Ask any AI marketing vendor whether their product invents clients, quotes or numbers, and they’ll say no. They mean it. But “no” can be kept in very different ways, and most buyers never ask which one they’re getting.

In most products, the guardrail is a sentence in the prompt: never invent facts. That sentence is real, and it helps. It’s also a request. A language model weighs it against everything else it’s been asked to do, including “make this persuasive”, and most of the time it complies. Most of the time is fine for a draft you’ll rewrite. It isn’t fine for something published under your name to the people you most want to trust you.

The same promise, kept four ways

When we design a rule in Groundworks AI, the first question isn’t what the rule says. It’s how strongly it’s kept. There are four answers, and “never invents a client” means something different under each.

How it’s keptWhat that meansHow far you can lean on it
InstructedThe prompt asks the model to follow the ruleUsually holds. Nothing stops it when it doesn’t
ReviewedA person approves before anything goes outAs good as the reviewer’s attention on the day
WithheldThe model is never given the thing it mustn’t useIt can’t misuse what it never saw
BlockedCode refuses the outputSame answer every time, whatever the model wrote

Most tools stop at the first two: tell the model, then hand the draft to you. That’s half a solution. The model is asked to be careful, and you’re asked to catch it when it isn’t, with no help finding the line that has nothing behind it. We use all four, and we put the rules that matter most at the strong end.

What we block

Blocking happens in code, on the path a draft takes to approval and publishing. No model is consulted.

  • Lines you’ve marked never to say. A draft that contains one can’t be approved, published or marked as posted until the line is edited.
  • Figures and names with no source. When the CMO drafts from its record and states a figure or a name that none of its sources carries, the draft is held. You add the proof, change the line, or release it with a reason, and the reason is logged.
  • Rewrites that add facts. When a draft is revised to read better, a revision that adds a fact fails in full, and the draft comes back as it was.
  • Opinions on the news you haven’t given. A news-led post carries no stance unless it’s built on your own experience or a view you’ve given.

What we withhold

The surest way to stop a model misusing something is to never hand it over. Claims you’ve marked as needing proof aren’t given to the writer at all. Proof that’s gone stale, and figures nobody has confirmed, stay out too. A customer you haven’t cleared to be named from your proof list reaches the writer only as “a customer in” their sector. It can’t name them, because it doesn’t know who they are.

What we still instruct, and what we ask you

We still tell the writer never to invent a meeting, a person, a quote, a customer, a result or a number. When a linked page can’t be read, it’s told not to claim it read it. The instruction is where the model starts. It isn’t where our guarantee comes from.

And some things only you can supply. Where a piece would be stronger with your own experience, it writes a complete line and asks you for the specific. That’s review used for what review is good at: adding what only you know, not hunting for what the AI made up.

The calls we made

No AI checks the AI. The checks are deterministic. The same text gives the same answer every time, and when in doubt they hold. A model checking another model’s claims shares the failure it’s checking for, and it might answer differently tomorrow.

The prompt is a help, not the control. That phrase is written in our own code, next to the instruction, so nobody mistakes one for the other.

The AI proposes; your team decides. The agent can suggest that a claim be banned. Only a member of your team can ban it.

Every strength is stated, not implied. Inside the company we mark each rule with how it’s kept. A rule we only instruct is never described to you as one we enforce.

Where the edges are

Code is exact, which is its strength and its limit. The checks catch figures, names and the lines you’ve listed, plus close matches. A banned line reworded cleverly enough can reach you, and a name you mentioned in your own website or notes isn’t withheld. That’s why approval stays with you, and why we keep moving rules toward the strong end of the table. Extending the same checks to every kind of piece the CMO makes is the work in progress now.

What you get

A guardrail you can reason about. When we say a never-say line can’t be published, that isn’t a hope about model behaviour. It’s a rule that runs the same way on every draft, on the worst day as on the best. You spend your review on the lines that need you, because the lines that can’t go out already don’t.

One question for any vendor

For each thing your AI “never” does, is that enforced in code, or only asked of the model?

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.

Start your free week More Insights