How we keep the AI from making things up.

When you imagine handing your marketing to an AI, the fear usually isn't that the prose will be clumsy. It's that it'll be confidently wrong. It'll state a price you don't charge. It'll name a feature you don't ship. It'll invent a statistic that sounds plausible and put it on your pricing page under your name. That specific failure — fluent, confident, false — is the one that ends the trial.

It's a real risk, and pretending otherwise would be the wrong way to earn your trust. So the design question that mattered most wasn't how to make the writing good. It was how to make sure the AI never states as fact something it doesn't actually know.

So we split everything the AI knows about your business into two kinds, and treat them differently. Facts are exact values — your price, your product names, your launch date, the number of customers you'll actually claim. When the AI needs one, it doesn't recall it from a blur of training; it retrieves the specific value you gave it and uses that. Facts are for quoting.

Context is the background — your foundation document, your voice, old blog posts, the way you talk about the problem you solve. The AI reads context to understand how to sound and what matters, but it doesn't quote it as if every sentence were gospel. Context shapes the writing. Facts get stated. Keeping those two jobs separate is the difference between an AI that sounds like you and an AI that impersonates you badly.

Both kinds come with a second switch: public or private. A public fact is one the AI can put on a page — your price, your product name. A private fact is one it can use to make a decision but must never print — a margin, an internal target, a customer's name you weren't given permission to publish. Same with context: some is fair to draw on openly, some is background it should learn from and never leak.

That's what lets you hand over real, sensitive material — the strategy notes, the honest internal framing — and trust that the AI gets smarter from it without ever spilling it into a blog post. It reads everything you give it. It only says the parts you marked as sayable.

I won't oversell this. Underneath, the AI is still a general-purpose model, and a model can still phrase a true fact in a misleading way, or reach for context when it should have reached for a fact. Grounding the AI in your real values makes the confident-and-wrong failure far rarer. It doesn't make it impossible.

Which is exactly why the approval gate isn't going anywhere. The grounding is what makes the drafts trustworthy enough to be worth reviewing quickly. Your eyes are what make them safe to publish. The two work together: the AI stays close to what it actually knows, and you stay the last check before your name goes out. That's the honest version of keeping an AI from making things up.