Train your knowledge

Building Procedures the agent can run safely

Procedures are the actions your Customer Agent and Team Assistant can take on behalf of a customer during a conversation. They turn a question-answering bot into an agent that books meetings, redirects to the right page, escalates to live chat, or runs an API call. The key to using them well is understanding how to build them so they stay within the boundaries you control.

The platform gives you seven procedure types out of the box: suggest a follow-up question, redirect to a URL, start a component, switch topic, hand off to live chat, make an API call, and run custom JavaScript. Each one is a defined move the agent can offer, and the customer always picks whether to run it. That keeps a human in the loop and prevents the agent from acting unprompted.

How Procedures are structured

Every procedure has three parts you configure: variables, triggers, and an optional escalation flag.

Variables are the inputs a procedure needs to run. If the agent does not already have a variable, it asks the customer for it. Variables can depend on previous answers, so you can collect an email address or a product name before the procedure fires.

Triggers define where and when a procedure shows up. You can link a trigger to a specific Source, fire it on negative feedback, after an unanswered question, or set it to always appear. This is how you control relevance: a pricing redirect only shows when the customer is asking about pricing.

The escalation flag marks a procedure as escalating the conversation. This is used in reports to track how often the agent reached its limit, which is useful for spotting where human hand-off is needed most.

Building and managing procedures

You create procedures in the dashboard under Train > Procedures. Click Create procedure, pick the type, give it a name, configure the variables, and set the triggers. The workflow is designed for business users, with no code and no engineering tickets required.

From the Procedures section you can also link a procedure to one or more Sources or FAQs, set trigger conditions, mark a procedure as an escalation, and test it in Preview before it goes live. Testing before shipping is important because you can simulate agentic skills against synthetic users and see how the procedure behaves.

Safety and permissions

The platform enforces one hard boundary: the agent never takes an irreversible action without a human approving it first. You decide which procedures count as irreversible during setup. Every procedure carries explicit permissions that define what the agent can do alone, what needs a human, and what is off limits.

This permission model extends beyond procedures. Every expansion action, for example, carries a permission boundary, and the agent runs only the procedures you have defined. Unless enforces compliance guardrails throughout, and every action is logged with a full audit trail. There is no autopilot.

Common use cases

Procedures are useful in several recurring scenarios:

When customers ask about pricing, redirect them to the pricing page
When customers need to book a call, let the agent handle scheduling
When the agent cannot resolve an issue, escalate to a human with full context
When you want to collect an email address, add a required variable to the procedure

You can also limit a procedure to certain customers by combining it with an Audience, so it only fires when the customer is in that group.

Measuring success

To see if a procedure is working, open Analyze \u003e Conversations and filter for conversations where the procedure ran. Check completion rates and customer reactions. This tells you whether the procedure is actually helping or just adding friction.

Conclusion

Procedures give your agent real capability while keeping the customer in control. The safety model is simple: you define the moves, the agent offers them, the customer decides, and nothing irreversible happens without approval. Build them in Train, test them in Preview, and watch the audit trail to keep them sharp.