My AI strategy tool fabricated the price of my own product on its home page

I opened the dashboard for the thing I've been building — Sensei, an AI strategy advisor for solo founders — and the hero quote read:

"You've built the $15K strategy call as a $99/mo product — now prove the weekly insights are worth it."

Cute line. Confident. Stops you when you read it.

It's also wrong. Sensei is $29/mo for founding members (1st 50 members) , $49/mo for everyone else. There is no $99/mo tier. There is no $15K consultant I'm benchmarked against. The model invented both numbers and shipped them to the most prominent surface in my own product.

If I'd taken a screenshot for a launch post and shipped it, the very first thing every founder reading my product would see would be a lie about my own business.

This is what's actually wrong with AI strategy tools — including the ones you and I both pay for. Tell an LLM to "be specific" and it will be specific. It will invent the specifics. Tell it to "give actionable advice" and it will give actionable advice — it will recommend doing things to numbers that don't exist. The fluency masks the fabrication. Most founders never check.

I caught this one because I was looking at my own dashboard. How many of these has Sensei shipped to other founders' dashboards? More than zero. That's the part that wakes you up.

So this week I rebuilt the layer underneath. Four pieces:

The grounding contract. Before the LLM is told what to write, it gets a contract block: "You may only make factual claims that trace to (a) the founder's stated facts, (b) live data from a connected integration, or (c) public knowledge about a confirmed competitor. Anything else, you say less." Then a structured FOUNDER FACTS block lists the finite set of things the model is allowed to cite — pricing on file, ICP, stage, MRR, integrations connected, competitors confirmed. If pricing isn't on file, the block reads "Pricing on file: NONE STATED — do not invent tiers or dollar figures."

The verifier. After the model returns, a regex pass catches every dollar figure, customer count, retention percentage, conversion rate, market-size claim, and unconfirmed competitor citation. Each one gets severity-ranked: flag (close miss, founder can correct), demote (downgrade confidence chip), drop (null the field — the model is making it up).

The data-model fix. The deeper problem was that my onboarding form captured a price band ("$20-100") and a price model ("freemium") but never an explicit pricing-tiers free-text field. The model had nothing to anchor to, so it invented. Added the field to onboarding, the settings modal, the validate route, the background-analysis route. Founders enter their tiers; the next analysis reads against them.

The inline correction. When the verifier flags a price the founder hasn't put on file, Sensei surfaces it inline in chat — small amber line, "Sensei flagged 1 claim." Click → expand → "Correct this." A text area appears pre-loaded with an example. Type the real tiers, hit Save, the canonical pricing field updates, and the verifier goes silent on that claim shape forever. Three clicks from caught fabrication to fixed source data.

The deeper take. The mistake every AI tool makes is assuming fluency equals correctness. It doesn't. A confident sentence about your business that's wrong is worse than no sentence at all — it costs trust the moment you notice, and most founders don't notice until a real customer points it out.

The fix isn't a smarter model. The fix is an opinionated layer that says "when in doubt, say less." That's not a constraint on the model's intelligence; it's a constraint on its confidence.

For founders building anything on top of LLMs: the grounding contract is ~100 lines of code. The verifier is another ~150. The data-model fix took three hours. The rebuild took a day. Worth it the moment you catch your own product fabricating about itself on a screenshot you almost shipped.

Sensei is free during the founding-member beta — closes May 31. Try it out!