If you use AI to write product copy, scripts, or replies, you will eventually see it state something that isn't true: a guarantee the product never made, a result nobody promised, or a brand name that doesn't exist. Telling the model "don't make things up" in the prompt helps a little. It doesn't solve the problem. In the AI writing pipeline we built for one of our own products, the fixes that actually worked all live in code.
1. Check the output against the source data
Keep a list of risky phrases for your domain: absolute claims like "guaranteed" or "100%", miracle words like "permanent" or "instant", and promised results. After the AI writes, scan the text. If a risky phrase appears and it is not in the product data you provided, the output is rejected. The important part is the comparison with the source: if the product data really says "waterproof", the word is allowed. If it doesn't, it isn't.
2. Catch invented names
Models like to make text sound specific by adding brand or product names. A simple rule catches most of this: a capitalised name in the middle of a sentence that doesn't appear in the input data is treated as invented. Keep a small allowlist for words that are legitimately capitalised, and reject the rest.
3. Reject, explain, retry, then stop
When a check fails, don't just regenerate blindly. Send the draft back with a short note naming exactly what was wrong ("the word 'permanent' is not in the product data") and ask for a rewrite. Allow a few attempts. If it still fails, block it and show a clear error instead of passing a bad result to the user. A visible failure is better than a confident mistake.
4. Examples beat rules
We learned this the slow way. Changing the written rules in a prompt often made little difference, because the example scripts we included were stronger than the instructions. When we replaced the examples with ones that showed the structure we wanted, the output changed immediately. Curate your examples as carefully as your rules, and retire old examples when your standards change.
5. Check the final result, not just the text
If the text goes through another step, check that step too. In a voice pipeline, for example, the generated audio can skip a sentence. We transcribe the result and compare it with the script, and any take where part of the script wasn't spoken never reaches the user. The principle applies to any automation: verify the thing the user actually receives.
What these checks don't cover
Code-level checks catch explicit claims. They are weaker at implied ones, like a sentence that suggests a benefit without using a risky word. For those, the practical answer is to let a person edit the draft before it's published, and to make editing easy. Good automation doesn't remove human review. It makes review faster and focused on what matters.
Share a short description of your business and what you want to improve. We'll get back to you.