What Your AI Agent Can Actually Do Inside a Conversation
Chat's AI agent capabilities are a short, fixed, inspectable list — here's exactly what's on it and what happens when it needs a human.
"AI-powered" support tools tend to describe themselves in terms of what they feel like rather than what they do. Chat's AI agent capabilities are the opposite: a fixed, named list of things a Skill is allowed to do inside a live conversation. There's no vague "the AI handles it" — there's a closed set of capability IDs, and every Skill is built from that set.
The capability list
As shipped, a Skill's capabilities are drawn from exactly four options:
crm_lookup— look up a record in your CRM related to the conversation.route_to_team— hand the conversation to a specific team.tag_conversation— apply a tag for categorization or reporting.escalate_to_human— end the AI's involvement and bring in a person.
That's the whole list. There's no capability for "send a refund," "modify an order," or anything that changes state outside the conversation itself — the AI agent's actions are scoped to understanding and routing the conversation, not executing business transactions on your behalf.
Escalation is a real exit, not a fallback message
When a Skill sets escalated: true, it triggers a hand-off with a specific, visible reason attached to the conversation: "A Skill escalated this conversation to a person." This isn't the AI agent quietly giving up and producing a generic "let me connect you with an agent" line — it's a structured event that shows up in the conversation record, so an agent picking it up can see exactly why the AI stepped back rather than reconstructing that from the transcript.
Who can configure this
Writing to Skills — defining what capabilities are wired up and under what conditions — is restricted to workspace owners and admins. The reasoning is stated plainly in the product itself: a Skill decides what the AI does to a live customer conversation, which is not a setting to leave open to every team member. Agents interact with the results of Skills; only owners and admins configure which capabilities are active.
Why a short list is the point
A support tool that claims broad, unspecified AI autonomy is hard to evaluate and harder to trust with real customers. A four-item capability list is easy to audit: you can read it in under a minute and know exactly what the AI agent is and isn't allowed to touch in a conversation. If a capability you need isn't on the list — say, issuing a refund automatically — that's a real, current limitation, not a matter of the AI "deciding" not to do it. The honest answer is that it can't, by design, until that capability is added.