Gaps show up as questions the help centre fails to answer. An agent can collect those signals from failed searches, routed contacts, repeat contacts and its own "no source" answers, group them into topics, check what the articles say, and rank each gap by volume and potential harm. The content lead sets priorities. Article owners write and approve every fix.
What is gap finding, and who owns it?
A help centre has gaps of several kinds. Missing: no article covers the question. Outdated: the article describes an old screen or rule. Conflicting: two articles disagree. Unclear: an article exists, but customers read it and still contact you. Misplaced: the answer exists but nobody finds it. Each needs a different fix, and the first job is telling them apart.
The content or knowledge lead owns the help centre and the backlog of fixes. Article owners in product teams own the substance. Support leaders know which contacts hurt most. The agent does the collecting and sorting, consistent with the role of support agents in our guide to the AI support agent that knows when to say no: it reports what it cannot answer instead of inventing an answer.
What are the steps, and who acts at each one?
- Agent, depth 1: collects signals from the sources you connect: help centre searches with no click or no result, questions an answering agent routed for lack of a source, contacts made shortly after viewing an article, negative article feedback, and themes in call and chat transcripts.
- Agent, depth 1: groups signals into topics by meaning, not wording, so "can't log in on new phone" and "moved to a new handset, app won't open" land together.
- Agent, depth 1: checks coverage for each topic: is there an article, is it current, does it conflict with another, does it cover every region and product version?
- Agent, depth 2: classifies the gap: missing, outdated, conflicting, unclear or hard to find, with examples of customer wording.
- Agent, depth 2: ranks gaps by volume and by potential harm. A rare question about a frozen account outranks a common one about changing a profile picture.
- Person, content lead: sets priorities and assigns owners.
- Agent, depth 2: drafts fixes: a new article, a correction, a merged article to end a conflict, a rewritten title to aid search.
- Person, article owner: approves and publishes, after checking substance with the product team.
- Agent, depth 1: monitors the topic after the fix and reports whether contacts and failed searches fell.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Collect | Gathers signals from sources | 1 | Choosing sources |
| Group | Clusters by meaning | 1 | Nothing |
| Coverage | Checks articles per topic | 1 | Nothing |
| Classify | Labels the gap type | 2 | Nothing |
| Rank | Orders by volume and harm | 2 | Setting priorities |
| Draft | Prepares fixes | 2 | Owner approves |
| Monitor | Tracks the topic after fix | 1 | Nothing |
Which controls matter, and what evidence does each leave?
Customer wording kept as evidence. Owners see what customers actually asked, not a summary. Evidence: sample messages stored with each gap, minus personal data.
Harm counts, not just volume. The ranking weights topics where a wrong or missing answer could cost a customer money, access or safety. Evidence: the ranking with the harm reason stated.
Drafts are drafts. Evidence: publication by named owners only.
Personal data is stripped. Signals come from customer conversations; gap reports should not carry names or account details. Evidence: redaction checks on the report.
Before and after. Every fix is measured against the signal that raised it. Evidence: topic monitoring results.
What should you measure?
- Gaps found and closed, by type.
- Time a gap stays open, especially high-harm ones.
- Contacts and failed searches on a topic before and after the fix.
- Questions routed for lack of a source by the answering agent, as the help centre improves.
- Drafts rejected by owners, and why.
Where does it usually go wrong?
Volume drowns harm. The top of the list fills with trivial topics. Weight by harm, and read the long tail.
Content fixes product problems. Customers cannot find a setting because the screen is confusing. An article helps, but the gap belongs with the product team too. Report both.
Conflicts are resolved by the agent. Two articles disagree, and the draft fix simply picks one. The agent cannot know which is right; it should present both, with the product owner deciding which describes how the product actually works.
Drafts promise too much. An agent-written article says "refunds arrive within three days" because customers asked. The owner must check what is true.
Accessibility gaps are invisible. Articles that rely on images or dense text fail some customers. Can an AI agent help a vulnerable customer? shows why plain language and alternatives matter as much as coverage.
Signals come from one channel. Searches show what web users want; call transcripts show what phone users need. Connect both, which depends on how you connect an AI agent to your help centre and to your contact channels.
Fixes do not reach the answering agent. A new article is published but the agent's index is stale, so routed questions continue. Close the loop with answering product questions from the help centre with an AI agent.
Where Pryme fits
Pryme files every "no source" answer from your support agents, which makes them a ready signal for gap finding. The gap-finding agent's scope is depth 1 and 2 only; drafting is allowed, publishing is not, and the owner's approval is the recorded act that puts a fix live.
Questions people ask
How often should gap finding run?
Continuously for signal collection, with a regular review cycle for priorities. After a product launch or policy change, look daily for a while.
Can the agent find gaps in internal knowledge too?
Yes, with the same method on staff questions, kept separate from the customer help centre. Internal gaps often explain why advisers give inconsistent answers.
Does gap finding need access to customer accounts?
No. It works on questions and article content. Keeping account data out reduces what the reports could expose.