An AI strategy tool hallucinated $400k+ from meeting notes with zero numbers in them
I ran a quick test: I gave a well-known AI business planning tool (Venturekit) a set of meeting notes that had no financial data in them at all — no revenue, no costs, nothing. It still generated a full revenue forecast: $400k+, broken down by product line, with 5-year projections.
That's when I decided the thing I'd been building for the past few months was worth taking seriously.
AnchorStrategy is a multi-agent AI pipeline that takes your meeting notes and runs them through the same frameworks a consultant would use (SWOT, Porter's Five Forces, STP, Ansoff, BCG, VRIO) — but every claim it makes has to trace back to something actually in your notes. If it can't, it says "insufficient basis" instead of making something up. I built specific guardrails to catch and block fabricated numbers before they reach the final report.
It outputs a formal report, a plain-language summary, and a pitch deck.
I'm opening it up as a small, free private beta — looking for founders, operators, or consultants who've been burned by AI tools that sound confident and turn out to be wrong, and who'd be willing to try it on real (or realistic) meeting notes and tell me honestly where it breaks.
If that's you: https://docs.google.com/forms/d/e/1FAIpQLSdvwMaqaD6RHX0TCYpd8Rd0tEBuriHXTtDwfyeejTm8QAE1AA/viewform
Happy to answer questions about the architecture too — happy to nerd out on how the guardrails work if anyone's curious.
The $400k forecast pulled out of notes with zero financial data is the detail I keep chewing on. It filled the gap because nothing told it what the report was actually for.
The habit that helped me most: write the one sentence core job before anything else, then judge every output against it. Mine reads "turn a meeting into decisions the notes can prove." An invented revenue line fails that sentence in a second, no guardrail debate needed.
I wrote up how I hold that line here: https://durablefoundations.gumroad.com/l/pyramid-reality-check
What is the one sentence job for AnchorStrategy?
Kael Voss / DurableFoundations
Good question — ours is: "Turn meeting notes into a strategy report where every number can be traced back to something someone actually said." Same instinct as your line — anything that can't point to a specific sentence in the notes doesn't get asserted as fact, it gets flagged as "insufficient basis" instead. The $400k case is exactly what happens when a tool skips that check.
This comment was deleted 17 days ago
The strongest part is the explicit evidence boundary. Refusing to produce a financial conclusion when the source contains no financial evidence is a much stronger proposition than simply making AI strategy generation more polished.
Exactly — that was the hard part to get right. It's tempting to let the
model "fill in the gaps" to make the output feel more complete, but that's
exactly where the hallucinations creep in. Building the guardrails to
catch that (and just say "insufficient basis" instead) took longer than
building the actual analysis pipeline. Appreciate you calling that part
out specifically.
That tradeoff is exactly where the credibility comes from. “Insufficient basis” is far more useful than a confident conclusion the evidence never supported.
Right there with you. "I don't know" is a much harder sell in a demo than
a confident-sounding number — but it's the only way the output stays
useful once real decisions start riding on it.
That’s exactly the distinction I was getting at. I have a few thoughts on the business side that are better discussed privately than unpacked in the thread. What’s the best email to reach you at?
hito1120@gmail.com — happy to dig into the business side. Looking forward to it.
Thanks! I’ve just sent it over.
Looking forward to hearing your thoughts whenever you have a chance.