Support knowledge base training: using answers correctly

Short answer

Training agents to use the knowledge base well teaches five habits: searching with the customer's words, checking the date and scope of the answer found, adapting the article to the customer's question rather than pasting it, not guessing when there is no answer, and reporting the gap. These habits need practice of their own, because experienced agents often trust memory while new ones get lost in search. The most useful drill is a scenario where a policy changed recently: an agent answering from memory gets it wrong; one who checks the base does not.

Why agents don't check the knowledge base

• Search is slow or the words don't match: the customer says "money back", the base says "refund policy".
• The base is out of date and the agent once found a wrong answer — they no longer trust it.
• The experienced agent says "I know this" — but the rule changed last month.
• The article is long, and finding the part relevant to the customer takes time.
• There is no clear rule for answering a question the base does not cover.

Two of these are problems with the base and three with habits. Training fixes habits; problems with the base need solving separately, or even the best habit fails on a poor base.

Five habits

1. 1Search with the customer's wordsThe customer's own phrase first, then synonyms and internal terms: "money back", "refund", "return payment".
2. 2Check date and scopeWhen was the article updated, and which product, channel and customer type does it cover?
3. 3AdaptGive the two or three sentences relevant to the customer in plain language — not the whole text.
4. 4Don't guessIf there is no answer: "I'll confirm and come back to you" — and escalate.
5. 5Report the gapSend missing or wrong information to the base's owner with a short note.

Adapt, don't paste

A knowledge-base article is written to cover every case: exceptions, conditions, legal notes. The customer has one question. An agent who sends the article as it is makes the customer find the answer themselves. The right way: pick the part that fits the customer's situation, say it in plain language, and add a link to the full rule if needed. The meaning must not change: deadlines, amounts and conditions stay as they are in the base.

When there is no answer

When the base has no answer, the agent has two bad options: guess, or say "we don't know" and close the conversation. The right option is an honest reply plus escalation: "I'm checking the exact information and will write to you today by 16:00." The request goes to the right address and the gap is reported to the base's owner. Escalation itself is covered in support escalation training.

Gap reports

Agents' gap reports are the main source of knowledge-base quality — but only get sent when they are easy. Three fields are enough: the customer's question (in their words), what was or was not found in the base, and how the agent answered. The base's owner reviews reports weekly and sends the team a short update on the outcome — agents need to see their reports making a difference, or they stop sending them. How gaps are found from conversations is shown in knowledge base gap analysis.

Practice scenarios

• A rule changed recently: the return period got longer, the old answer is wrong.
• The answer exists but under different words: the customer uses a synonym the agent must find.
• The answer is right for one product but not another: checking scope.
• The answer is not in the base: escalate without guessing.
• A long article: pick the two sentences relevant to the customer and put them plainly.

Building scenarios is covered in detail in AI customer simulation for support.

What to measure

• The share of practice answers that match the base — the "accuracy of the answer" criterion.
• Cases of the old answer being given in the first week after a rule change.
• The number of gap reports from agents, and how many were added to the base.
• Cases where guessed answers are spotted — in practice and in reviews of real chats.

An illustrative example

This is an illustrative example. An electronics retailer changes the warranty period for some products. Two weeks later, a review of real chats shows some experienced agents still quoting the old period. The team adds a "warranty changed" scenario to the simulation: the AI customer asks about a product covered by the new rule.

In the first round several agents answer from memory and get it wrong. After the feedback conversation, the team starts tracking the "check the base first" habit separately in weeks when a rule changes.

Common mistakes

• Teaching the base only to new agents — experienced ones work from memory.
• Sending articles to customers as they are.
• Guessing when there is no answer.
• Making gap reports hard — nobody sends them.
• Using an out-of-date base as the "right answer" in training.

Limitations

Training does not fix a poor knowledge base: if articles are out of date, the problem is wrong information, not a wrong habit. The base needs an owner, an update rule and dates. In simulation the "right answer" also comes from the base — if the base is wrong, so is the evaluation.

The knowledge base in Vexvon

In Vexvon, the knowledge base is built from five sources: website crawling (up to 4,500 pages), PDFs (up to 25 MB), manual and bulk entry (up to 200 per request), the product catalogue and the Instagram archive. Each item can be targeted at the chatbot, the call agent or both, and grouped by category and type. For every bot reply, the knowledge pieces it used are stored — which helps find outdated information. In AI Training, customer profiles can be drafted from the company's knowledge base. More on the knowledge base.

Next step

Pick two rules that changed last month and write a simulation scenario for each. You will see what share of the team gives the new rule. How a knowledge base is prepared for simulation is covered in a knowledge base for training. More articles are in the customer service training section, and we can build your scenarios together during a demo.