The assistant reads your docs. AI Actions let it read your APIs too.

Register your own endpoints — order lookups, subscription changes, refund issuance, status checks — and BeforeQuery calls them from chat when a question needs a real answer, not a documented one. Every action runs under an approval rule you define, so nothing sensitive ships without a human in the loop.

Start FreeAgent Platform

What AI Actions do

Turn 'what's my order status?' into a real answer — with an audit trail and approval gates for anything that changes state

Your APIs as tools

Callable Functions

  • Register REST or webhook endpoints as callable actions
  • JSON schema per action defines inputs and outputs the model must fill
  • Actions can be read (lookup / status) or write (refund / update)
  • Called only when the retrieval + reasoning path decides they're relevant
  • Every invocation logged with inputs, outputs, and the message it answered
Human in the loop, by policy

Approval Rules

  • Auto-approve rules for safe read-only actions (status, lookup)
  • Human-approval queue for write actions (refund, subscription change)
  • Per-action or per-KB approval rules; policy travels with the action
  • Approve, reject, or edit-then-approve from the dashboard queue
  • SLA on pending approvals reported in analytics
Every action recorded

Safety & Audit

  • Every invocation is a trace — inputs, retrieved context, outputs, and the deciding turn
  • PII masking still runs before the model sees anything
  • Actions inherit workspace and per-KB permissions
  • Rate limits per action to bound blast radius
  • Revoke an action to cut off any future call instantly
The difference between an answer and a resolution

'The status is delivered' beats 'here's how to check'

Docs can tell a customer how to look up their order. An action can look it up for them. The first is a Q&A assistant; the second is a support agent that saves the ticket. AI Actions turns your APIs into tools the assistant can reach for — with an approval policy for anything that mutates state, so the safety story stays intact.

  • Answer questions from your APIs, not just your docs
  • Approval rules distinguish 'safe to run' from 'needs a human'
  • Every action call is auditable via traces and the invocation log
  • Works inside chat, Support Ticket AI drafts, and agent workflows
  • Enterprise plans add SLA on approval queues and dedicated review roles
What teams say

Deployed in production, cited by the buyers who chose it

“Deployed on our docs site in an afternoon. Every answer has citations, and the abstention gate means we've never had a customer complain about a made-up answer.”
PN
Priya Nair
Head of Customer Support · Supabase
“The knowledge base connected to our Slack, Confluence, and helpdesk in one setup. On-call teams get the same cited answer whether they ask in chat, in the widget, or from Cursor.”
TR
Tom Richter
IT Operations Manager · Grafana Labs
“The gap analytics turned into a real docs backlog. Deflection went up because we finally knew which pages were missing — the AI told us.”
AC
Ana Castillo
VP of Customer Experience · Clerk

Frequently Asked Questions

Common questions about AI Actions

Read: order status, account lookup, subscription details, invoice fetch, feature-flag state. Write: password reset, plan change, refund issuance, ticket reassignment, order cancellation. Any REST endpoint your team maintains is a candidate — you describe it with JSON schema, the model calls it when the retrieval path decides it's the right tool.
Each action carries an approval rule. Read-only actions (status, lookup) can auto-approve. Write actions default to requiring human approval; sensitive invocations land in a review queue in the dashboard, where an agent approves, rejects, or edits before the call fires. The rule is per-action or per-knowledge-base — the policy travels with the action, not the caller.
Two guards: retrieval decides whether an action is even considered (it has to be indexed as relevant to the question), and the invocation records the retrieved context and reasoning trace so a wrong call is auditable and correctable. In practice you tune this the same way you tune retrieval — start with a small set of actions, watch the trace log, expand.
Function calling is the model-side mechanism BeforeQuery uses under the hood. AI Actions is the platform on top: the schema registry, the approval queues, the invocation traces, the analytics, the per-KB scoping — the parts that make function calls deployable in production instead of a demo.
Yes — through the Agent Platform. Actions are the units; agent graphs orchestrate coordinator → specialist → action sequences with human-approval steps between them. Simple 'answer + one action' cases don't need the agent platform; complex multi-step resolutions do.
Each end-user question that reaches the model counts as one AI question, whether or not it triggers an action. Actions themselves are not separately metered on Free or Pro; Enterprise pricing may include separate action-volume tiers.

Give the assistant your APIs, not just your docs

Register your first action in the dashboard and set an approval rule. The assistant reaches for it the moment retrieval decides it's the right tool.

Get Started Free