Skip to main content
An agentic workflow turns one long prompt into a map of steps that the call moves through while it is happening. Each step does one job, such as talking to the caller, capturing details, checking their identity, or transferring them, and the arrows between steps decide where each caller goes next. You build it in your agent’s Workflow tab.

​
What it is

An agentic workflow is a flowchart your agent follows during a live call. It starts at Start, where your agent says its greeting, and moves from step to step along arrows called paths. Every step and every path runs while the caller is on the line. Nothing in the workflow runs after the call ends. For work after the call, see Automate what happens after a call. A single prompt is enough for most agents. A workflow earns its place when different callers need clearly different handling and you want that routing to be predictable, not left to the agent’s judgement in one long set of instructions.

​
When you would use it

  • A front desk that sorts callers. New patients go to intake, existing patients asking about billing go to the billing team, and everyone else keeps talking to the front desk.
  • A check before private information. Callers asking about their account must pass a texted code before the next step reads anything out.
  • A transfer that must happen on a rule. Every caller who says it is urgent goes straight to a person, every time.

​
How it works

The workflow builder offers these steps. You add one by hovering over a step and choosing +, then picking from the list. The table lists each step and what it does on the call. Extract Information is not the same as the extraction fields in Post Call Analysis. Extract Information fills values at the moment the call reaches it, so the call can be routed on them. Post Call Analysis fields are filled after the call ends, for reporting. See Pull the details out of every call.

​
How paths decide where the call goes

A path is the arrow between two steps. Choose a path to open its Edge Condition panel, then answer When should the call take this path? with one of three choices:
  • When the AI decides. You describe the situation in words, and the agent judges it from the conversation. Good for intent, such as “The caller wants to book an appointment”.
  • When a condition matches. You compare a value against something: What to check, How to compare (for example equals, is one of, is set), and What to compare against. Choose Add condition for more than one check, and Match with All of them or Any one of them. The same values always give the same answer.
  • Otherwise. Taken when nothing else leaving the step matched. There is nothing to fill in.
One step can have as many paths leaving it as you need. After an Extract Information step, Brilo checks the paths in this order: rule paths against the values it captured, then the paths left to the AI, then Otherwise. When two rule paths could both match, you cannot choose which wins, so write rules that do not overlap. Rule paths leaving a Sub-Agent are checked when the caller arrives at that Sub-Agent, not again later. To route on something the caller says during a conversation, send the call through an Extract Information step and put the rules on the paths leaving it. A new path shows red until you save a condition on it, and you cannot publish your agent while any step or path is red.

​
When nothing matches

A step with no Otherwise path can reach a point where none of its paths fit. When that happens after Extract Information, Send SMS or Tool Calling, Brilo sends the caller back to the last Sub-Agent they spoke to, which asks one short clarifying question and then tries again. If the second try still matches nothing, the agent apologises, tells the caller it cannot help on this call, and ends the call. The same happens when a path left to the AI is chosen but the Extract Information step captured nothing it was meant to: the caller is asked for the missing detail once. Give every Extract Information step an Otherwise path. It is the simplest way to make sure no caller ever reaches a dead end. Point it at a Sub-Agent that carries on helping, or at a Call Transfer to your team. A step can only be revisited a limited number of times on one call. A path that loops a caller back to an earlier Sub-Agent again and again ends the call politely, so send callers forward to a new Sub-Agent rather than in a circle.

​
The Verify Identity step

Verify Identity confirms a caller is who they say they are before a later step reads out personal or account information. Under Setup, How is SMS, and Send the code to is one of:
  • Contact record, the phone number saved on the contact matched to this caller, or another contact field you pick.
  • Use dynamic variable, a phone number an earlier step filled.
  • The number they’re calling from. Only safe when the contact was matched from that same number.
Advanced sets Code expires after, Tries allowed, Resends allowed, and Let them use the keypad. What the caller hears holds the lines your agent says during the check. Verify Identity has two exits instead of conditions. Verified is for callers who gave the right code. Not verified is for everyone else, including callers who ran out of tries or had no number to text. Both must lead somewhere, such as a transfer or a polite ending, before you can publish.

​
A worked example for a clinic front desk

A clinic front desk workflow sorts callers by why they are calling, while the front desk conversation carries on for everyone the rules do not catch.
  • Start leads to a Sub-Agent called Front desk. Its Conversation goal asks the caller whether they are a new or existing patient and why they are calling.
  • The path from Front desk is set to When the AI decides, with “The caller has said whether they are a new or existing patient and why they are calling”. It leads to Extract Information.
  • Extract Information captures caller_status as a Choice of new or existing, and call_reason as a Choice of appointment, prescription refill, billing, or other.
  • The billing path is When a condition matches, with call_reason equals billing. It leads to a Call Transfer to your billing team.
  • The refill path is When a condition matches, with call_reason equals prescription refill and caller_status equals existing, and Match set to All of them. It leads to Verify Identity. Verified leads to a Sub-Agent that handles refills, and Not verified leads to a Call Transfer to the front desk.
  • The Otherwise path leads to a Sub-Agent called Front desk, continued, with the same goal of helping the caller, so the conversation carries on for everyone else.
The values you compare against must match the Allowed values of the Choice exactly as you typed them.

​
When not to use one

  • Your agent answers questions and takes messages. A single prompt does this well and is easier to change. See Write what your agent says.
  • You only need a transfer when a caller asks for a person. A transfer action does this without a workflow. See Pass the call to a person.
  • The work can wait until the call ends. Use a post-call workflow instead.

​
What it costs and what it needs

  • The workflow builder needs a desktop screen. It is not available at phone width.
  • Extract Information needs an agent with versions turned on. See Save and switch agent versions.
  • The workflow is off until you choose Enable Agentic Workflow in the Workflow tab. From then on your agent follows the workflow on every new call. The Agentic Workflow switch in the same tab turns it off again.
  • On a workspace with HIPAA support turned on, Extract Information does not capture values, so rule paths that check them will not work as intended. Use paths set to When the AI decides instead. See Healthcare and HIPAA.

​
What else you can do here

​
FAQ

Yes. Add as many paths leaving the Extract Information step as you need, each with its own condition, for example one checking caller_status and another checking call_reason. Brilo checks the rule paths first, then the paths left to the AI, then Otherwise. Write rules that do not overlap, because you cannot set which one wins.
The Otherwise path is taken, if the step has one. Without an Otherwise path, the caller goes back to the last Sub-Agent they spoke to for one clarifying question, then Brilo tries again. If nothing matches the second time, the agent apologises and ends the call. Add Otherwise to every step to avoid that.
During the call. Extract Information reads the conversation at the moment the call reaches it, fills its values, and the call moves straight on along the matching path. It never speaks to the caller. The extraction fields in Post Call Analysis are different, because those are filled after the call ends.
Point the Otherwise path, and any path that should not end or transfer the call, at a Sub-Agent. That Sub-Agent picks up the conversation where it left off. Send callers forward to a new Sub-Agent rather than back to the one they came from, because a step can only be revisited a limited number of times.
Put the rule on a path leaving Extract Information and point it straight at a Call Transfer or a Verify Identity step. For Verify Identity, connect Verified to the step that reads out private information and Not verified to a transfer or a polite ending. Add Otherwise for everyone else.
Keep the Extract Information step if you route on exact values, such as a reason chosen from a list. Remove it if the routing is about what the caller wants, because paths set to When the AI decides on the Sub-Agent can judge that directly. On a workspace with HIPAA support turned on, use paths left to the AI.
Extract Information only appears for agents with versions turned on. Open your agent and choose Enable Versioning at the top of the agent page, then return to the Workflow tab. See Save and switch agent versions for how versions work.
A prompt is one set of instructions your agent follows for the whole call. An agentic workflow splits the call into steps with paths between them, so routing follows rules you can see. Use a prompt unless different callers clearly need different handling. See Write what your agent says.