Most advice about AI tells you what it can do. This is the other list, and it is shorter, more specific, and more useful, because the failures are predictable and expensive.
AI should not be used for marketing tasks that require knowledge of a specific customer, that involve regulated or unsupportable claims, that respond to an upset customer, or that set pricing. In each case the failure is not that AI produces poor writing. It produces confident, plausible output with no way of knowing it is wrong, and the cost of that error falls on the business.
The pattern behind all four
AI is never uncertain. That is the root of every item on this list.
A person who does not know something usually signals it. They hedge, they check, they say they will find out. A model states a fabricated figure, an unsupportable claim, or an inappropriate response in exactly the same confident register it uses for everything true.
Which means the danger is proportional to how good the rest of the output is. Fluent, well-structured, entirely wrong is harder to catch than obviously bad.
1. Anything requiring knowledge of a specific customer
A model knows how customers in your category generally behave. It does not know that this client has been with you six years, had a bad experience in 2023 that you fixed, and always asks about timelines first.
Do not use it for: personalised outreach to a named individual, responses referencing account history, or anything where being wrong about the relationship is worse than being generic.
Why it fails: it will invent plausible context. "As a long-standing customer" to someone who bought once. "Following up on our conversation" when there was none. Small errors, and they signal that nobody was paying attention.
The workable version: use it to draft the structure, then supply the specifics yourself. Never let it fill in facts about a person.
2. Regulated claims and anything you cannot support
Health, finance, legal, insurance, and anything with a professional body attached.
Why it fails: a model will produce a claim that sounds standard for the category because it has read thousands of similar pages, without any knowledge of what your licence, jurisdiction, or evidence permits you to say.
The output is not obviously wrong. It reads exactly like competitor copy, which is precisely why it slips through.
The workable version: write down explicitly what claims your business can and cannot make, and include that in every brief. This is one of the five sections of a brand standard and it is the one that eventually saves you real trouble.
Anything customer-facing in a regulated field gets human sign-off. No exceptions worth making.
3. Responding to an upset customer
Complaints, refund disputes, bad reviews, anything where someone is angry.
Why it fails: the output is usually technically appropriate and emotionally hollow. It acknowledges, it apologises, it offers to make it right, and it reads as a form response, which to someone already annoyed is worse than a slow reply.
There is also a judgment problem. Knowing when to refund, when to hold the line, and when the customer is right despite being rude is not a language task.
The workable version: AI can help you calm down a draft you wrote angry, which is genuinely useful. It should not generate the response.
Your reply to a bad review is read by every future customer. That is not a place to save fifteen minutes.
4. Pricing decisions
Why it fails: a model has no knowledge of your costs, your margins, your capacity, your local market, or what you can defend in a negotiation. It will produce a confident recommendation derived from general patterns.
It is also disconnected from the strategic question. Price signals position, and a number generated from an average has no view on where you are trying to sit.
The workable version: use it to structure how you present pricing, and to pressure-test your reasoning by arguing against it. Do not use it to produce the number.
The three-column exercise
The practical way to make this concrete for your business, and it takes ninety minutes.
List every marketing task your team does in a normal week. Sort each into three columns: AI does this, AI assists with this, AI stays away from this.
| AI does it | AI assists | Never |
|---|---|---|
| First drafts from a brief | Editing your writing | Complaint responses |
| Research and summarising | Structuring an argument | Named-person outreach with history |
| Adapting one piece to formats | Brainstorming angles | Regulated claims without review |
| Product and meta descriptions | Pressure-testing a decision | Pricing |
| Categorising and sorting | Improving a draft you wrote | Anything with an invented figure |
The arguments this produces are the actual value. Someone will insist customer replies belong in column one and someone else will explain why they do not, and that conversation builds shared judgment in a way no policy document does.
The escalation rule
Whatever you decide, write down one rule and make it non-negotiable: anything that reaches a customer with a claim, a number, or a promise in it gets a human read first.
Not a full review of everything. A read of the things that carry risk.
That single rule catches most of what would otherwise go wrong, and it costs a couple of minutes per item.
The one that actually costs money
If you take one thing from this: never let AI supply a number.
Not a statistic, not a percentage, not a benchmark, not a study result. It will produce plausible figures with total confidence and no source, and your name is on the page.
Every number you publish should have come from you or from a source you personally checked. That rule alone prevents the most damaging category of error, and it is the easiest to enforce, because you can search a draft for digits.
Eric Howard
Founder of SteelAIQ and a chief marketing officer with more than twenty years running marketing organizations across enterprise software, manufacturing, and consumer brands. Based in Pittsburgh.
