Learn

How voice AI triggers actions after a conversation

A finished call runs the post-call workflows on that agent. Each one walks an ordered list of steps: filter on an analyzer result, then send the email, write your system, or run a browser automation. The actions the customer has to hear happen earlier, while they’re still on the line.

14 min read

  • Zach KamranCo-Founder & CTO
In this guide

Voice AI does the work after a conversation in a post-call workflow. The workflow is an ordered list of steps on the agent. When the call ends, Simple walks those steps: a filter on what an analyzer found, then the email, the write to your system, or a browser automation. The actions the customer has to hear happen earlier, while they’re still on the line.

Mixing the two up is how a call feels finished and the work never lands. A prompt that says “update the CRM when you’re done” is not an action. The agent will often say that it did. An action is a request your own systems receive and answer. Tool calling is that request during the call. After-call work is everything that starts because the call ended.

Two moments, two kinds of action

While they’re on the lineAfter they hang up
Who decidesThe agent, from what the customer just saidThe workflow you wrote, from analyzer results on the finished call
What it can doLook up an order, check a slot, book it, take a payment, transferWrite the record, send an email, run a browser automation, score the call
What the customer hearsThe result, in the same conversationNothing on the line. The workflow can send an email
If it failsThe agent says so, and doesn’t claim it workedThe run stops. You retry the failed step, and steps that succeeded stay done

A job in the wrong column fails in a familiar way. Booking after the customer has hung up means they leave believing they’re set, and the slot may already be gone. Writing the CRM note while they’re still talking makes them wait in silence for work they never needed to hear.

During the call, the agent calls a tool

Function calling is how a model asks for a specific action instead of describing one. You give the agent a short list of tools. Each has a plain description of when to use it, and the fields it has to fill in. “Look up an order when the customer gives an order number” is a tool. “Book this time, and only after the customer has confirmed the address” is another.

On the call it goes like this. The customer wants Thursday morning. The agent picks the booking tool, fills in the day and the address it already collected, and your system does the booking. The tool answers with the time that was booked, the times that are open, or an error. The agent says that answer. It does not invent a confirmation the tool never returned.

The description does most of the work. A vague one, such as “use this for orders,” gets used at the wrong moment, or never. A precise one gets used when the customer is actually asking. Keep each answer short. The agent has to say it out loud, and a dump of every field on the record makes the line go quiet.

Tools fail. The calendar times out, the card is declined, the order number doesn’t match. The agent should tell the customer what happened and take the path you allowed, which might be another attempt or a handoff to a person. Guardrails keep it inside that list. It can call only the tools you attached, and only with the fields you defined.

After the call, the conversation is an event

Hang-up is when the paperwork used to start. A rep would write the note, set a disposition, send the confirmation and flag the callback. None of that needs the customer on the line. It needs the finished conversation.

The platform marks the end with an event. The event names the call and says it finished. It does not carry the transcript, the recording or the outcome. Those live on the call, and the system that receives the event goes and reads them. Treating the event itself as the record is how a follow-up goes out with a blank summary.

The same event can arrive twice, and not always in order. A delivery that failed is sent again, so one finished call can show up again later. Anything you do from it has to be safe to run twice: one CRM activity, one confirmation, one task. When the second delivery finds the work already done, it stops.

Read the call, then decide

The event is the knock on the door. The decision uses what the call contains once you fetch it.

  • The transcript and the recording, so a person can check what was actually said.
  • The outcome: answered, voicemail, or never connected. A voicemail is not a booking.
  • Structured fields pulled from the conversation after it ends. On Simple these are analyzers. Each one asks a specific question and returns a specific kind of answer: yes or no, a number, one choice from a list, or a set of fields such as the customer’s name and whether the issue was resolved. A yes or no is something a workflow can branch on. A paragraph is not.
  • Tags, short labels such as “follow-up needed” or “billing,” applied from a description of when each one fits. A person can add or remove them afterwards.
  • Sentiment, a read on the caller’s tone across the call.

When you change the question an analyzer asks, or the description on a tag, you can run it again over older calls. A new field gets filled in for last month without anyone re-listening.

Post-call workflows and actions

A post-call workflow is the ordered work after the call, attached to the agent. Simple runs every workflow you turned on for that agent. The agent does not pick the steps. You wrote them, and the analyzer results on the finished call decide which ones run.

The steps run in order. These are the ones.

  • A filter keeps going only when an analyzer result matches. If it doesn’t, the workflow stops there.
  • A branch picks a lane from an analyzer result. The first match runs its steps, and the workflow continues with whatever you put after the branch. If nothing matches, it runs the fallback you wrote, or it simply continues.
  • Send an email, with the call and the analyzer results filled into the message. Simple sends that email.
  • Run an action. That is an HTTP call, or your own code, with values from the call filled in. This is the CRM note, the ticket, the update to your system.
  • Run a browser automation you already set up, filled with those same values.

If a step uses an analyzer, the workflow waits until that result is in. It does not send a half-read call. A step that fails stops the run. You retry it. Steps that already succeeded are left as they are, and a step that timed out is not sent again on its own, because the other system may already have done the work.

There is also one action on the agent that runs as the call is ending, separate from the workflow. It runs when the call ends, and not when the call was transferred to a person. If the agent is the one hanging up, that action finishes before the line drops. It can read why the call ended. It does the job and stops. There is no conversation left to steer.

One call, from hang-up to the next step

  1. The customer hangs up, or the agent ends the call because the job is done.
  2. The recording and the transcript are stored with the call.
  3. Analyzers read the finished call and store their results on it.
  4. The post-call workflows on that agent run. They wait for the analyzer results they use, then walk their steps. A filter that doesn’t match stops that workflow.
  5. A completion event can also go to your own system. It carries the call’s id and the event name, and it’s signed. Your system fetches the call. That path is for work you run yourself. The workflow above is the one Simple runs.
  6. Follow-up that spans days, the day-before reminder and the “we missed you” text, is a journey. It runs in the customer’s hours and stops when the goal is met. A journey can attach its own post-call workflow, and that one runs after each call the journey places.

Accept the event, put the work on a queue, and answer. Heavy work inside the request is how deliveries time out and get sent again, which is how the CRM ends up with two notes for one call.

Which job goes where

JobWhenWhy
Check a live order, price or slotDuring the callThe customer asked, and the answer has to be true right now
Book, change or cancelDuring the callThey should hear the time before they hang up
Take paymentDuring the callA decline has to be handled with them still there
Transfer to a personDuring the callThey’re waiting, and the rep needs the summary as the call arrives
Write the CRM note and dispositionAfter, as an action in the workflowThis is a record, not part of the conversation
Send the confirmation emailAfter, in the workflowSimple sends it, filled in from the call
Open a task or alert the ownerAfterIt depends on how the whole call turned out
Score the callAfterScoring needs the full conversation
Remind them the day beforeLater, as a journeyIt isn’t this call’s job

A confirmation the customer asked to hear, “can you text me that?”, can go out as the call is ending. The note, the score and the reminder sequence still wait until the call is over.

How Simple does it

During the call, each tool is a request the agent can make. That might be an HTTP call to your API, a connection to a system such as your CRM, or your own logic that combines several lookups and applies your rules before it answers. You write when to use it. The agent fills in the fields from the conversation, your system does the work, and the agent says what came back. If the request fails, the conversation continues, and silence is not treated as success.

When the call ends, the agent can run one action of its own, separate from any workflow. It runs on a normal ending, and not when the call was transferred. If the agent hangs up, that action finishes before the line drops. It writes or updates. It does not speak.

The post-call workflow runs once the call is over. You build it in Workflows, on that agent: a filter or a branch on an analyzer result, then an email Simple sends, an action that calls your API or your own code, or a browser automation. Every workflow turned on for the agent runs. A step that needs an analyzer waits for the result. A failed step stops that run, and a retry skips steps that already succeeded.

If the work should start in a system you run, Simple also sends call.completed to the webhook you configured. The body is the call’s id and the event name. Your system fetches the call for the recording, the transcript, the analyzer results, the tags, and the sentiment. That can be your own endpoint, or a tool such as Zapier. The API and webhooks page is the short version of that connection. The same event can arrive twice, so the handler has to notice when its job is already done.

Before the first word, an optional pre-call step can look the caller up and hand the agent their name, an open order, and where the last conversation left off. It has a few seconds. If it doesn’t answer, the call still starts.

Journeys are the days after this call: a text, then a call if they don’t reply, in the customer’s local hours, stopping when the appointment is confirmed or they opt out. A journey can attach a post-call workflow, and that workflow runs after each call the journey places, with the step and the result. Develop scores the call afterwards against your checklist. That score is after-call work for the team, not something the customer waits on.

How to tell it’s working

  • Book a real slot on a test call, hang up, and check the calendar has that appointment once. Then make the booking system fail, and check the agent says so instead of “you’re all set.”
  • Deliver the same completion event twice, if you receive it in your own system. You should still have one CRM activity.
  • Set the analyzer result the filter checks to no. That workflow should stop, and the email after the filter should not send.
  • Fail the action on a test call. The run stops. Retry it, and a step that already succeeded does not run again.
  • Read the note a rep would see. It should have the outcome and the next step, and it should match the transcript.
  • Change a tag’s description and run it again on last week’s calls. The ones that fit should pick it up. The ones that don’t should stay clear.

If the agent books on calls where the customer only asked a question, the tool’s description is too loose. If the CRM is empty the next morning, look at the workflow’s run. A filter may have stopped it, or the action failed and nobody retried it.

Frequently asked questions

How does voice AI trigger an action after a call?

When the call ends, Simple runs the post-call workflows on that agent. Each workflow walks its steps in order. A filter or a branch looks at an analyzer result, and the steps that continue send an email, call your system, or run a browser automation.

What’s the difference between a tool call and a post-call action?

A tool call happens while the customer is on the line. The agent asks your system for something and speaks the result. A post-call action runs after the call, inside a workflow, and it does the job without anyone left to tell. The agent can also have one action that runs as the call ends, except when the call was transferred.

What is a post-call workflow?

It is an ordered list of steps Simple runs after a call on that agent. A filter stops the run unless an analyzer result matches. A branch runs one lane or another from those results. The other steps send an email, call an HTTP action or your own code, or run a browser automation.

Does the agent update the CRM during the call or after?

Both happen, for different jobs. Booking a time the customer has to hear can write the record during the call, through a tool. The note and the activity on the contact usually run after the call, as an action in the post-call workflow, so the rep who follows up starts with the full conversation.

What if the action fails?

During the call, the agent should say it failed and follow the path you allowed, which may be another attempt or a transfer. After the call, a failed step stops the workflow. You retry that run, and steps that already succeeded are not run again. If your own system receives the completion event, a failed delivery can arrive again, so that handler has to recognize work it already did.

Can one call start several actions?

Yes. Every post-call workflow turned on for the agent runs. A booked call can write the CRM and send the confirmation email. A journey can keep going over the following days, and its own workflow runs after each call it places. A call whose analyzer result doesn’t match the filter stops there.

Hear a voice agent handle a real call