seamless even when an AI agent can’t handle the entire flow themselves.

What is end-of-turn detection?

End-of-turn detection is how a voice AI system knows a caller is actually finished talking, not just pausing to think. If your agent waits too long to speak, the conversation feels stilted and robotic; if it jumps in too early, it interrupts the caller mid-sentence. This is one of the areas where voice AI performance differs, and is a metric systems are frequently graded on.

What is barge-in?

Barge-in is when a caller interrupts AI mid-sentence, requiring it to stop speaking and listen instead of finishing its own line. Handling barge-in well is one of the clearest tells that you’re talking to a system built for real conversation, as opposed to a recording with a menu attached.

What is voice AI latency, and why does it matter?

Latency is the delay between when a caller finishes speaking and when the AI responds, measured in milliseconds. Anything much above 800ms starts to feel unnatural to a caller, even if they can’t articulate why. It’s one of the most underestimated factors in whether a voice AI system feels like talking to a person or talking to a machine.

What is RAG?

RAG (retrieval-augmented generation) is a technique where an AI system looks up relevant information from a knowledge base in real time before generating its response, rather than relying only on what it was trained on. This is what lets a voice AI agent answer accurately when asked about an order status, current inventory, or company policies.

What is SOC 2, and why does it matter for voice AI vendors?

SOC 2 is an independent audit that verifies a company has real controls in place for data security, availability, and confidentiality. For a voice AI vendor handling customer calls, often including payment or account details, a SOC 2 report is one of the first things a security or compliance reviewer will ask for.

What is WISMO, and why do contact centers care?

WISMO stands for ‘where is my order,’ and is the single most common call type for any business that ships physical goods. It’s cited constantly in contact-center benchmarks because it’s high-volume, low-complexity, and one of the easiest call types to automate well.

What is knowledge-based authentication?

Knowledge-based authentication verifies a caller’s identity by confirming details only they should know, such as an order number, billing ZIP code, or date of birth, before an agent shares account information. This is often established for both human and AI agents, and is a standard security layer for any call that touches personal or account data.

Three ways companies build voice AI

Not every voice AI system is built the same way. Here are the three most common approaches.

1. Flowchart-style automation

Flowchart-style automation layers voice on top of the same logic as a traditional IVR: a caller’s words trigger the next branch in a decision tree, rather than being genuinely understood. It’s the fastest and cheapest option to stand up, but breaks the moment a caller says something the script didn’t anticipate.

2. Do-it-yourself infrastructure

Do-it-yourself infrastructure means assembling your own system from individual pieces: a speech recognition provider, a language model, a text-to-speech engine, and the plumbing to connect them. This gives you full control over every part of the stack, but also means your team owns integration work, latency tuning, and any ongoing maintenance that comes with running in production.

3. Full conversational platforms

Full conversational platforms handle the entire pipeline—recognition, understanding, response, and voice—as one connected system, along with the orchestration and analytics layered on top. The tradeoff is less granular control over any single piece, in exchange for not having to build or maintain the pipeline yourself. Some companies, like Simple, offer bespoke deployments that make sure the system you’re implementing answers all of your business needs.

None of these options are universally right. The best fit usually comes down to how much engineering time your team wants to spend on the voice AI itself, and the complexity of the problem you’re trying to solve.

Why teams start looking at voice AI in the first place

Voice AI isn’t right for every contact center. But a few situations come up again and again for teams that truly benefit from it.

Call volume spikes hard during specific seasons or hours. If your busiest weeks mean hiring and training a wave of temporary agents, only to scale back down a few months later, you know how expensive, time-consuming, and ineffective the process is. Voice AI can operate at scale for contact centers handling seasonal volume.

Callers hang up before they get help. A high call abandonment rate, especially during peak hours, usually means people are giving up after waiting too long for a live agent. It signals more than just a bad experience: it’s a missed sale or an unresolved issue. AI agents help reduce hold times, and ensure your abandonment rate reduces alongside it.

A huge share of calls are the same simple question, asked constantly. ‘Where’s my order’ is the classic example: high-volume, low-complexity, and one of the easiest call types to automate well. Having voice AI handle these requests frees up live agents for callers who actually need them.

Nobody can answer the phone after hours. Missed calls outside business hours are lost opportunities that competitors with round-the-clock coverage don’t have to give up. If a potential customer calls after hours, voice AI can still pick up. It can book the appointment, close the sale, or answer the question instead of sending them to voicemail.

You don’t fully trust your own numbers. If getting a straight answer on your containment rate means reconciling two dashboards or a spreadsheet someone still updates by hand, that’s usually a sign the underlying reporting needs attention. This is often the moment teams start looking at their contact center stack more broadly.

If a few of these sound familiar, it’s worth a closer look. These are the exact problems voice AI is built to solve.

Wrapping up

Voice AI sounds complicated because of the vocabulary, not the ideas. Containment, latency, agentic, WISMO: once you know what these terms actually mean, vendor conversations get a lot easier to follow.

We created this guide to make sure you’re never nodding along to a term you don’t actually understand. Bookmark it, come back to it, and, if this sounds like the right solution for you, reach out to the Simple team for a demo.

Voice AI’s biggest challenge is timing. For a conversation to feel natural, an agent has to recognize the moment a customer is done speaking and craft their response quickly and accurately. Most models still struggle with this, either interrupting mid-sentence, or pausing too long to confirm their turn to speak. Customers are left with conversations that feel robotic rather than human. It’s the biggest obstacle standing between voice AI and mainstream adoption.

We tried all of the end-of-turn models on the market and couldn’t find one that actually worked well in production. In Simple AI fashion, we chose to train our own.

Today, we’re introducing Tango, the system behind Simple’s most natural voice agents yet. In production, its end-of-turn decision lands a median 300ms before a conventional voice-activity detector would even detect a pause — measured on real telephony audio, not just clean VoIP.

How Voice AI Used to Work

Most voice AI on the market uses a cascaded approach: transcribe what the customer says, generate a response from that transcript, convert it back to audio. Everything the agent needs has to survive being flattened into text first, and a lot gets lost in translation. The result is both slower and less natural: no matter how fast your agent is, it’s waiting on that transcript before it can say anything.

Tango skips that step for the decisions that matter most, reading end-of-turn and interruption directly from audio. Transcripts still get generated for every call, but they no longer sit in the critical path.

Benefits of Tango’s Audio Processing

Improved end-of-turn accuracy

Tango isn’t just faster than competitor end-of-turn models, it’s also more accurate. This is crucial, since false positives mean more interruptions and false negatives mean uncomfortable pauses. With improved end-of-turn, conversations flow more freely. On the public LiveKit benchmark, it’s the fastest of 11 systems once you require at least 90% detection, missing just 15 of 400 true turn endings versus 54 for Soniox and 197 for Deepgram Flux. Against Smart Turn specifically, it fires 3.7 to 4.7 times less often mid-speech at equal turn-end coverage.

Fewer accidental interruptions

Genuine interruptions and quick backchannel cues like “mm-hmm” can look identical in a transcript, but they call for opposite behavior: stop talking for one, keep going for the other. Tango tells them apart in real time, so it doesn’t cut off a customer who’s simply agreeing. That’s cut interruptions by 64% versus our previous agent in production, recognizing a real interruption about 100ms after the customer starts talking.

Hear what text can’t show

Type “thank you” and there’s no way to tell if it was said with a smile or through gritted teeth. Text strips out tone, prosody, and pacing: the exact signals that make an interaction feel human. Audio-native processing keeps that signal intact.

Speed, without the translation loss

The old pipeline adds roughly one second of latency before your agent starts thinking. Tango skips it: audio comes in, decisions come out, with both agent and customer channels processed at once. By going straight to streaming audio models, we can make assessments live every 80 ms, and run inference in less than half that time. There’s no transcript to wait for, and no second pipeline to run.

What This Looks (and Sounds) Like

A transcript-based model sees silence and guesses. Tango hears the customer still thinking, and responds at just the right time.

Tango can tell a real interruption from a conversational nod, so your agents don’t stop talking every time a customer says “okay”.

What This Means For Your Call Center

In practice, the difference between a natural-feeling agent and a robotic one comes down to about one second of latency. This gap shows up directly in the metrics you’re already tracking.

• Customer Satisfaction Score (CSAT): A customer’s satisfaction is shaped by their overall experience rather than resolution alone. Fewer awkward pauses and false interruptions means fewer calls that feel broken, even when the agent is correct.
• Average Handle Time (AHT): Every pause, interruption, or repeated “sorry, go ahead” adds seconds to your call. At call center volume, these seconds compound fast.
• Abandonment: The customers you lose aren’t the ones who got a wrong answer. They’re the ones who hung up in frustration before you got the chance to help them at all.

How We Compare

A system near the top of this chart catches more end-of-turns, while a system to the left catches them sooner. Tango is the most accurate model for its speed, and the fastest model for its accuracy. Because the best, most natural conversations have both.

Built for the Calls You Actually Get

Most end-of-turn models are trained and benchmarked on clean, high-bandwidth audio. Real customer calls are not. They’re compressed, 8kHz telephony audio, running through phone systems built way before there were AI agents on the other end of the line. That’s why models that perform well in a demo often degrade the moment they meet a call center’s actual call volume.

Tango is built dual-channel and telephony-native from the ground up. It plugs directly into the inbound and outbound calling systems you already run, without asking you to upgrade your infrastructure first. That’s the difference between a benchmark number and a number you’ll actually see in production.

What’s Next for Simple Tango

Tango is already live across Simple’s production traffic, delivering lower latency, higher accuracy, and more natural conversations for the calls our customers handle every day. It’s already shaping how we think about every agent we build from here.

They say it takes two to Tango. Schedule a demo today to see how Tango could sound for you.

Just a few weeks ago, Simple founder and CEO Catheryn Li spoke at Genesys Xperience about what actually separates voice AI agents that work from ones that just sound good in a demo. For those of you who missed it, here are a few of our key takeaways.

Voice AI comes down to two things: infrastructure and behavior

Cat split the problem of building voice AI into two broad categories. The first was infrastructure, the engineering that happens behind the scenes and that most people never think about directly. This is what determines whether a conversation feels natural. Does the voice sound human? Does it know when it is actually its turn to speak? How quickly does it respond, and can it handle being interrupted without falling apart?

In our system, every single turn of conversation runs multiple models in parallel: speech to text, voice activity detection, end-of-turn detection, interruption detection, and noise cancellation. That output feeds into an inference layer that decides what knowledge to access, what guardrails to apply, and what the agent should say next, after which a text-to-speech model turns the agent’s response back into audio.

If you were to chain together off-the-shelf APIs for this task, you would likely end up with response times of one to three seconds per turn—and a stilted, robotic conversation experience. At Simple, we train and self-host our own models across the stack, and we are proud to have reduced turn latency to an industry-leading 550 milliseconds.

The behavior side is where businesses have real control

The second half of the equation is behavior, or what your agent actually says and does. Cat framed this in three tiers:

1. Does the agent route and triage correctly?
2. Does it answer accurately?
3. Does it take the right action when it matters?

Answering accurately depends on the agent knowing how and when to pull from the right knowledge base, especially when a company has several disconnected ones. Taking the right action is equally crucial, since placing an order incorrectly creates significantly more work than doing nothing at all.

Good documentation matters, but so do your call recordings

The better and more complete a company’s SOPs and knowledge base, the better the results. This is not surprising on its own, but what is less obvious is that even strong documentation only shows a fraction of the picture inside a contact center. We ask clients to let us train on past call recordings and transcripts so we can learn from how their best agents actually handle situations the SOPs do not cover. Those recordings also become the basis for evals, where we build customer personas to continuously test that the agent handles real scenarios correctly as it evolves.

A different way to think about improving customer experience

One of the more interesting points from the talk was a mindset shift. Traditionally, contact center leaders spend a lot of energy coaching their weaker reps to close the gap with their best ones. With AI, the question becomes different: how do you get the system to perform as well as your best rep? Once that is done, how do you make your best rep even better? AI also makes A/B testing in call centers quick and effortless. Instead of taking months to roll out a change across a human team, you can test a new message, tone, or even voice and see results tomorrow.

The results

Across deployments, Simple AI sees an average resolution rate of 70 percent, along with higher conversion and upsells compared with live reps. We use the word resolution intentionally, rather than containment or deflection. At Simple, the goal is not to get customers off the phone. It is to solve their problems.

Last week, we spent three days at Genesys Xperience in Las Vegas, talking to contact center leaders from around the world. A few themes kept coming up in booth conversations, one-on-one sessions, or during our talk. Here’s what stuck with us:

1. The best AI rollouts start small and expand deliberately
2. Integration experience is just as important as the AI itself
3. Voice quality is the most important part of your customer’s experience with AI

1. Starting small: crawl, walk, run

We’ve worked with contact centers of all sizes, across different industries. The teams that have the most success with implementing AI aren’t the ones trying to automate everything on day one. Rather, they start with simple wins, get familiar, and expand from there.

Crawl looks like intent based routing, FAQ handling, and basic account questions like checking a balance. It’s low risk, high reward, and pretty quick to implement.

Walk starts handling tasks like updating customer information or locating a customer’s order. The agent is still contained, but leans on more context and more complex routing.

Run is the fun stuff: placing or modifying an order, walking a customer through a lending application, or anything that touches multiple systems and real business logic. This is where AI agents become revenue powerhouses, rather than customer support un-blockers.

On the Simple team, we’re always ready to run. But if you want to build confidence and trust in your agents, this is the best way to do so.

2. Successful integration makes or breaks the customer experience

At Genesys, most contact center leaders we talked to had a similar worry: what happens when an AI agent hands off to a human? A successful handoff makes your customer’s interaction feel simpler, while a mishandled one creates a disjointed experience (even if the AI agent did everything right). A proper screen pop and warm transfer ensure your customer feels taken care of, every step of the way.

We built Simple to integrate cleanly with the tools contact centers already use, including Genesys. When a call moves from Simple’s agent to your live team, the context comes with it. Your agents aren’t starting cold, and your customers aren’t repeating their account number for the third time.

3. Voice quality is still the most important piece to the customer experience

An awkward pause, a reply that comes a beat too slow, a voice that sounds a little too perfect. It doesn’t matter how accurate your agent is if the conversation feels stilted and unnatural. We’ve found, over and over again, that voice quality is the most important aspect of truly seamless voice AI.

We recently announced our own proprietary model, Tango, that understands the natural rhythm of conversation. It knows when someone’s done talking, when they’re pausing to think, and when they’ve been interrupted. The result is an agent that responds at just the right moment, and feels intrinsically human. Read more about Tango.

Simple Wishline, and the rest of our Genesys Xperience

If you were on the Strip during Genesys, you may have run into one of our Wishline trucks or taxis. You may have even called in, or had your wish granted! Anyone who saw our Wishline number in the wild was welcome to call in and make a wish with Wendy the Wishline Agent. Wendy took in details and the wildest of wishes (including F1 tickets and seats to an upcoming Nashville Panthers game), while showing first-hand what Simple AI’s voice agents can do.

In case you missed it, our CEO spoke to a full house about voice AI, how we build it, and how it can actually move the needle for contact centers today. She walked through the infrastructure that most people never see and the results it adds up to: a 70% resolution rate across our deployments, and higher upsell and booking conversion than live reps. Her favorite reframe of the talk: stop asking how to fix your worst agents, and start asking how to make your best agents even better. Then use AI to get everyone there.

In all of our conversations, we kept hearing how eager companies are to bring AI to their contact centers. They want to harness the power of Voice AI, but haven’t yet found the product that fits their needs. It’s the problem we’re here to solve, and all of your insights brought us that much closer. A big thanks to Genesys for an event well run, as well as everyone who stopped by our booth, made a wish, or attended my talk. I look forward to the many conversations we’ll continue to have.

Want to talk through what crawl, walk, run could look like for your team?

Five tests for evaluating a voice AI vendor

“Revenue agent” and “deflection bot” are useful sales shorthand. They are poor product categories. The same platform may handle either role; the workflow, permissions, and scorecard determine what it actually does. Buyers should ask whether the agent can finish valuable work safely in their systems and whether a production pilot proves the economics.

Customer-care priorities have already widened beyond cost reduction. In McKinsey’s 2024 survey of more than 340 customer-care leaders, one-third named revenue generation as a priority, while 11% named reducing contact volume. An evaluation centered on containment can miss the work closest to growth.

1. What does the scorecard reward?

A vendor’s default dashboard shows what its product was designed to optimize. Extra charts prove little. Ask which outcomes the vendor can calculate from your order, payment, contact-center, and CRM data, and which events sit behind each metric.

For service work, useful measures include first-contact resolution, repeat-contact rate, customer satisfaction, and cost per completed task. For sales work, use completed orders from eligible calls, average order value, gross margin, upsell acceptance, and revenue recovered from callers who would otherwise abandon.

Containment still matters because it shows how often the agent finishes without a person. On its own, it says little about customer value or additional revenue. A contained call may lead to a repeat contact, refund, cancellation, or lost sale. Track those outcomes for long enough to see whether the original call was truly successful.

Choose one number that will decide the pilot before it starts. Then name the warning signs that could outweigh a gain in that number. A sales pilot might optimize profit per eligible call while limiting increases in refunds, repeat contacts, transfers, complaints, and abandonment. Define an “eligible call” in plain terms so nobody can change the denominator after seeing the results.

Revenue handled by the agent is different from revenue the agent added. To measure the difference, keep some comparable callers on the current call path. The cleanest setup assigns eligible callers at random and keeps each caller on the same path throughout the test. That split gives both groups a similar mix of customers and demand; NIST explains why random assignment produces more dependable comparisons. If caller-level assignment is impractical, split comparable time blocks or customer groups and document the compromises.

Build the business case on profit from retained orders. The calculation can stay simple:

Eligible calls × the difference in retained-order rate × average profit per retained order, less the added platform, telephony, implementation, and service costs.

Finance should be able to reproduce every term from company data. The vendor’s dashboard can help explain performance, but it should never be the only place where the result exists. A broader voice AI scorecard should connect containment to dollars before the pilot starts.

2. What work can it complete in your systems?

An integration logo proves that two systems can connect. Revenue work requires the agent to use catalog, inventory, pricing, promotions, cart, payment, order-management, and customer-history data. It also has to make changes safely when a call goes off the happy path.

Give each vendor scenarios drawn from real calls: a clean purchase; an invalid promotion or unavailable item; and a service call that develops into a sale. Add a failure case, such as a payment timeout after the caller confirms an order.

Require a live sandbox demonstration and inspect the records afterward. Check whether the agent:

• verifies price, inventory, address, and payment status before confirming the order;
• prevents a retry from creating a duplicate charge or order;
• has access only to the systems and actions needed for the task;
• records what it did and when;
• handles timeouts, partial failures, and unclear caller responses safely; and
• gives the human agent the transcript, information already collected, and exact point of failure.

These checks show whether the agent can finish the job cleanly. A fluent conversation or successful API call covers only one part of the transaction.

In Simple’s Omaha Steaks deployment, a narrow sales flow expanded to roughly 100 searchable products and wrote completed orders into the company’s homegrown order system. That supports the claim that the agent can execute the workflow. A separate comparison is needed to show how many additional orders it produced.

3. How does it perform when a caller is ready to act?

A production pilot should test the complete sales motion: recognize what the caller wants, make an appropriate offer, complete the action, and preserve the customer experience.

Agree on the test rules before routing traffic. Decide which callers qualify, how calls will be split, which number determines success, which warning signs can stop the test, how long later outcomes will be tracked, and how much improvement is worth buying. Keep the offer mix, season, and caller population comparable between the AI and current paths.

Show the total number of calls and the absolute lift alongside percentage rates. Continue the test until there is enough traffic to distinguish a repeatable improvement from normal week-to-week variation.

For a sales deployment, track completed-order rate, profit per eligible call, upsell acceptance, abandonment, transfers, repeat contacts, cancellations, refunds, complaints, and payment failures. Read the measures together. A system that increases order count through heavier discounts may reduce profit. Higher containment can also create more work later if repeat calls rise.

Omaha Steaks provides a useful starting point. Its temporary holiday agents attached upgrades on about 22% of eligible calls, while Simple reports that its deployed AI runs at 28% to 30%. The comparison covers one retail workflow, and the source is a vendor case study. The published account does not provide the traffic count or enough detail to show whether the callers, offers, and time periods were directly comparable. Buyers should rerun the comparison on their own traffic under fixed rules.

Abandonment needs the same discipline. Omaha Steaks reports a decline from 16% two years ago to 2% today. The improvement matters, although staffing, routing, demand, or other changes during those two years may explain part of it. Count how many additional callers complete profitable orders, then compare that result with callers kept on the existing path.

4. How does the contract count success?

Pricing model is a weak product category. Per-resolution, usage, capacity, and outcome-based contracts can all support service or sales. Buyers need to understand what each model rewards and whether they can check the invoice against their own data.

Stripe’s guide to outcome-based pricing lays out the essentials: both sides need one definition of success, a shared data source, clear rules for shared credit and edge cases, and a regular way to compare counts. Before comparing rates, define:

• the exact event that triggers a charge and the system that records it;
• how long after the call a transfer, reopened issue, repeat contact, cancellation, refund, or partial completion can change the count;
• who receives credit when a customer starts online, changes channels, or finishes with a person;
• what happens during fraud, vendor outages, customer-system outages, and test traffic;
• how long records are kept and how disputes, audits, and credits work; and
• how minimums, caps, overages, and peak-season volume affect the bill.

Record operational success and billing status separately. A useful call may fall outside the contract’s billable definition, and a charged event may later be reversed. Keeping both records makes disagreements easier to resolve.

Recreate the invoice using the previous 12 months of traffic, including the busiest weeks. Run low, expected, and high cases. This exercise exposes vague billing units, expensive minimums, and bad peak assumptions before the contract reaches production.

5. What can reference customers prove?

Case studies show a vendor under favorable conditions. A reference becomes useful when its operation resembles yours. Look for similar call types, transaction complexity, systems, regulation, and seasonality.

For each performance claim, ask for:

• the starting point, what the AI was compared with, and how traffic was divided;
• the raw counts behind the rate, total traffic, and measurement dates;
• which calls were included or excluded and how long later outcomes were tracked;
• the company system that supplied the result;
• changes in staffing, routing, offers, pricing, or marketing during the same period; and
• performance after launch, including the weakest month.

Speak with the operator who owns the metric. Ask finance or analytics to confirm where the numbers came from. A conversion or recovered-revenue claim is impossible to judge when the vendor cannot show who was counted, what they were compared with, or how credit was assigned.

The reference call should end with a one-sentence pass/fail test for your pilot. For example: “For callers eligible to place a catalog order, the AI must raise profit per call by at least 8% while keeping any increase in refunds and seven-day repeat contacts below one percentage point, compared with similar callers on our current path over six weeks.” If the vendor cannot turn its strongest case study into a test like this, the case study is positioning.

Choose for the work, then prove the value

A support-heavy line may benefit most from faster resolution at lower cost. Measure repeat contacts and customer satisfaction alongside containment so failures cannot disappear into later calls.

A sales-heavy line should test completed transactions, recovered demand, and profit against the existing path. Mixed lines need separate rules and scorecards by call type; a blended average can hide gains in one area and losses in another.

Choose the vendor that can complete the priority task in the system of record and show the added value under counting rules your team can audit. Keep the category label as shorthand. Let the pilot determine what the product is worth. The broader voice AI buyer’s guide covers the security, latency, implementation, and governance checks that sit beside these five commercial tests.

If you have strong engineers, building your own voice agent is tempting. Vapi provides an orchestration layer for speech-to-text, an LLM, and text-to-speech. ElevenLabs offers an agent framework built on some of the best-known voice models in the market. A competent backend engineer can have a prototype answering calls in days.

We compete with these tools every week. Teams usually discover that the prototype is the easy part. Production work sits in turn-taking, business logic, contact-center integrations, reporting, capacity planning, and ongoing operations.

What you get by building

Building gives you control. You choose each component, tune every prompt, and pay providers directly without a platform margin. That can be the right trade for a narrow internal use case, a developer-heavy team, or a product company embedding voice into its own offering.

You also get to a demo quickly. On either stack, a capable engineer can usually build one over a weekend.

What production requires

The voice layer is only one component of a contact-center agent. A prototype proves that the agent can hold a conversation. It does not prove that the system can handle real callers, business rules, peak traffic, or the needs of the people running the contact center.

• The system has to recognize when a caller has finished speaking. Poor turn-taking makes the agent interrupt people or leave long silences, and both drive callers away.
• p95 latency matters more than the average. If the pipeline chains four vendors together, its worst delays compound across four services you do not control.
• The agent needs your business rules: promotion validation, address requirements, loyalty tiers, campaign pricing, and catalog quirks. Omaha Steaks has millions of possible campaign combinations. An API key does not encode any of them.
• Your contact-center staff need warm transfers, screen pops, partial-containment tracking, and QA workflows. The deployment includes what employees see and do when the agent cannot finish the call.
• Operations and finance teams need intent categorization, containment reporting, and anomaly alerts. “The calls seem fine” will not satisfy a CFO asking for proof of ROI.
• Peak demand tests every dependency. If December traffic is 12 times the steady-state volume, every provider in the chain must scale on the same day, at the service level you purchased.

What two customers learned

Before working with us, Omaha Steaks spent six months building a virtual agent in-house with its CCaaS tools. The team had experienced developers, but the agent contained only 20% of calls for one use case and created a maintenance burden nobody wanted. As the team put it, “None of us were experts in AI, and it showed in the timeline.”

Another customer built a DIY agent on a voice API. The voice sounded great, but the team still had to create the orchestration, analytics, agent-assist tools, and iteration workflow. They were doing that work alongside their regular jobs. The project stalled after the demo because the surrounding production system demanded more time than the team could give it.

Four questions to ask before you build

1. Is voice AI part of your product? A company embedding voice into its own product may benefit from owning the stack. If voice AI supports contact-center operations, compare the full cost of engineering and maintenance with the platform fee.
2. Who will own the system two years from now? Models, providers, and telephony infrastructure will keep changing. The work continues long after launch.
3. What would an outage cost during peak season? If a December outage would cost millions, decide who should own capacity planning and the pager.
4. How often will you need to make changes? With a platform and dedicated agent engineers, Omaha Steaks ships changes in hours and runs experiments every week. A DIY stack moves on the engineering team’s sprint schedule.

Building on Vapi or ElevenLabs can be the right choice, especially when voice AI is part of the product and the company is prepared to own the stack for years. The cost model should include the production system and the people who will maintain it, not just provider bills.

Many teams we meet build the demo, discover what production requires, and then buy a platform. If your requirements already point in that direction, evaluating a platform before a six-month internal build is usually cheaper.

“We just finished a CCaaS migration. We are not doing another one.”

That is a reasonable response. A contact-center migration touches phone numbers, routing, workforce management, recordings, reporting, agent desktops, security reviews, and training. Adding voice AI should not force you to reopen every one of those decisions.

It usually does not have to. You can add AI to an existing contact center as a new call-handling layer. The CCaaS still routes calls to people, manages the workforce, records human conversations, and runs the agent desktop. The AI answers selected calls first, completes the work it is allowed to complete, and sends the rest to an existing queue.

This is an overlay deployment. It changes the path of a call without replacing the platform your agents already use.

How the overlay architecture works

Start with a phone number or a defined slice of traffic. The carrier routes those calls to the voice AI agent before they reach the current IVR. The agent greets the caller, identifies the reason for the call, and attempts the approved task.

If the agent can finish the task, it confirms the outcome and ends the call. If the caller asks for a person, the task falls outside the agent’s scope, or a system fails, the agent transfers the call to the appropriate queue in the CCaaS.

The connection can use ordinary telephony components: number forwarding, PSTN transfer, or SIP trunks, depending on the environment. SIP REFER is a standard way to move an active call from one endpoint to another; Twilio documents the transfer flow, including transfers to SIP and PSTN destinations. Existing CCaaS products also support third-party telephony connections. Genesys, for example, documents BYOC options for connecting external carriers and SIP trunks.

The exact design depends on your carrier, CCaaS, geography, compliance requirements, and whether the AI needs to remain on the call during a handoff. None of those choices inherently requires a CCaaS replacement.

A transfer and a contextual handoff are separate things

A phone transfer moves the caller. It does not automatically move the information collected during the conversation.

To give the human agent useful context, the AI also needs a data path. That might create a CRM activity, add a note to the customer record, populate a screen pop, or send a structured summary to the agent desktop. The handoff can include the caller’s intent, authentication status, information already gathered, actions completed, and the reason for escalation.

This distinction matters during vendor evaluations. A vendor may demonstrate a successful transfer while leaving the human agent blind. Ask to see the full handoff: the call arrives in the right queue, the context appears before the agent answers, and the caller does not repeat the opening five minutes.

Warm transfers need their own failure logic. What happens if the queue is closed, the target does not answer, or the transfer request fails? The AI should tell the caller what happened and follow a defined fallback, such as another queue, a callback request, or voicemail. A dropped call is not an escalation strategy.

What stays in your CCaaS

Your current platform can keep the jobs it already performs well:

• human queue routing and skills-based distribution
• workforce management and scheduling
• the agent desktop and supervisor tools
• recordings and quality workflows for the human leg
• existing reports on queue performance, handle time, service level, and agent activity

The AI platform adds a different set of operating data: why people called, which tasks the AI attempted, completion and escalation rates by use case, where conversations failed, and what happened before a transfer.

Do not assume both platforms will show the same interaction from end to end. The CCaaS will usually have detailed telemetry for the transferred leg. The AI platform will have the full AI conversation and transfer outcome. Decide before launch which system owns each metric and how the teams will reconcile call IDs. Otherwise, the first business review turns into an argument about denominators.

This division of labor is consistent with the role of a CCaaS. Gartner describes CCaaS as the platform that orchestrates self-service and employee-assisted engagement across channels. An overlay changes one part of that engagement model without asking you to discard the rest.

How much integration do you need?

The answer depends on the task, not the existence of AI.

A first deployment can run with little or no CRM integration when the caller supplies the necessary information and the outcome can be handed to a person or delivered through an existing channel. Common examples include routing, lead qualification, after-hours intake, basic FAQs, and callbacks.

Read access lets the agent personalize the conversation or check information such as an order status, appointment window, or account detail. Write access lets it finish transactions: book an appointment, update a record, place an order, or process a payment through an approved workflow.

More integration expands what the agent can complete, but it also expands the security review, test surface, and consequences of an error. Start with the minimum access required for a valuable outcome. Add write permissions only after the team has defined validation rules, rollback behavior, audit logs, and human escalation.

This is especially useful in environments with closed or heavily customized systems. A home-services operation can begin with after-hours intake or speed-to-lead while deeper CRM work continues. The launch and the ideal end state do not need to happen on the same day.

A rollout that does not disturb the contact center

Choose one entry point where intent is predictable and success is measurable. A dedicated campaign number, after-hours line, overflow queue, or narrow service flow is easier to test than the main support number.

Omaha Steaks followed that pattern with a dedicated mailer line. Callers were usually ordering the package shown in the mailer, so the first flow was narrow but commercially meaningful. The team proved the transaction end to end, then added products and use cases. The full Omaha Steaks case study covers the rollout and results.

Before the first live call, document five things:

1. Which calls reach the AI, and how traffic can be returned to the old path.
2. Which tasks the AI may complete and which require a person.
3. Where each escalation goes during business hours and after hours.
4. What context reaches the human agent and where it appears.
5. Which metrics decide whether the pilot expands, changes, or stops.

Run a small share of traffic first. Review recordings and transcripts daily, then inspect completion, transfer, abandonment, latency, and customer outcomes by intent. An average containment number can hide a broken high-value flow, so segment the results.

Keep a rollback path that the contact-center team can use without waiting for the vendor. If the deployment began by pointing one number or changing one routing rule, reversal should be equally bounded.

What changes for agents and supervisors

Agents keep their desktop, queues, and QA process. The work arriving in those tools changes. Routine calls may disappear; escalations should arrive with more context and a clearer reason for transfer.

That shift affects staffing forecasts and coaching. A lower volume of easy calls can raise average handle time because the remaining calls are harder. It can also change quality scores if supervisors compare the new call mix with the old one. Track complexity and intent alongside handle time so the team does not mistake a healthier queue for declining agent performance.

Supervisors also need visibility into the AI’s queue. Someone should own failed intents, escalation spikes, incorrect actions, and prompt or workflow changes. Adding an AI layer does not remove operational work; it moves part of that work from scheduling people to inspecting automation.

When an overlay is the wrong choice

A replacement may be sensible when the CCaaS contract is already ending, the platform cannot support required routing or data access, the telephony vendor is sunsetting a critical product, or the organization wants to consolidate channels and reporting anyway.

Treat that as a platform decision with its own business case and timeline. Bundling it into the AI launch increases scope and makes it harder to tell which change produced the result.

An overlay is also a poor fit if the vendor cannot explain call ownership, failover, data retention, recording responsibilities, emergency routing, and transfer behavior in your environment. “We integrate with everything” is not an architecture.

Questions to ask before you sign

Ask the vendor to draw the proposed call path using your carrier, numbers, CCaaS queues, CRM, and reporting tools. Then ask them to show the failure paths.

• Can we start with one number or a small percentage of traffic?
• Who controls routing, rollback, and after-hours behavior?
• Does the transfer use SIP, PSTN, or another method?
• How does context reach the human agent?
• Which system records each leg of the call?
• What appears in our existing CCaaS reports?
• What can launch before the CRM integration is complete?
• Which actions require read or write access?
• What happens if the AI platform, an API, or the transfer target is unavailable?

These questions belong beside the broader commercial and operational checks in The CX Leader’s Guide to Voice AI.

You should be able to add voice AI without reopening the entire contact-center stack. A narrow routing change, a clear handoff, and a reversible pilot are enough to learn whether the agent can complete useful work. If a vendor cannot describe the deployment while leaving your CCaaS in place, the migration is part of their product strategy, not a technical requirement of voice AI.

Two years ago, Omaha Steaks lost 16% of callers before anyone could help them. Last year, the rate fell to 9%. Today it is 2%, measured on a line that handles roughly 12 times its normal call volume in December.

That change carries real revenue weight. Roughly half of Omaha Steaks’ inbound calls are sales calls. When one of those callers hangs up, the company loses a customer who was already close enough to buying that they picked up the phone.

Call abandonment rate is easy to bury among service-level metrics. On a sales line, it belongs next to conversion rate. Both tell you how effectively the business turns existing demand into revenue.

What changed on the line

The improvement came from three practical changes to how calls entered and moved through the queue.

First, answer time went to zero. An AI agent picks up on the first ring and can handle calls concurrently, including the holiday surge. The queue that used to form in front of a fixed number of people no longer forms before the greeting. That kind of capacity is particularly valuable for retail voice AI, where a few peak weeks can carry an outsized share of annual revenue.

Second, most callers no longer need the human queue. Omaha Steaks reached about 60% overall containment, so the majority of callers finish what they came to do with the AI agent. People who need a person still get one, but they enter a much shorter line.

The third change was less obvious. Callers who speak with the AI and then choose to wait for a person stay on the line about 30 seconds longer than callers did before. Being greeted, understood, and told what will happen next appears to make the wait feel more tolerable. Across a holiday queue, another 30 seconds is enough to keep a meaningful number of calls from disappearing.

“Customers will, once they’ve talked to the AI, they’ll wait on hold a little bit longer. And yeah, on average it’s about 30 seconds more for us, which is a really nice thing to learn for when we staff.”

Grant Young

Senior Manager, CEC Operations, Omaha Steaks

Why a two-percent abandonment rate is worth more than it looks

Consider a sales line receiving 10,000 calls in a peak week. At 16% abandonment, 1,600 calls end in the queue. At 2%, that number falls to 200. The difference is 1,400 conversations that the business now has a chance to complete.

That is an illustration, not Omaha Steaks’ reported weekly volume. But the revenue logic is straightforward. Multiply rescued sales calls by the conversion rate for answered calls, then by average order value. If one-third of those 1,400 callers buy, the change produces about 467 additional orders from demand the company had already acquired.

Peak weeks expose the real constraint

An annual average can hide a queueing problem. At Omaha Steaks, call volume rises to about 12 times steady state during the holiday period. Hiring enough people for that peak meant recruiting months in advance, training thousands of seasonal workers, and still accepting that queues would lengthen when demand arrived faster than agents could finish calls.

Concurrency changes that equation. The AI answer layer does not need a December staffing plan to pick up the first ring. It can complete routine work immediately and reserve the human queue for calls that genuinely need judgment or escalation. The bigger gain is structural. It removes the sharpest mismatch between demand and capacity.

Companies can test this without replacing the rest of the contact-center stack. A narrow number, overflow queue, or defined slice of traffic can route through an AI agent first, then transfer to the existing CCaaS when necessary. The CX Leader’s Guide to Voice AI covers the questions buyers should ask about integrations, handoffs, security, pricing, and operating ownership.

How to measure call abandonment without fooling yourself

Start by agreeing on the denominator. Some teams count all offered calls; others remove callers who hang up within a few seconds. Either method can be defensible, but a before-and-after comparison is useless if the definition changes during the project.

Then segment the number. Separate sales from service, and review abandonment by intent, hour, day, queue, and transfer outcome. A single blended rate can improve while a valuable sales flow gets worse. Omaha Steaks learned a related lesson when conversational routing showed that its phone tree had overstated sales intent for years. The details are in the analysis of why phone tree data misreads customer intent.

Finally, measure the whole journey. Track how quickly the AI answers, how many calls it completes, how many it transfers, how long transferred callers wait, and whether those callers ultimately buy or resolve their issue. A low abandonment rate is only valuable when customers reach a useful outcome.

A practical way to reduce call abandonment

Begin with the part of the queue where abandonment is most expensive. On a sales line, that may be a dedicated campaign number or a narrow ordering flow. On a service line, it may be after-hours intake or a high-volume status request. Choose a slice with predictable intent, enough traffic to measure, and a clear outcome.

Route a small share of calls through the AI answer layer and keep the existing queue as the fallback. Review recordings and transcripts daily. Watch abandonment alongside completion, transfers, latency, conversion, and repeat calls. If the overall rate falls but one intent starts producing bad handoffs, fix that flow before expanding traffic.

Omaha Steaks’ result came from removing dead time at the start of the call, completing more work before a queue was needed, and giving transferred callers enough context to keep waiting. The drop from 16% to 2% is the headline. The more useful lesson is that queue performance can improve without manufacturing more demand or rebuilding the entire contact center.

Every rescued call is a second chance to convert demand that marketing already paid for. That makes abandonment one of the clearest places to connect contact-center operations with the economics of AI-assisted sales.

Most voice AI evaluations fail before the demo starts, because the buyer walks in without a list of questions that separate a really good product from a well run demo. This guide is that list: twelve questions you can ask your prospective solution providers, the point of each question, and what a strong answer sounds like.

1. Is the agent conversational or a flowchart?

Deterministic bots walk a tree: ask, capture, branch. They break the moment a caller says two things at once. Ask the vendor to demo a caller who interrupts, changes their mind mid-order, and gives information out of sequence. A strong answer is a live demo that survives all three, and an architecture explanation that doesn’t include the phrase “dialog flow.”

2. Can it complete tasks, or just talk?

"Agentic" means the system takes actions: placing orders, updating records, booking appointments, processing claims. Ask exactly which systems the agent will write to on day one, and what a completed task looks like in your stack. A strong answer names APIs and shows a task finishing with no human touch.

3. What happens on day one if our CRM has no open API?

Some deployments stall for months waiting on integration work. Get specific: what can launch with zero integration? Inbound flows often can, because the caller supplies the information the agent needs. One of our home-services customers went live while its CRM vendor kept the API closed, and Omaha Steaks exposed APIs from a homegrown AS400 stack in about a week with our engineers alongside theirs.

4. What’s the realistic time to first live call?

Six months is an answer you should question. Eight days is an answer you should verify. Omaha Steaks signed a proof of concept on June 2 and took first live calls June 10; ask any vendor for their equivalent story with dates and a reference customer you can call.

5. How fast does it respond, and how does it know when to speak?

Voice tolerates far less delay than chat. Response latency around 800 milliseconds is where conversations start feeling natural; beyond that, callers talk over the agent and loops multiply. Ask what the vendor’s median and p95 response times are on real calls, and whether the turn-taking and voice-activity models are their own.

6. Who owns the model stack?

If the vendor is a thin wrapper on someone else’s transcription, inference, and voice APIs, their latency, cost, and peak-season capacity are all someone else’s decisions. Ask which pieces they host and auto-scale themselves. If they rent everything, ask how they handled last December’s peak, and pay attention to whose engineers did the work.

7. How does it hand off to a human?

Every agent needs to escalate sometimes. The difference is whether the context survives to sets the human up for success. A strong answer includes a warm transfer, a screen pop showing the human agent everything gathered so far, and partial-containment reporting so you get credit for the half of the call the AI already did. The customer should never repeat themselves.

8. What does the analytics layer show me?

You want intent categorization across 100% of calls, containment and conversion by use case, and alerting on anomalies. Mid-peak last December, our monitoring caught a spike in callers saying the add-to-cart button was broken on the Omaha Steaks website; their web team shipped the fix before the QA vendor had transcribed a single call. Ask the vendor for the story their analytics caught that nobody was looking for.

9. How do we test changes safely?

Prompt edits, voice changes, and script experiments need versioning, a sandbox, and ideally A/B testing on live traffic. Ask how long a change takes to ship, and how you’d measure whether it helped. “Hours” is a good answer. “Submit a ticket” means every experiment waits in the vendor’s queue.

10. What’s the security and compliance posture?

SOC 2 Type II at minimum. HIPAA if health information can come up. Real-time PII redaction in transcripts. Ask where audio and transcripts are stored, for how long, and who can access them. A strong vendor answers without checking with legal.

11. What does it cost, and what makes the invoice grow?

Per-resolution pricing puts the vendor in charge of deciding what “resolved” means. Usage pricing is more predictable but demands volume estimates. Whatever the model, ask for the invoice math on your actual call volume, peak month included, and get the overage terms in writing.

12. Who works with us after launch?

The deployments that hit 70% containment and hold it are the ones where vendor engineers meet with the customer team weekly and treat the roadmap as shared. Ask who your named engineers are, how often you’ll meet, and what the last customer shipped with them in month six. If the answer is “documentation and a support portal,” plan accordingly.

Scoring the answers

No vendor aces all twelve. Weight the ones tied to your economics: if your inbound line sells, questions 1, 2, 5, and 9 decide revenue. If you’re regulated, question 10 is a gate. If your stack is homegrown, question 3 will decide your timeline.

Make every vendor answer the same twelve questions in the same order. By the third demo, you will be able to sense which ones were rehearsed.

0–3: Run.
4–8: Worth a pilot.
9–11: A serious contender.
12: Simple.

For years, Omaha Steaks believed 70% of its inbound calls were sales calls. The evidence looked solid: the IVR said "press one for sales, press two for support," and 70% of callers pressed one.

Then an AI agent replaced the phone tree and started asking callers, in plain language, what they needed. The real mix turned out to be only 45 to 50% sales.

A quarter of the traffic had been misclassified for years. The company had built its entire seasonal training program on that number, so thousands of temporary agents were drilled on selling holiday packages, then spent much of December fielding claims and delivery questions they weren't trained for.

Why key presses lie

Callers don't press buttons carefully. They want a human, and they've learned the fastest route is the first option, or zero, or saying "representative" until something gives. Menu order, wording, and how long the hold message runs all shape the numbers, so what the IVR records is escape behavior.

Every downstream decision inherits that error: staffing models, training curricula, skill-group ratios, and the business case for automation all get built on data that measures which button was easiest to press.

What honest intent data changes

When the routing layer understands what the caller actually said, the numbers underneath it move. Misrouted calls carry a transfer cost and usually a repeat-yourself cost, and Omaha Steaks cut both by letting the agent classify intent from the caller's own words. Knowing the real sales-to-service split also lets you train and schedule against it, which matters most in peak season, when every misallocated trainee is money spent on the wrong skill.

Intent captured at conversational resolution shows you what a phone tree can't: which products drive calls, which policies confuse people.

What else is in there?

If twenty points of your call mix can hide inside your IVR data, what else is hiding there? Most contact centers can't say, because the phone tree is the only intent instrument they have.

Fixing it doesn't require automating a single conversation. Replacing menu buttons with an agent that asks "how can I help?" and routes on the answer is the smallest AI deployment available, and it pays for itself in data quality. Omaha Steaks found a 20-point gap in week one. Yours will be a different number in a different place, and you won't know which until you measure it.

Two years ago, Omaha Steaks hired 5,000 seasonal agents to get through the holidays. This year, the plan is 692.

That number is the spine of the conversation our CEO Cat Li had on stage at Customer Contact Week in Las Vegas with Grant Young, Director of CEC Operations, and Rob Bradshaw, Quality Manager at Omaha Steaks. The full fireside ran 45 minutes, but we've added a 9-minute highlight reel for the masses.

The staffing math that used to run the holidays

In the few weeks around December, Omaha Steaks’ call volume runs about 12x normal. Staffing for that meant hiring in September: recruit 5,000 people to net roughly 2,000 still on the phones by the time peak arrived.

Six months of building it themselves

Before Simple, the team spent six months building an IVA in-house on their CCaaS provider's tools. Grant, an analyst, and a group of developers got it to 20% containment on a single use case. It worked well enough to prove the idea and cumbersome enough to prove they shouldn't be the ones building it.

Eight days to live calls

The proof of concept with Simple was signed on June 2. By June 10, the AI was taking real customer calls on a dedicated order line: full orders placed in the system, no human on the line. Containment on day one was 52%. Grant's reaction, verbatim from the stage: "Holy shit, are we hanging up on customers?"

Meet Simone

Omaha Steaks named the agent Simone, after the Simon family that built the company. Callers hear the AI disclosure in the first sentence or two of the call. A few minutes in, plenty of them forget. They call her ma'am. One customer invited her to his daughter's birthday party.

The add-to-cart save

December 13, right at the peak of peak season. Zach, Simple's CTO, noticed a spike in callers saying the add-to-cart button on the Omaha Steaks website wasn't working, and reached out directly. At holiday volume that bug was worth millions in exposure. Omaha's team had it fixed before their QA vendor had even finished encoding the calls that would eventually have surfaced it. Nobody's contract required that call. That's the point of the story.

What about the people

Rob's QA team was about 20 people whose jobs the AI would clearly change. Omaha Steaks told them so on day one, then invested in training them toward more analytical work. Two took buyouts. Everyone else is still with the company a year later, in roles built around directing and improving the AI rather than manually reviewing calls.

The results, two years on

Call abandonment went from 16% to 9% to 3%. Seasonal hires upsold on 22% of calls; the AI runs at 30%, a 36% higher rate and level with Omaha Steaks' most experienced reps. And seasonal hiring went from 5,000 to 1,500 to a planned 692.

Watch the nine-minute reel above for the story in Grant and Rob's own words. The full 45-minute conversation is here: https://www.youtube.com/watch?v=ZCmXLwly76c

The Bureau of Labor Statistics projects employment of customer service representatives to decline 5% between 2024 and 2034. According to the 2026–27 benchmarking report from CMP Research, the group behind CCW Digital, the frontline customer contact workforce fell below that projected 2034 level in 2025: 2,595,750 people, against a projected 2,673,300 for 2034. More than 218,000 frontline roles disappeared between 2024 and 2025 alone.

The official models gave the industry a decade to absorb this shift. It took about a year.

The trust gap

Numbers like that usually come with a wave of confident technology adoption behind them. CMP found the opposite. When it surveyed contact center executives, improving agentic and generative AI ranked among the most important issues facing their operations and among the hardest to solve. Managing the change of an AI-augmented workforce sat right beside it.

CMP reads the drop as a signal that automation may be changing the workforce faster than anyone projected. The people running contact centers apparently agree, and they still don't trust the current generation of customer-facing AI products to deliver. Every sale in this category has to cross that gap.

Nobody can tell the products apart

CMP's awareness data explains some of the skepticism. The best-known tool in the category is recognized by just 42% of the buyers it's sold to, and awareness of three out of four tools falls under 20%. The takeaway is this: no one understands what vendors are selling. After years of funding and noise, even the highest profile products still draw a blank from most of the market.

Vendors own a share of that fault, according to CMP: buyers and sellers often name the same thing differently. A vendor sells an "intelligent virtual agent" or "agentic voice." The buyer is shopping for a "voicebot" to replace a tired IVR. One business function has been split across three product categories, so buyers end up guessing, and a buyer who has to guess usually stalls.

What to do with this if you run a contact center

Ignore the category names. Evaluate what a system completes on your calls, in your stack. A product that finishes an order or books an appointment is worth paying for whatever its vendor calls it. A product that can't survive a live demo on your hardest call type isn't.

Treat the workforce shift as a management problem you have now. The teams handling it well tell their people early and retrain them toward the analytical work that automation creates.

And hold vendors accountable for the trust gap. The industry's skepticism is earned, which means the burden of proof sits with vendors. Ask for named customers and real numbers, and insist on a pilot measured in days. The vendors who can't produce those are part of the reason awareness tops out at 42%.

The future of the contact center workforce stopped being a forecast. Every planning assumption downstream of that number deserves a fresh look.

Omaha Steaks has been selling premium steaks and seafood for 109 years. Half of its annual call and chat volume arrives in a single month. From the week before Thanksgiving through mid-December, volume runs 12 times higher than steady state, because everyone wants to send steaks as a holiday gift.

For years, the company solved that spike with bodies. It hired around 5,000 seasonal workers starting each September to end up with roughly 2,000 actually taking calls by December. Training started in mid-September and ran as a revolving door: hire, train, quit, rehire. Supervision stretched from one supervisor per 20 agents to, at the worst point, one per 212.

Grant Young, who runs contact center operations at Omaha Steaks, describes the old holiday season simply: a disaster every year. Getting it right was never optional.

The first attempt: build it in-house

Before working with Simple AI, the Omaha Steaks team built its own virtual agent on the tools inside its CCaaS platform. Grant, an analyst, and a group of developers spent six months on it. The model was deterministic, so every path had to be mapped by hand, and a caller saying “my address is” in the wrong spot could throw the whole flow off.

The result was 20% containment on a single use case. Useful, but nowhere near the seasonal problem, and expensive to maintain for a team whose members were not AI specialists.

The pilot: signed June 2, live June 10

When Omaha Steaks evaluated vendors, the bar was set by its customer experience standards: the agent had to hold a natural conversation rather than walk a script tree, it had to complete real work end to end, and it had to sound good enough to represent a premium brand.

The team started narrow on purpose. Omaha Steaks runs dedicated phone numbers for its mailers, so a caller on one of those lines is almost certainly ordering the package from the mailer. That system offered Simple a contained first use case that still finished the job: take the order, place it in the order system, no human touch.

Simple flew to Omaha and worked alongside Grant’s team for two weeks. The proof of concept was signed on June 2. First live calls happened June 10. That eight-day window included new APIs on the Omaha Steaks side, since the company runs their business on a homegrown order and CRM stack rather than an off-the-shelf platform.

Day one containment was 52%. Grant’s first reaction was alarm: was the new system hanging up on customers? (It wasn’t.) Three weeks in, containment held steady at 60%.

Crawl, walk, run

From there the rollout followed a waterfall: while one use case moved to production, the next was in configuration. The single-package line expanded to roughly 100 products, searchable by item number or name. A month of testing later, in September, the team turned the agent loose on 100% of traffic.

The agent, named Simone after the founding Simon family, now greets every caller, understands why they’re calling, and either handles the call or routes it with full context. Sales came first. Service calls that need a more careful touch, like claims and delivery problems, are rolling out now.

The numbers after one year

• Containment: 60% across all calls, 70 to 75% on the dedicated sales lines, and 75% on chat.
• Abandonment: 16% two years ago, 9% a year ago, 2% today.
• Upsell rate: Simone runs an upsell rate of 28% to 30%, compared to the 22% of seasonal agents.
• Seasonal hiring: from 5,000 hires down to 1,500 trained last holiday, with 692 planned for this year.
  

Average order value tells the same story. Callers who complete an order with Simone match the numbers of tenured agents, and because far fewer calls abandon, more calls become orders at all.

Why it worked

Three decisions mattered more than any feature.

First, Omaha Steaks framed the project around revenue rather than deflection. The inbound line is where the company sells, so the goal from day one was completed sales rather than contained calls.

Second, both teams treated it as a partnership. A year in, the teams still meet two to three times a week, and each side has flown to visit the other. Deployment in ten days only happened because Omaha Steaks gave the project engineering resources and decision-makers, and Simple put dedicated agent engineers on their account.

Third, the rollout was incremental. One contained use case, proven end to end, then the next. Nobody just flipped a switch on day one.

If your busiest month is also your most error-prone month, the Omaha Steaks playbook is worth copying: pick the narrowest use case that still finishes a revenue task, prove it in weeks, and expand from there.

Customer Contact Week Vegas ran on a contradiction this year. Companies are more open than ever to implementing AI, but are also more skeptical than ever of the solutions on the market. Nowhere was this more apparent than in the voice AI category. According to a CMP survey, 4 out of 5 CX leaders say they're open to investing in voice AI, but only 37% are prioritizing it — a 10% decrease from the previous year.

The exception was on our own stage: Grant Young, Director of CEC Operations at Omaha Steaks and Simple AI customer, shared a containment number most of the audience could only dream of: 75% on their highest-value calls, compared to 20% with the system they used a year ago. Back at the booth, almost every conversation circled the same gap between what voice AI buyers were promised and what most of them got.

The most common conversation at our booth: searching for a new voice AI vendor

Rather than shopping for their first voice AI, the buyers we spoke to were most often looking to replace their existing solution. These companies had bought an early agentic voice product, as a module within a legacy platform or one of the first-generation voice vendors, and the results fell short of expectations.

We heard the same two things all week: hopeful about voice AI, unhappy with their current setup. These buyers have been burned, and they show up expecting the next vendor to miss the nuances of their business too. They won't be won over by another pitch, only by results.

Results are exactly what Omaha Steaks brought to their session.

Omaha Steaks and the containment numbers that shook the room

Grant Young and Rob Bradshaw of Omaha Steaks told their story on stage Thursday. They shared their containment on inbound sales calls: 55% in the first week, around 60% across the board now, and peaking near 75% on certain use cases.

When they dropped those numbers, the room gasped. Grant and Rob also could not believe it when the first numbers using Simple's voice agent came in:

Some of the other contact centers in the room bought voice AI hoping to achieve a fraction of that result. For example, Omaha Steaks' earlier in-house build from Five9's IVA studio had stalled at 20%, and it took six months to get even that far. This time, the timeline looked very different:

What they have now is an agent customers are happy to stay on the line with, one that solves whatever they called about in the first place.

The revenue most voice AI still leaves on the table

The companies that sell on the phone came looking for one more thing and mostly could not find it: an agent that captures revenue on the call instead of only containing it. Real upselling at volume is still rare across the industry, and it is the clearest place older solutions leave money on the table.

Omaha Steaks put numbers to it on stage:

That puts the Simple on par with their most experienced reps and roughly 30% ahead of the seasonal staff who carry the holiday load. This difference in seasonal performance translates into a huge amount of revenue over the holiday season. On Father's Day, their biggest sales day of the year, the AI outperformed even their steady-state agents on the most expensive upsell.

Good voice AI changes the job, not just the cost

Grant and Rob also raised an issue that gets less attention than it deserves. When a reliable voice agent takes the highest-volume, most repetitive calls, the work left for your people is the work that needs judgment. Reps handle harder problems and less of the routine grind, and the role starts to look like an engaging career instead of a volume quota. The QA team stops spot-checking a thin sample and starts working from every call, which turns them into the analysts who tell the business what its customers are really saying.

The fear underneath all of this is that automation just means layoffs. At Omaha Steaks, that isn't what happened:

"We challenged them with additional training … giving them mentored opportunities to show us that they could work in that more analytical sort of capability rather than just checking forms while they listen to calls. They were more engaged because they recognized the investment that we'd made.And the people that really didn't have the propensity to work in that capacity, we found other places for them. Of about 20 people impacted, two took buyouts and left. Everybody else is still with the company and still productive a year later." — Rob Bradshaw, Quality Manager, Omaha Steaks

Their bottom line improved, and so did the experience of the people doing the work.

What the trends at CCW Vegas 2026 mean for voice AI

One thing was certain by the end of the week. Buyers are done grading voice AI on a curve. They are judging its impact on the metrics they have always cared about—CSAT, self-service, first contact resolution, containment—, and are ready to leave vendors that try to get them to expect less.

To see what this means for your contact center, book a demo at usesimple.ai.

Leads reach out on their own schedule. Many homeowners fill out quote requests in the late evening or early morning, but most sales teams won’t see it until office hours. That delay holds a lot more significance than you might think. Speed to lead measures the gap between a lead and a call back, and the brand that closes it first is often the one that gets the deal.

What is speed to lead?

Speed to lead is how long it takes a new lead to reach a real conversation with your brand. The clock starts when the lead arrives, by form, quote request, text or call, and stops when they reach a live rep or an AI agent. An auto-reply or voicemail doesn’t count.

Why does speed to lead matter?

The best-known research on this dates to 2007, but the industry still treats it as a foundational study. MIT’s James Oldroyd and InsideSales.com studied more than 15,000 web-form leads and over 100,000 call attempts across six companies. The odds of qualifying a lead fell 21 times between a call at 5 minutes and a call at 30 minutes, and dropped more than 6 times within the first hour. The study measured qualification rather than closed sales, and it looked at web forms rather than texts or inbound calls. We read it as a direction. For a consumer brand, that direction says to count response time in minutes, rather than hours.

What is a good speed-to-lead time?

For a consumer brand, a good speed-to-lead time is seconds to a few minutes. The 2007 study used 5 minutes as its fastest window. AI agents can beat that because they don’t have to finish another call first. The lead’s clock keeps running at 9pm, on weekends and through a holiday rush: an agent’s response time stays consistent no matter what.

Which businesses feel speed to lead most?

Speed to lead matters most for consumer brands that sell over the phone, which is who we build for. A homeowner asks for a window quote, a shopper calls about an order, an applicant starts a loan application. Each one is waiting on a reply, and each often reach out long after your office has closed.

Why do leads end up waiting?

Leads wait when your employees are busy with something else. With many human reps juggling multiple responsibilities and limited resources, this makes a lot of sense. They can only talk to one customer at a time, after all. After hours, no one is there to answer. At Renewal by Andersen, the overnight overflow went to a third-party call center, and unfortunately that meant poor lead quality and misrouted leads.

The gap is widest in two moments. Who picks up when a lead arrives at 9pm, after the last shift has gone home? And who calls back when several leads land in the same busy hour? Either way, your leads are left waiting.

How can you shorten your speed to lead?

Teams usually have three levers. The first is more coverage: staffing more people across more hours, which works until volume spikes or the evening shift is the thin one. The second is smarter routing, sending each new lead to whoever is free first rather than whoever owns the account, which trims the wait. The third is putting an AI agent on the first response, so every lead gets an answer right away and a rep steps in when the conversation calls for one.

How does an AI agent respond in seconds, on the channel the lead chose?

Leads come in through ads, quote forms, click-to-call buttons and aggregator lists. Simple responds on whichever channel they used, across phone, text, email and web, and keeps one memory of the conversation so no one has to repeat themselves. A sales conversation with Simple moves through four stages.

1. The lead arrives. A form, a quote request or a text comes in, and Simple starts the conversation within seconds (or the timing you decide on).
2. Simple reaches them on their channel. A lead who called gets a call, and a lead who texted gets a text. If there’s no answer, Simple tries another channel and builds on what the lead already shared.
3. The follow-up continues. If the lead can’t talk, Simple works out a better time and calls or texts back then, remembering the first conversation.
4. A rep steps in when it matters. Your reps see who the lead is, what they asked and what Simple has already done, so they can take over without starting from scratch and hand back when they’re done.

How does speed to lead connect marketing to sales?

Marketing pays to get a lead’s attention, and speed to lead decides what happens in the minutes after it works. Leads don’t always stay on one channel, either. Someone might click an ad, text a question that evening and call the next day, but you always want them to feel they’re speaking to the same person. Simple keeps one memory of the conversation across phone, text, email and web, so each response picks up where the last one left off.

What should you measure?

Start with time to first contact, measured from the moment a lead arrives to the first real conversation and broken out by lead source and by hour of day. That breakdown will show you which sources get answered fastest, which hours get answered slowest, and whether leads convert differently once they’re reached sooner. From there, follow what happens next: contact rate, appointments set, conversion and revenue per lead. Revenue per lead matters most, because it shows whether a faster response is turning into more sales.

How do you start with speed to lead?

Pick one lead source, such as web quote requests or an aggregator list, and run Simple on it alongside your current process. Before it starts, agree on what counts as the first real conversation and when the clock starts, so both sides are measured the same way. Then compare time to first contact, appointments set and revenue per lead against the same source before, and run it long enough to cover evenings and weekends, when leads are most likely to be left waiting. Once the first response is fast, the next question is what happens to leads who don’t answer or aren’t ready yet.

The takeaway

Speed to lead comes down to being the first brand a new lead talks to, which for phone, text and web leads means answering in seconds, at any hour, on the channel they used. We run that first response for consumer brands like Renewal by Andersen, and a single lead source is enough to start comparing. Book a demo to talk to our team about yours.