A customer looks at something your team produced and says it reads like AI wrote it.
The instinct is to go fix the tells. Strip the em dashes. Kill the stock phrases. Add guardrails so the model stops sounding like a model.
Do all of that. It is worth doing, and it takes an afternoon. Then understand that it will not solve the problem the customer is describing.
They Are Not Reacting to the Formatting
When people say something reads like AI, formatting is the evidence they reach for, because it is the part they can name. What actually triggered the reaction is that the work reads generic. It could have gone to anyone.
Generic is not a model failure. It is what any writer produces, human or otherwise, when they do not know enough about the specific situation in front of them. Hand a talented person one half-answered form about a company they have never spoken to and ask for a strategy document, and you will get something that reads exactly like AI wrote it, because the constraint was never the writer.
The AI did not fail you. The missing context did.
What Context Actually Means Here
Context does not mean a longer prompt. It means knowledge of this customer that would not be true of any other customer.
Which is worth being concrete about, because there is a hierarchy and most operations only touch the top of it.
What they told you when you asked. The intake form, the questionnaire, the onboarding survey. The thinnest layer, because people edit themselves when they write, and because a form only collects what you already knew to ask.
What they told you in conversation. The kickoff, the sales call, the strategy review. Richer, because you could follow up, and because the things people say out loud are less curated than the things they type into a box.
What they told you without being asked. The offhand remark forty minutes into a call. The support ticket that reveals what they actually use the product for. The thing they complained about once and never repeated. This is the most valuable layer and almost nobody feeds it into anything.
The Test
Before something goes to a customer, ask one question about it.
Could this have been sent to a different customer with the names swapped out?
If yes, that is exactly what they are going to react to, and no amount of tone adjustment will fix it. If no, if there are three or four specifics in there that only make sense for this account, then nobody is going to accuse it of being machine-written, whether or not a machine wrote it.
The Uncomfortable Version
Here is the part worth sitting with rather than deflecting.
If a customer says your work looks like AI, the finding is not that you used AI badly. The finding is that you did not know enough about them to produce anything else.
That gap existed before you automated. Automation did not create it. It industrialised it and then showed it to the customer at speed, which is why it feels new.
Fix the input and the output stops reading generic, because it stops being generic. That is not a prompt problem and it is not a model problem. It is a listening problem with a delivery schedule attached.
If you want to find out where your own operation is producing work faster than it is gathering context, the audit takes about two minutes.
