How to answer product questions from the help centre with an AI agent

The help centre is a safe source only if the agent treats it as the whole truth. It answers from published articles, cites the one it used, and says plainly when none covers the question. It never stitches two articles into a feature that does not exist. Gaps and contradictions become draft changes for a named owner to approve.

What is the process, and who owns it?

Most firms already have a help centre: articles explaining how products work, how to do common tasks and what to do when something goes wrong. An agent answering product questions from it can resolve many contacts in seconds. But help centres are written to be browsed, not queried. Articles overlap, go out of date, differ by region and assume the reader has seen the page before.

The knowledge or content team owns the help centre. Each article should have an owner, usually in the product team. Customer support owns the conversation. The agent is a reader and a reporter of gaps, not an author, which fits the stance in our guide to the AI support agent that knows when to say no.

What does the agent prepare, and what does a person approve?

1. Person, content lead: defines the source set. Published articles only, for the regions and products in scope. Drafts, internal notes and retired articles are excluded.
2. Agent, depth 1: identifies the product, region and plan the customer is asking about, since answers often differ.
3. Agent, depth 1: retrieves matching articles and checks their review dates and regions.
4. Agent, depth 3: answers from one or more articles, citing each, in plain words for the question asked.
5. Agent, depth 3: says when no article answers. "I cannot find that in our help articles. I will pass your question to the team." No guessing.
6. Agent, depth 1: logs gaps and conflicts. Questions with no answer, articles that contradict each other, and articles past their review date.
7. Agent, depth 2: drafts article changes: a new article for a common gap, a correction where two articles conflict, a regional note where one is missing.
8. Person, article owner: approves, edits or rejects each draft before it is published.
9. Person, adviser: answers routed questions the help centre does not cover.

Which controls matter, and what evidence does each leave?

Published means approved. Only articles that passed your content approval are sources. Evidence: article IDs and versions cited in each answer.

No synthesis beyond the text. If article A says a feature exists and article B says how payments work, the agent does not infer that the feature works with payments. Certification tests for this. Evidence: grounded-gate results.

"I don't know" is a valid answer. Evidence: routed questions with the reason "no source".

Region and plan are checked. An answer for one country's product can be wrong in another. Evidence: region and plan recorded with each answer.

The agent cannot publish. Article changes are drafts. Evidence: publication events attributed to named owners.

What should you measure?

• Questions answered with a citation, against questions routed for lack of a source.
• Gaps logged, by topic, and how many led to a published article.
• Conflicts found between articles, and time to resolve them.
• Answers later found wrong, traced to the article or to the agent.
• Articles past review date that the agent is still citing.

Where does it usually go wrong?

Stale articles answer confidently. An article describing an old screen or a retired feature is still published. Review dates matter more once an agent is reading every word.

The agent fills the gap. Asked something the articles do not cover, it answers from general knowledge about similar products. That is the most common failure and the most damaging.

Help content assumes a capable reader. A customer who is confused, distressed or has difficulty reading needs more than a cited article. Can an AI agent help a vulnerable customer? sets out when to hand over.

Account questions are answered as product questions. "How long does a transfer take?" has a general answer, but "where is my transfer?" needs account data, as with order status questions.

Screenshots carry the answer. Many articles explain a task through images of the screen, with little text. The agent cannot cite what it cannot read, so it either skips the article or guesses at the steps. Ask article owners to put every instruction in text, with images as support.

Old and new versions sit side by side. During a product change, both the old and new article are published for a while. The agent must know which customers are on which version, or it will give half your customers the wrong steps.

Gaps are logged but never fixed. The log fills up and nobody owns it. The method in finding gaps in the help centre with an AI agent turns the log into a work queue.

Where Pryme fits

Pryme lets you mark the published help centre as the agent's only source for product questions, and Attest's grounded gate checks on your own past questions that every answer traces to an article. The connection itself, and what the agent may read or write there, is covered in how to connect an AI agent to your help centre.