Knowledge gaps: turning the questions your AI could not answer into answers

A knowledge gap is a question a customer asked that the AI had no approved fact to answer. A grounded AI says "I am not sure" and logs the question; the gaps log is the list of those. Reviewing it monthly, grouping similar questions, and adding a fact or an FAQ for each group is how the AI gets better at your specific business. Most businesses find the log is long in the first month, short by the third, and a steady trickle after that.

Why gaps are good news

Each gap is a real question from a real customer, in their words, that your front desk answers every week and nobody had written down. The alternative to a gap is a guess, and a guess is worse. A log of honest "I am not sure" replies is the AI telling you exactly what to teach it next.

What the log shows

The question as asked, the date, the channel, what the AI replied, whether a person then answered, and what they said. That last column is the draft answer: if Dana replied "yes, we take Delta, bring your card", the FAQ writes itself. Similar questions grouped together ("do you take Delta", "is Delta Dental accepted", "Delta?") so one fact closes ten gaps.

The fifteen-minute monthly review

1. Sort by frequency. The top ten groups are most of the volume.
2. For each: is there a fact or an FAQ that answers it? Write it, or approve the draft from the person's reply.
3. If the question needs judgement (a medical, legal or pricing-by-visit question), it is not a gap; it is a handoff. Add it to the escalation list if it is not there, and mark it so it stops appearing as a gap.
4. If it is a one-off ("do you validate parking at the mall garage on Sundays"), answer it in a fact anyway if it is true; it costs nothing.
5. Check the facts you added last month against this month's gaps. If the same question is still appearing, the fact did not match the phrasing; add the customer's wording as an alternative.

Gaps to leave as gaps

Questions the AI should never answer. "Is this lump serious?" "Do I have a case?" "Will insurance cover it?" These belong on the handoff list, permanently. Filling them with a fact ("we cannot say without an exam") is fine as the handoff wording, but the handoff itself must remain: a person, promptly.

Gaps that reveal a business problem

"Why has nobody called me back?" is not a knowledge gap. "Is the Northgate branch still open?" after a move is a website problem. "Why was I charged twice?" is a billing problem. The gaps log is also a list of things customers are confused about, and some of them are not the AI's to fix.

Drafting from conversations

A good product proposes FAQs from the gaps and the human replies, in the customer's phrasing, for you to approve. Approving ten drafts takes five minutes. The approval is the important step: a draft that goes live unapproved is a guess with better formatting.

Measuring progress

Gaps per hundred conversations, by month. It should fall steeply and then flatten. The flat level is new questions that genuinely need a person; if it rises, something changed (a new service, a new location, a price change nobody loaded).

Frequently asked questions

Who should review the gaps?

Whoever answers the phone. They know the right answers. The owner approves anything about policy or price.

Should I fill every gap?

Every gap that has a factual answer, yes. Gaps that need judgement become handoffs. Gaps that reveal a business problem get fixed elsewhere.

What if the AI marks something as a gap that it should have answered?

The fact exists but did not match the phrasing. Add the customer's wording to the fact as an alternative phrasing and it will match next time.

What Frontiva does here

Frontiva's knowledge gaps log is deduped and counted, ranked by how often each question is asked, so the top of the list is the next FAQ to write. FAQs are written and approved by your team; the product does not draft them for you yet. See grounding explained for why the AI says it is not sure in the first place.