Write Docs That Teach Your AI When Rules Bend


TL;DR
When a customer cites an article your AI has already indexed but your team has since invalidated, the agent will confidently repeat wrong information. The fix is a three-layer protocol: expire content before it spreads, detect the mismatch at handoff, and close the loop so the correction reaches the next customer too.
A customer opens a chat, pastes a paragraph from one of your help articles, and asks the AI agent to confirm it. The article is eight months old. Your pricing changed, your UI was redesigned, or that feature was sunset entirely. The agent — working from whatever it last indexed — validates what the customer said, and now you have a support ticket, a confused user, and a correction that has to happen twice: once for this customer and once in the knowledge base.
This is not a fringe case. Any team that ships frequently and publishes docs will hit it. The question is whether you have a protocol before it happens or a scramble after.
Chattering's agent answers from your indexed content and cites its source. That's the feature. The consequence is that if stale content stays in the index, the citation looks authoritative even when the underlying information is wrong. The agent has no way to know that the article it's reading was accurate in Q3 and is wrong in Q1 of the following year. It reads what you gave it.
This means the correction problem is entirely a content-ops problem, not a model problem. The model will do exactly what you configured it to do.
Every article you publish should carry a review date in its metadata, set at the time of writing. For fast-moving areas — pricing, integrations, UI flows — that window might be 60 days. For stable conceptual content it can be longer. When the review date passes, Chattering should not be serving that article until a human has confirmed it's still current.
In practice this means:
review_by field in your help center.A gap in coverage produces an "I don't have enough information to answer this" response, which Chattering handles by offering a human handoff. That's a recoverable situation. A confident wrong answer is harder to undo.
When a customer quotes a specific article and the conversation escalates to a human agent, the handoff context Chattering passes should surface which article the AI cited. Your inbox view shows the full conversation thread plus the citation. The support rep's first job before replying is to check whether that citation is still live and accurate.
If the rep finds a mismatch — the customer quoted something real but wrong — the correction workflow is:
Step three is the one teams skip. They fix the customer, forget the article, and the same failure recurs the following week.
Chattering's feedback portal lets you log recurring mismatches. When a customer quotes stale content, that's a signal worth capturing: what article, what claim, what the accurate answer is. Over a month you'll see patterns — certain articles generate a disproportionate share of correction tickets, usually because they cover something that changed recently and attracted inbound traffic right after the change.
Those high-signal articles should get permanent review cadences shorter than your default. If your pricing page generated three correction tickets in six weeks, it gets a 30-day mandatory review, not a 90-day one.
The language matters. Don't blame the AI, and don't make the customer feel like they did something wrong by finding your article. A direct, low-drama response works best:
"That article is out of date — our apologies for the confusion. Here's how [topic] actually works today: [correct answer]. We're updating the article now."
Short, takes ownership, gives them the right information, and signals that you're acting on it. Customers who see that response tend to drop the frustration quickly because you've shown the problem is being fixed, not just acknowledged.
Every protocol above is damage control. The upstream fix is treating documentation as a release artifact, not an afterthought. When a feature ships, the doc update and the index re-sync happen on the same day. When a price changes, the help center and the Chattering knowledge base are updated before the change goes live for customers.
That discipline shrinks the window in which a customer can find stale content and quote it back to your agent. You won't eliminate it — someone will always find a cached version of something somewhere — but you make it rare enough that the correction protocol is a light touch, not a weekly firefight.
Chattering lets you control which URLs or content sources are included in the knowledge index independently of your help center's publish state — remove the article from the index sync while keeping the public URL live, then re-add it once the content is updated.
Chattering responds that it doesn't have enough information to answer confidently and offers to connect the customer with a human agent, so the gap in coverage produces a clean handoff rather than a wrong answer.
Not automatically — the agent matches on content it has, not content it lacks; that's why the human-review step at handoff is essential for catching the mismatch and triggering the correction workflow.