Skip to main content
When a message comes in, the AI agent looks at your active Procedures, decides whether one applies, and if so, follows its steps. This page explains the rules that govern that decision so your Procedures behave predictably. Most surprises with Procedures come from these rules: a Procedure that is skipped on one channel, a flow that switches partway through, or an Action that never fires on its own. Knowing how a turn runs makes those cases easy to diagnose.

​
What happens on one turn

Here is one turn of a conversation on WhatsApp, where you have a refund Procedure that references @Get_Orders and @Create_Return:
1

The customer sends a message

The customer writes “I’d like a refund for my last order.”
2

The AI agent filters the Procedures it can use

Only Active Procedures are considered. Any Procedure that references an Action that is unavailable on WhatsApp is skipped, so the refund Procedure stays in the running only if both Get_Orders and Create_Return are on for WhatsApp.
3

The AI agent picks at most one Procedure

The refund Procedure’s trigger matches the request best, so the AI agent engages it. If no Procedure applies, the AI agent replies without one.
4

The AI agent works through the steps in order

It runs each step top to bottom. At a branch step, it evaluates the conditions from the top, runs the first one that matches, then continues after the branch. When a step references an Action, the AI agent calls it, including Actions set to Only use in Procedures.
5

The AI agent pauses when it needs the customer

If a step needs something from the customer, such as which items to return, the AI agent asks and stops. On the next turn, the refund Procedure stays in flight and picks up where it left off, unless the customer clearly switches topics.

​
Only one Procedure runs per turn

Within a single turn, the AI agent runs at most one Procedure: the one whose trigger best matches what the customer is asking for right now. It never executes two Procedures in the same reply. This is a per-turn limit, not a lock. The active Procedure can change from one message to the next: if the customer is partway through Procedure A and then sends a message that matches Procedure B’s trigger, the AI agent can switch to Procedure B on that turn.
A Procedure stays in flight across turns while the customer is mid-flow (for example, replying “yes” or providing a value the AI agent asked for). It ends when the situation is handled, or when the customer clearly switches topics, at which point a different Procedure can engage on the next turn.

​
Procedures are dropped when an Action isn’t available

A Procedure can reference Actions (with @action_name) in its steps. Before a Procedure is even considered for a conversation, the AI agent checks that every Action it references is available on the current channel. If even one referenced Action is unavailable, the whole Procedure is skipped for that conversation. An Action counts as unavailable when it is:
  • Disabled or removed,
  • Disabled on the current channel, or
  • Not supported on the current channel/medium (for example, a Widget-based Action on a channel that can’t render Widgets).
The check is all-or-nothing and covers steps inside conditional branches too, even branches that might never run on a given conversation. If a deep branch references an Action that’s off on, say, WhatsApp, the entire Procedure is dropped on WhatsApp. This is intentional: a half-runnable Procedure is worse than none.
What this means for you: if a Procedure works in one channel but seems to be ignored in another, check that all of its referenced Actions are enabled on that channel.

​
Procedure-only Actions

By default, the AI agent can call any enabled Action on its own whenever the Action’s When to use matches. Sometimes you want the opposite: an Action the AI agent should never reach for by itself, and only ever run as a deliberate step inside a Procedure, for example a “process refund” Action you only want fired after the refund Procedure’s checks have passed. Turn on Only use in Procedures in the Action’s settings to enforce that. The Action stays enabled and fully functional, but:
  • The AI agent won’t select it on its own judgment, so its When to use field is hidden (it no longer needs one).
  • It runs only when a Procedure step references it with @action_name.
This keeps sensitive tools on rails: they can’t fire outside the flow you designed. See Actions for where to find the setting.

​
Waiting on the customer

If a step calls an Action that needs user input (an account picker, a form, a confirmation), the AI agent treats it as a pause, not a failure. It tells the customer it’s waiting, then stops. It resumes the same step once the customer responds. It won’t retry or jump to another Action.

​
Troubleshooting

Confirm it’s Active, and that its trigger describes the customer’s intent clearly and distinctly from other Procedures. Overlapping triggers can cause a different Procedure to win.
A referenced Action is likely unavailable on that channel, which drops the whole Procedure there. Enable every referenced Action on that channel.
Only one Procedure runs per turn. Combine the flow into a single Procedure (using branches) rather than expecting two Procedures to run together.
An Action with Only use in Procedures on runs only when a step references it with @action_name. Check that a Procedure references it, that the Procedure is Active, and that every Action in that Procedure is available on the channel.
The customer’s latest message matched another Procedure’s trigger better, or they changed topics. Make each trigger describe a distinct situation, so triggers don’t overlap.
A step called an Action that needs the customer’s input, such as a form or a confirmation. This is a pause, not a failure: the AI agent resumes the same step once the customer responds.
A step references an Action that doesn’t exist or is disabled. The editor flags the reference. Fix or remove it, then enable the Procedure.

​
Frequently asked questions

Yes. A Procedure isn’t a lock. If the customer clearly switches topics, or sends a message that matches another Procedure’s trigger, the AI agent can move to that Procedure on the next turn.
Yes, by default. The AI agent can also call the Action on its own whenever its When to use matches. Turn on Only use in Procedures to stop that.
No. With Only use in Procedures on, the AI agent never decides when to call the Action, so its When to use field is hidden. The Procedure step that references the Action decides when it runs.