Why does the AI SDR make up answers when the knowledge base doesn't cover something?
Why does the AI SDR make up answers when the knowledge base doesn't cover something, and what to check next.
The AI SDR isn't lying on purpose — it's doing what language models do when asked a question they don't have context for: guessing plausibly. That's fine for a trivia game and terrible for a sales conversation where confidence about pricing or features becomes a promise you have to honour.
The fix is a combination of three things: make the KB comprehensive, tell the AI explicitly what to escalate, and monitor the Needs Attention folder so you catch drift early.
Why it happens
When a prospect asks something the KB doesn't cover, the AI has three options:
- Say "let me check" — which requires it to be told that's a valid move.
- Escalate the thread — which requires the AI to recognise the gap.
- Generate a plausible-sounding answer — the default fallback.
Out of the box, option 3 wins too often. Fixing that is deliberate work.
Fix 1 — Fill obvious KB gaps
Prospects ask the same questions over and over. Cover the high-volume ones explicitly.
| Topic | Example KB entry |
|---|---|
| Pricing | "Q: How much does this cost? A: Our packages start at $X. Full pricing depends on scope — happy to send a proposal after a discovery call." |
| Timeline | "Q: How long until I see results? A: Most clients see first replies in week 2, first booked calls in week 3–4." |
| Guarantees | "Q: Do you guarantee results? A: We don't guarantee specific outcomes, but we do guarantee [the delivery promise that's actually yours]." |
| Team / company basics | "Our company is [Name]. Founded [year]. Based in [location]. CEO is [person]." |
The rule of thumb: if you've been asked the same question three times in the inbox, it belongs in the KB.
Fix 2 — Tell the AI what to escalate
Add explicit escalation rules to the KB so the AI knows when to step back:
Escalate to human if the prospect:
- Asks about custom pricing or deals
- Mentions legal, compliance, or security questions
- Wants to negotiate scope
- Asks a question that isn't answered aboveA short list like this reduces fabrication noticeably. The AI respects the "isn't answered above" line as permission to punt.
Fix 3 — Watch the Needs Attention folder
Escalated threads land in AI SDR Inbox → Needs Attention. Check it daily. Every escalation you see is either:
- A legitimate handoff (good — the AI did its job)
- A question you could have pre-answered in the KB (good — now you have a new KB entry to write)
Both outcomes reduce fabrication over time.
What the KB can't fix
The KB gives the AI facts. It doesn't change the AI's underlying writing style, persona, or behaviour patterns — see the KB tone/persona FAQ for what's in scope and what isn't.