What to Put in an Outreach Tool's Knowledge Base, and What Happens When You Don't
· 3 min read
Ask a general-purpose AI model to write a cold email mentioning your pricing, and it will cheerfully invent a number if you didn't give it one. Not out of malice — it has no concept of "I don't actually know this," only a pattern of what a plausible-sounding answer looks like. The fix isn't a smarter model. It's not asking the model to know things it was never told.
Confidence is not one setting, it's a result
A knowledge base search doesn't return a single yes-or-no on whether it found an answer — it returns a match quality. A question that lands close to something you've actually documented is a strong match. One that's loosely related but not really the same question is a weak match. One with nothing close in the knowledge base at all is no match. Those three outcomes should produce three different drafts: state it plainly, mention it but hedge, or leave it out entirely. Treating a weak match the same as a strong one is how a draft ends up stating something as fact that was really a guess dressed up as a lookup.
What's worth actually documenting
The highest-value content for this isn't a polished "About Us" page — it's the specific facts a draft would otherwise have to invent or omit:
- Pricing, even roughly ("starting from," a typical project range) — the single most common thing a model will confidently fabricate if it isn't given anything.
- What you actually do and don't offer — a short, plain list beats a long services page a search can't match cleanly against.
- A real result or two, with enough specifics that it reads as a fact rather than a claim — "cut their no-show rate from roughly a third to under one in ten" is checkable in spirit even if the exact source isn't quoted; "we get great results" isn't.
- Anything you'd be annoyed to see a draft get wrong — if a wrong guess here would embarrass you, it belongs in the knowledge base, not left to a model's judgment.
What this doesn't fix
A knowledge base stops a draft from inventing facts about your business. It does nothing about a draft inventing facts about the lead's business — a separate problem that has to be solved by only ever quoting what the lead's own listing or site actually says, never by having the model guess at a detail that would make the message feel more personal.
How AyePin handles the "no match" case
AyePin's Knowledge page chunks whatever you upload and compares a question against it by similarity, in two tiers: a close enough match is used directly, a loosely related one is used but marked low-confidence, and anything with no real match at all is refused rather than guessed. A draft can mention a price, a result or a specific claim only when it traces back to something you actually gave it; anything it can't find is left out of the message, not filled in with something plausible-sounding instead.