On this page
When an AI agent gives a poor answer, the instinct is to blame the model. Most of the time, the problem is upstream: the information it was given was vague, scattered, outdated or written for a different reader.
The good news is that fixing the material is something any team can do — no engineering required. Here are seven habits that make the biggest difference.
1. Write the question the way customers ask it
Internal teams name things one way; customers name them another. Your help center says “Shipment status inquiry”. Customers type “where’s my parcel”, “has my order shipped”, “tracking number?”.
When you add a Q&A pair, write the question in the customer’s words. Better still, look at real conversations and copy the phrasing people actually use. In AskCub, an exact match on a Q&A pair is sent word for word — and doesn’t use an AI answer — so getting the wording right pays twice.
2. One topic per entry
A long FAQ entry that covers shipping costs, delivery times, international customs and returns in one block is hard for anyone to use — human or AI. When the AI retrieves it, it has to pick the relevant sentence out of a wall of text.
Split it. “How much does shipping cost?” and “How long does delivery take?” are two entries. Short, focused entries match more precisely and produce cleaner answers.
3. Put the numbers in
The details customers ask about are almost always specific: how many days, what price, which countries, until when. Vague answers (“delivery is usually quick”) force a follow-up question every time.
Compare:
- “We ship quickly to most countries.”
- “Orders ship within one business day. Delivery takes 2–4 business days in the EU and 5–8 elsewhere. Shipping is free over €49.”
The second one answers the next three questions too.
4. End with what to do next
An answer that explains a policy but doesn’t say how to act on it leaves the customer stuck. End entries with the next step: a link, a button name, an email address, a form.
“You can return items within 30 days” becomes “You can return items within 30 days. Start the return from your order page and we’ll email you a free label.”
5. Keep one source of truth
If your returns policy lives on a web page, in a PDF and in three FAQ entries — and they disagree — the AI may quote any of them. Pick one place for each topic. If you import web pages into the knowledge base, update the page and re-import rather than editing copies.
6. Remove what’s no longer true
Outdated material is worse than missing material. A missing answer leads to a hand-off; an outdated one leads to a confident wrong answer with a source attached. When something changes — a carrier, a price, a holiday schedule — update or disable the entry the same day.
Temporarily untrue? Disable rather than delete. AskCub ignores disabled entries until you turn them back on.
7. Let real conversations tell you what’s missing
You don’t have to guess what to write next. Two places tell you:
- The not-answered list — questions the AI couldn’t answer, grouped and counted.
- Your team’s resolved conversations — the answers your agents already gave. AskCub can turn the last 30 days of resolved conversations into Q&A drafts for you to review.
Answer the top three gaps each week. Within a month the knowledge base covers what customers actually ask, not what you imagined they’d ask.
Q&A pairs or documents?
Use both, for different jobs:
- Longer policies and guides
- You want the AI to pick the relevant part
- Searched first; summarised with a source
- The twenty questions asked every day
- You want the exact same wording
- Instant; uses no AI answers
A simple starting plan
- Export your last month of chat or email questions.
- Group them into topics and count them.
- Write Q&A pairs for the top twenty.
- Import your five most visited help pages as documents.
- Test with a private conversation, phrasing each question three different ways.
That’s an afternoon of work, and it will do more for your AI agent’s answers than any setting.
Related articles
AI customer service4 min read
When should an AI agent hand a conversation to a person?
Five hand-off triggers that keep customers from getting stuck with a bot — and the three mistakes that make AI support feel worse than no support.
Read article
Cross-border & multilingual3 min read
Seven questions overseas shoppers ask before they buy
Shipping, customs, payment, returns, sizing, trust and timing — what cross-border buyers need to hear, and how to answer each one in chat.
Read article
AI customer service3 min read
Measuring an AI agent without fooling yourself
“The AI resolved 70%!” sounds great until you ask how it was counted. Four numbers that tell you whether AI support is actually helping.
Read articleTry it on your own site
A free plan with no time limit, AI agent included. One line of code.