Browse documentation
Support agent and approved knowledge
Review grounded drafts, publish knowledge and control a narrowly bounded automatic mode.
Configure the app first
Open Settings → Support agent for the selected app. Assistance starts disabled and requires a server-configured provider, explicit workspace allowlist, verified spend budget and app enablement. The direct OpenAI adapter uses the server’s OPENAI_API_KEY and OPENAI_MODEL; an optional gateway adapter has separate configuration. Never copy keys into the SDK or setup prompt. Human replies remain available when AI is unavailable. The approved five-request synthetic canary has consumed its full budget; the agent and automatic replies were disabled afterward. Its bounded results are described below, not treated as general quality or production readiness.
Draft and summary are human tools
Request a reply draft or summary from a conversation using its current revision. Drafts require published app knowledge; summaries cite public message references. Bounded context excludes private notes, other customers, images and arbitrary fetched URLs. Approved metadata keys are explicit. Source quotes are checked against the captured source and immutable knowledge revision. Treat model confidence as a model claim, not measured accuracy. Review or edit a draft before inserting/sending it; generating, inserting or viewing an assist never sends a customer reply.
Publish knowledge deliberately
Owners/admins can keep up to 20 bounded documents, save draft revisions and publish immutable versions with an optional expiry. An unpublished edit does not silently replace the published source. Deletion, publication, expiry or settings changes fence a result generated against old knowledge. Missing sources, conflicting policy, unsupported financial statements and changed conversations require review or escalation. Never put credentials, private customer content or instructions to bypass permissions in knowledge.
Automatic mode is deliberately narrow
Only an owner may enable it, with an authorized sender and an eligible published navigation answer. The answer must be the complete immutable document, at most 500 characters/1,000 bytes, using the canonical form below. The model may select that whole answer with a valid citation; it cannot remove a negation or condition, combine excerpts or invent wording. General prose in any language is human-review only. This mode sends labelled AI assistant text; it never opens a screen, invokes a paywall, requests a review, buys, restores, refunds or changes an entitlement.
Open “Settings” and choose “Reminder time”.Handoff, quotas and uncertain results
Newer activity, assignment, handoff, blocked/deleted customers, revoked sender access, changed knowledge/settings and expired reactive events stop an old automatic job. A human can take over; resuming requires admin authority. Existing Needs reply work is preserved. Durable leases and operation identities prevent blind repeated generation or sending after an uncertain outcome. Workspace allowance and explicit provider daily/total call budgets apply atomically. One attempt uses bounded text-only structured output without model tools, browsing, images or automatic transport retries. Provider usage/model records are diagnostic evidence, not purchase attribution.
Verify in development
Use synthetic questions with known answers, unsupported claims and handoff cases. Inspect citations, current thread revision, author labels, usage and exact customer-facing text. A local injected model or passing authorization test is not live model-quality proof. The recorded hosted OpenAI canary used five synthetic requests: one grounded human-review draft, three summaries escalated for review, and one successful exact approved navigation reply. The canonical automatic reply was fetched through scoped customer HTTP and Needs reply was preserved; this is not native display proof. Summary verification remains incomplete. The agent was disabled and no request budget remains. Production entitlement, broader model evaluation, the owner’s own installed app and physical devices remain separate release gates.