Why AI marketing teams make things up — and the two gates that stop it

Ask any team that has put AI into their marketing what went wrong first, and the answers rhyme: a pricing line that was almost right, a feature the product does not have, a statistic with no source that sounded plausible enough to ship. The model did not lie — it did what models do: it completed the pattern. The failure is that nothing stood between the completion and the publish button.

Hallucination is a workflow bug

Treating hallucination as a model problem leads to model-shaped fixes: better prompts, bigger context, a "be accurate" instruction. These lower the rate and change nothing about the outcome, because a lower rate still ships falsehoods — just less often, and less predictably.

Treating it as a workflow bug leads to a different question: what would a competent human department do? It would not let the copywriter fact-check their own copy. It would keep an authoritative record of what the company may claim. And it would make sign-off someone's explicit job.

The structure that stops it

A generated Growics department ships with that structure built in:

• Truth files, with provenance. Every company fact lives in a file with a source and a date. A fact without a named source is not a fact. The interview seeds these files on day one, and your reviewer owns them.
• "Needs verifying" beats an invented answer. When a fact is not in the truth files, the system says so — outward-facing text never gets an improvised yes.
• A never-say list that is enforced, not suggested. Claims your company must not make are checked mechanically, not remembered optimistically.
• Quarantine for new facts. Anything learned mid-task from an unverified source waits in a holding area, invisible to the team's memory, until a human approves it. Useful information gets in; contamination does not.
• The facts gate outranks the writer. Before anything publishes, a separate seat — the Product Expert — checks every claim against the truth files. A fail comes back with a note naming the claim, and the deliverable does not move until it passes. The quality gate then judges everything else: structure, voice, fit.

The two gates are not optional and not configurable away. They install in every generated team, and the telemetry they produce — pass rates per agent, per gate — becomes the team's strength score, computed from evidence rather than self-graded.

What this costs, honestly

Gates add latency: a deliverable that fails the facts gate goes back for rework instead of going out. That is the point. The cost of a re-draft is minutes; the cost of a false compliance claim in a paid ad is a different order of magnitude, and it compounds — retractions, refunds, and the quiet damage of a customer who caught you once.

The bet behind Growics is simple: in marketing, trust is the product. A department that occasionally invents things is not a cheaper department — it is a liability with good grammar. Structure, not vigilance, is what makes the difference durable.

See the gates standing in the live demo — including a real failed verdict and the note that sent it back.