04 · AI agent
It cites its sources. Or it declines.
You have been burned by a confident chatbot. The behaviour that matters is not how much it answers — it is what it does when it does not know.
How it works
Retrieve, cite, or hand off — and log what was missing.
- 01Retrieve from your contentGuides, articles, crawled pages and uploaded documents, scoped to the audience the asker belongs to.
- 02Answer with citationsEvery claim names the source it came from, linked, so the reader can check the answer instead of trusting it.
- 03Below threshold, declineHand off to a human with the full transcript, the retrieval set, and what the agent already tried. No confident guessing.
- 04Log the gapThe miss becomes a ranked knowledge gap rather than a shrug — which is where the loop picks it up.
In the product
A grounded answer, and an honest refusal
The same conversation showing both behaviours: a cited answer, and the handoff that happens when confidence drops below the threshold. The refusal is the feature.
agent-answer-cited.pngAn agent answer with two cited sources, and the threshold handoff beneath it
The detail
Why we think you can trust this one.
Grounding policyAnswers must cite retrieved sources. Ungrounded generation is refused rather than softened into a hedge — the agent does not have a mode where it improvises.
The eval harness gates changesPrompt and model changes are regression-tested on groundedness, refusal and escalation. A change that makes the agent more confident and less correct does not ship.
Red team in CIPrompt-injection and cross-audience leak probes run on every build, with a nightly scan. "Ignore your instructions" is a test case, not a surprise.
Credit metering, visibleAbout 5.5¢ per answered conversation, 1¢ per copilot draft, half a cent per summary — with a live balance and a hard auto-recharge cooldown.
Tool calls are permissionedEvery tool the agent can call carries a permission rule, argument constraints, and an audit trail. It cannot do something you did not allow.
Escalation with contextThe human who takes over sees the retrieval set, the confidence score, and the reason for the handoff — not just the transcript.
Where this sits
The agent is the last forward stage — and the source of both return paths.
01Capture
02Knowledge
03Engage
04AI agent
Capture flows to Knowledge, to Engage, to the AI agent; a miss becomes a work order and a resolution becomes a draft, feeding back into Knowledge.
Honest limits
It is a support agent, not an assistant.
You will find this section on every feature page. If a constraint would annoy you in week two, we would rather you knew in minute one.
- Ask it about the weather and it will decline. It answers from your knowledge base and nothing else, which is a deliberate ceiling on what it will attempt.
- It cannot answer what you have not documented. That is the point of the loop: it tells you what is missing rather than inventing a plausible step.
- Deflection rate depends entirely on your content. We report yours, measured before and after. We will not quote somebody else’s number at you.
Next
Executable guidesStop describing the steps. Show them.Four tiers of helpfulness instead of one. A cited answer, a guide card, an on-screen walkthrough on the customer’s real page — or, with consent, the agent doing it for them.Self-healing loopDocumentation doesn’t rot because people are lazy. It rots because nothing tells you it’s wrong.Two return paths nobody else has. A miss becomes a work order. A resolution becomes a draft. Usage improves the asset instead of wearing it out.