Blog

What your help centre says vs what a support agent needs

Numair Khan · 24 October 2026 · 5 min read

A help centre states the rule. A support agent, human or AI, gets the case that does not fit the rule. The answer to that case is usually in someone's head, not in an article.

If your AI support agent gets refund exceptions wrong while your help centre is up to date, the likely cause is not stale articles. It is that the articles never covered the exception. This post walks through that gap with a refund case and gives a way to find your own gaps in an afternoon.

We build a tool for the last part of this post, so weigh that. The method in the middle works without us.

The pattern: the article covers the normal case

Most help centre articles describe what should happen. Support work is mostly the other cases: the customer who cancelled mid-term, the contract with different terms, the credit above the usual limit. Experienced people handle these from memory. They know who to ask and what is usually granted.

A human agent who does not know asks a colleague. An AI agent has no colleague unless you give it one. It answers from the article, guesses, or escalates.

A typical refund article

Here is its help centre article, in full:

Refunds
Monthly plans can be cancelled at any time. You will not be charged again.
Annual plans can be refunded within 30 days of purchase.
Contact support to request a refund.

It is clear and correct as far as it goes. Now here are five questions a customer could plausibly ask:

1. "I'm on an annual plan and want to cancel in month 7. Do I get anything back?"
2. "Can I get a credit instead of a refund?"
3. "The amount is more than my account manager can approve. Who decides?"
4. "Our company signed a custom contract. Does the 30-day window apply to us?"
5. "I was charged twice after a failed payment. Is that a refund or a correction?"

The article answers none of them. Suppose a senior support agent at Acme says this in an interview:

"Officially it's no refunds on annual, but if they're past the window we do a prorated credit for the months left. Anything over five hundred I check with Dana in finance first."

That quote answers questions 1, 2 and 3. It is not in the article. Questions 4 and 5 need other people, which is the point of the afternoon method below.

Written as rules, with each reason taken only from the quote:

Notice two things. First, the useful knowledge has a situation, an action, a threshold and a person, and it cannot be found by rewording the article. Second, the interview did not give a reason for the approval step. A rule with an unknown reason should say so, so the reviewer can ask. In Keeplore the reason is left empty when the person did not give one.

An agent that answers from the article alone either quotes the 30-day rule to someone past it or escalates every one of these cases.

What vendors say causes wrong answers

Support-tool vendors Macha and Mava both publish advice on this, covering conflicting articles, knowledge gaps and escalation: Macha, Mava. If your help centre has duplicate or outdated articles, fix those first.

How to find your own gaps in an afternoon

This takes about half a day and does not need any tool.

1. List the 10 questions your team escalates most. Ask the people who handle escalations. Or read the last month of escalated tickets. Take the cases, not the categories.
2. For each, check if the help centre answers it. Search it as a customer would. Does an article answer this exact situation, or only the normal case beside it?
3. Note who answers it today. Write down the name. If three of the ten point to the same person, you have found where the unwritten rules live.
4. Mark which answers are risky. Refunds, credits, contract terms and anything with an amount are the ones to write down first.
5. Test your agent with questions whose right answer is "we don't know". If it answers confidently, it is guessing.

You now have a list of gaps and the people who can fill them. That list is useful even if you never use Keeplore.

Getting the answers out of people

The person who answers those escalations usually cannot write it all down on request. They do not think of it as knowledge. They think of it as Tuesday.

What works better is a conversation. Ask for the last time it happened, not what usually happens. Ask what made it an exception. Ask who approved it. Then write each answer as a rule: when this, do that, because of this. Keep their own words next to it, so a reviewer can see where the rule came from. Have a manager approve it before an agent uses it.

Our question list for knowledge transfer interviews is free and written for this. It works for someone leaving, someone changing role, or the colleague everyone already asks.

Keeplore does these steps. It interviews the person in a text chat for 30 to 45 minutes, keeps asking about exceptions, turns the answers into rules with a condition, an action, a reason and the exact quote, and drops any rule whose quote is not in the transcript. A manager approves each rule. Your agent then calls an API and gets the approved rules and who they came from. There is no integration with any particular support tool today. How it works.

When the answer is "we don't know"

Some questions have no rule. The right behaviour is for the agent to say so and hand over to a person, not to guess. Keeplore returns an empty answer in that case, so your agent can escalate on an empty result. Whether your agent does that depends on how you set it up.

Related

• Knowledge transfer interview questions
• How Keeplore works
• Keeplore for customer support teams