From answers to action

A support agent trained only on your documentation can tell a customer what your refund policy is. It cannot tell them whether *their* refund has been processed. That second question is the one people actually ask, and it is the one that ends in a support ticket.

An action closes that gap. You describe one endpoint you already own, the agent works out when a question needs it, extracts the parameters from what the customer said, calls it, and answers from the response. The agent is not guessing. It is reading your system, at the moment it is asked.

This is what separates deflection from automation. A chatbot that deflects sends people to a help article. A chatbot with actions resolves the thing they came for and closes the conversation.

Most platforms give you a catalogue of integrations and you get what is in the catalogue. Actions invert that. If your system has an HTTP API, the agent can call it, whether that is a Rails monolith, an internal microservice, a legacy PHP endpoint, or a third-party product nobody has heard of.

You get all five methods (GET, POST, PUT, PATCH and DELETE), custom headers, and typed parameters including nested objects. Bodies can be encoded as JSON, form fields or multipart, which matters more than it sounds: plenty of older systems still expect form encoding and will reject a JSON body outright.

Authentication is configured on the action, not on the agent. Bearer token, API key, basic auth, or nothing at all. One agent can hold a read-only key for your orders service and a separate, narrower key for your billing service, and neither can be used to reach the other.

Not everything needs to be wired by hand. Connect Calendly and your event types are pulled in as a dropdown, so booking a meeting from inside a conversation is a few clicks rather than an API contract. Connect HubSpot and your real pipelines, ticket stages and priorities appear the same way.

The custom-action builder and the pre-built ones are the same machinery, so you can start with a connected tool and drop down to a raw endpoint the moment you outgrow it. See the full list on the integrations page.

Actions are included from the Basic plan upward, with the number of actions per agent rising by tier and becoming unlimited on Business. See pricing for the current limits.

Two things are worth thinking about before you expose a write endpoint. First, scope the credential: give the action the narrowest key that lets it do its job, because an agent with a broad admin token is an agent that can do broad admin things. Second, if the action can act on a specific customer record, pair it with verified user identity so the agent knows the person asking is genuinely that customer, rather than someone who typed their email address into a chat box.