Your support AI can now act
Answering from articles is table stakes. Today Iris AI can look things up in your systems — orders, bookings, account state — through tools you define in a form or bring via MCP. Safely: verified identity, allowlisted responses, human-approved changes.
There are two kinds of support AI. The first kind answers questions from your help center — useful, and most vendors stop there. The second kind can tell a customer "your order shipped yesterday, here's the tracking link" — because it can actually look. Today Iris crossed that line. We call it Actions.
Tools in a form, or bring your own server
Most businesses define a tool in two minutes: a name, a sentence describing when to use it, and a request against their own API — "GET api.yourstore.com/orders?email={{email}}". Add an auth header (encrypted at rest, never shown again) and declare which response fields the AI may see. Done: your AI can now answer "where's my order?" with the truth instead of a shrug.
Teams with engineers can go further: connect an MCP server (the open protocol your systems may already speak) and Iris discovers its tools. Each one requires your explicit approval before the AI may touch it — and if the server ever changes a tool's definition, it's automatically pulled back for re-review.
The safety model is the feature
We run this exact architecture on our own support, at production scale, so the guardrails aren't theoretical:
Identity is never guessed. Values like the customer's email or user id are injected server-side from the verified identity of the person in the chat. The AI cannot look up someone else's account — structurally, not by prompt-politeness.
The AI reads, humans change. Read lookups run automatically. Anything that would change an account — a cancellation, a refund, an update — becomes a proposal card in your inbox. A teammate approves; only then does it execute. Every call, either way, lands in an audit trail you can read.
Responses are allowlisted. You declare exactly which fields the AI may see. It never receives your raw API responses, so it can't leak what it was never shown.
What this costs
Actions ships with paid plans — Growth includes three HTTP tools, Scale raises the limit and adds MCP servers. Tool-assisted answers use two AI credits instead of one, because an answer with your live data in it is worth more than a lookup.
Support AI that only deflects is a cost line. Support AI that resolves is a team member. Today Iris became the second kind.
Put this playbook to work.
Create a workspace, paste one snippet, publish a few articles. Free to start — live before your coffee cools.