An AI knowledge base gap is information a customer needs that your company has not documented in sources available to the assistant. Build the gap list from actual conversations rather than imagining everything you could publish. The result is a backlog of intents, owners, and acceptance rules; keep cases where correct information already exists but was not retrieved in a separate diagnostic queue.
Collect missing intents from conversations
Review a defined period and record the customer question, intended action, and the part of a human reply that actually resolved the information need. Each entry links to a conversation and describes the missing fact without copying personal details into the worksheet.
Example: five people ask about weekend service; the offer lists office opening hours but never explains when the service is delivered. Do not copy entire histories or assume that every human handoff proves a knowledge gap.
Zendesk recommends finding knowledge topics in tickets, grouping issues, and documenting recurring questions; the backlog here is a manual proposal. Zendesk: finding customer issues for knowledge.
Existing-content diagnosis belongs in a chatbot that cannot find an answer; keep that work separate from missing topics.
Group questions by the customer decision
Combine different phrases only when the answer supports the same decision under the same offer conditions. A group contains sample questions and one working description of the customer’s goal, not a broad label such as “information”.
Example: “do you work on Saturdays?” and “weekend delivery” may share an intent, while “can I contact support on Saturday?” needs another answer. Do not merge service availability with agent availability just because both questions mention weekends.
Scroll the table sideways to see all columns.
| Check | Action | Evidence of completion |
|---|---|---|
| Collect missing intents from conversations | Review a defined period and record the customer question, intended action, and the part of a human reply that actually resolved the information need. | Each entry links to a conversation and describes the missing fact without copying personal details into the worksheet. |
| Group questions by the customer decision | Combine different phrases only when the answer supports the same decision under the same offer conditions. | A group contains sample questions and one working description of the customer’s goal, not a broad label such as “information”. |
| Separate missing content from failed matching | Check whether an approved answer already exists in the sources before commissioning a new article, and label diagnostic cases separately. | The editorial backlog contains undocumented facts; existing correct content sends the case to publication or retrieval review. |
| Assign priority and an information owner | Consider intent frequency, decision significance, and handling effort, then assign a person authorized to approve the fact. | A chosen item includes a priority reason, preparation date, and source for the business decision rather than only a conversation count. |
| Write the answer and its boundaries | Prepare content that resolves the intent with conditions, exceptions, and a next action, then identify one source version to maintain. | Ready means an approved fact, clear scope, and an answer an agent or assistant can use without further guessing. |
| Verify the gap is closed on new questions | After publication, test the original wording and review later conversations, separating new exceptions from instructions that remain unclear. | Close the item after checking a correct answer and preserve its source and acceptance evidence in the record. |
A backlog sheet can hold intent, sample questions, observed case count, missing fact, owner, source, and editing status. Counts describe the selected period, not market demand. Attach one typical conversation and one exception so the editor understands the boundary. The broader topic-selection method is covered by building a knowledge base and FAQ.
Separate missing content from failed matching
Check whether an approved answer already exists in the sources before commissioning a new article, and label diagnostic cases separately. The editorial backlog contains undocumented facts; existing correct content sends the case to publication or retrieval review.
Example: an address-change procedure already exists, so writing another copy would increase contradiction rather than close a gap. Avoid adding synonymous documents as your first response to every failed assistant test.
A sample backlog item reads: “The customer wants to know whether this service is delivered on Saturday.” The missing-fact field identifies delivery hours for that service, and the owner is the person responsible for its schedule. Give the editor two real question variants and a public-holiday exception. Acceptance requires a confirmed ordinary rule and the exception boundary.
Only then is the content ready to become an assistant source. If the owner explains that availability is individual, the correct result may be contact instructions instead of an availability table; a backlog must not force the company to publish a rule it does not have.
Assign priority and an information owner
Consider intent frequency, decision significance, and handling effort, then assign a person authorized to approve the fact. A chosen item includes a priority reason, preparation date, and source for the business decision rather than only a conversation count.
Example: a rare question about an important exception can deserve faster work than a common question about an already visible address. A missing policy decision means waiting for approval, not permission for AI to invent the answer.
- Confirm the fact is absent from approved knowledge.
- Keep the original question without unnecessary customer data.
- Separate similar wording that leads to different decisions.
- Name the person authorized to approve the information.
- Test the published answer and record its source.
For contradictory documents, improve chatbot knowledge editing instead of enlarging the backlog of new pages.
Write the answer and its boundaries
Prepare content that resolves the intent with conditions, exceptions, and a next action, then identify one source version to maintain. Ready means an approved fact, clear scope, and an answer an agent or assistant can use without further guessing.
Example: a weekend answer names the service type and exception instead of making a vague promise of flexible customer care. Keep internal negotiations, individual order details, and unapproved alternatives out of general knowledge.
Once an answer is approved, Meet AI lets you add a supported source, publish knowledge, and use a private test with citations. This needs an account, permissions, an active application, and an eligible plan. The gap list described here is your team’s own worksheet; it does not imply a built-in automatic report of missing intents.
Verify the gap is closed on new questions
After publication, test the original wording and review later conversations, separating new exceptions from instructions that remain unclear. Close the item after checking a correct answer and preserve its source and acceptance evidence in the record. To check whether the new answer reduces contacts on that topic, use the before-and-after measurement described in the guide on the ticket deflection rate.
Example: the basic weekend question is answered, but a later holiday question introduces a separate schedule exception. Retain completed entries so a later policy change can explain why the answer was created.
For intents needing a person, prepare a chatbot-to-human handoff instead of replacing every individual decision with a public answer.
Questions that come up in practice
Is every AI refusal a knowledge gap? No. The question may require individual account data, an action outside the assistant’s scope, or existing information that was not matched. Classify the cause before creating content.
How many conversations justify a topic? There is no universal threshold. Repetition helps, but one important exception can justify work. Record the reason for your decision instead of claiming statistical certainty.
Can I reuse an agent reply unchanged? Treat it as material rather than approved policy. Remove individual details, confirm the rule with its owner, and check whether the answer includes conditions that apply to other customers.
