When someone tries an AI that writes in their name, the first question is nearly always voice. Does it sound like me? It’s the wrong question to ask first. A wrong word is easy to catch: you read the draft, wince, and change it. A wrong claim is not. It reads fine, it goes out under your name, and you find out when the person who knows better reads it.
A language model knows exactly what a persuasive marketing piece looks like. It opens with a client conversation, backs the point with a number, and ends on a confident view of this week’s news. Without the real material, it writes the shape anyway. The client is plausible. The number is round. The view sounds like something you might hold. Then the client you never spoke to reads it, or the colleague who knows the real figure, or the prospect who asks you to say more about an opinion that was never yours.
This is the most expensive mistake an AI can make for you, and almost nobody shopping for one asks about it.
Most tools treat it as a writing problem
The industry’s answers fall into three camps, and each one solves half the problem.
Tell the model not to. Put “never invent facts” in the prompt and hope. It helps, and we do it too. But an instruction is a request, not a control. A model under pressure to write something good will still reach for the plausible detail.
Ask another AI to check. Run the draft past a second model that looks for made-up claims. That’s a model checking a model. It shares the failure it’s checking for, and it can give a different answer to the same text tomorrow.
Leave it to you. Every draft goes to you for approval, and you’re the fact-checker. That works until you’re reading your fortieth draft on a Friday afternoon. Approval is a signature, not a search. If the tool hasn’t told you which line has nothing behind it, you’re looking for a needle you don’t know is there.
We made it the core of the product
Groundworks AI is an AI CMO: a marketing function you manage, not a writer you prompt. A function you manage has to be one you can trust with your name. So we built it on one principle, and we built everything else around it:
A piece says only what its substance supports.
A fact needs a source. A view needs someone who actually holds it. A customer is named only when you’ve said they can be. Here’s what that changes, claim by claim.
| The claim | What a typical AI writer does | What Groundworks AI does |
|---|---|---|
| A figure | Writes the round number that fits the sentence | Holds the draft until the figure has a source, or you record why it’s fine |
| A customer’s name | Names whoever makes the story work | Names only customers you’ve cleared. Everyone else is “a customer in” their sector |
| An opinion on the news | Gives you a confident take | Reports the news, and waits for your view before taking a side |
| Your experience | Invents the client who “told me last week” | Writes a complete line and asks you for the specific |
| A polished rewrite | Adds detail to make it read better | Throws away any rewrite that adds a fact |
A fact needs a source
You tell the CMO what it can and can’t say, in three lists, and the lists are enforced where a draft is approved and published.
| List | What goes in it | What happens in a draft |
|---|---|---|
| Allowed | Claims you’re happy to make | The CMO uses them |
| Needs proof | Claims that are true but need evidence behind them | Flagged for you. When the CMO drafts from its record, held until the proof is there or you release it with a reason |
| Never say | Lines you never want in your name | The draft can’t be approved or published until the line is gone |
On top of the lists, when the CMO drafts from its record and a figure or a name has no source behind it, the draft is held. Three ways out: add the proof, change the line, or release it with a reason, and the reason is logged. Nothing slips out on a shrug.
A view needs someone who holds it
A take is the easiest way to make a news post feel personal, which is exactly why AI invents opinions so freely. Ours reports the news for you, and it won’t invent your opinion of it. Taking a side waits until you’ve given it one.
The same goes for your experience. Where a piece would be stronger with something only you know, it doesn’t make the moment up. It writes a complete line and asks you for the specific. Skip it, and the post still stands. Answer it, and the post becomes yours.
A customer is named only when you say so
Every customer you list is marked as nameable, not nameable, or not decided yet. We don’t guess from your website. Until you decide, the answer is unknown, and unknown blocks: the customer appears as “a customer in” their sector, and a case study about them waits for you.
Why the check is code, not AI
We made four calls here, and each one was deliberate.
The checks are plain code. The same text gives the same answer every time, and when in doubt they hold. No model gets a vote on whether another model told the truth.
The prompt is a help, not the control. We still tell the writer never to invent a client, a quote or a number. The block sits on the approve and publish path, where an instruction can’t be ignored.
The AI never decides what’s true. It can suggest that a claim be banned. A person on your team decides.
We killed our own quality score. We built a score for how substantive a piece was. A reader panel couldn’t agree on what it meant, so we retired it rather than publish a number we couldn’t defend. We hold ourselves to the same rule we hold the drafts to.
What you get
You stop reading drafts as a fact-checker and start reading them as the person who signs them off. The lines that need you are flagged, the lines with nothing behind them are held, and every release has a reason on the record. The checks catch figures, names and the lines you’ve listed; a cleverly reworded line can still reach you, which is why approval stays yours. And we’re extending the same check to every kind of piece the CMO makes.
The result is marketing you can put your name to without wondering what it said. A client you never met, a number you can’t stand behind, a view you don’t hold: none of them goes out under your name without getting past you first.
One question for any vendor
If a draft states a number or names a client you never gave it, what stops it being published?
See what it writes from what you know
Give it your website and a short interview. Every draft is built from what you’ve given it.