When not to use AI in your marketing

I build with AI constantly. That's exactly why I think the honest version of this topic is worth writing: the failures are more instructive than the wins, and they're less commonly published.

Where it genuinely doesn't belong

Anything requiring accountability you can't verify.

If you can't check the output, don't automate it. AI-generated financial figures, legal wording, medical claims or compliance statements are risks with a plausible tone attached.

The failure mode is specific: confident, well-formatted, wrong. That combination survives casual review, which is what makes it dangerous.

High-volume public content with no editing.

AI can produce fifty blog posts. It cannot produce fifty blog posts anyone wants to read, and publishing them tells visitors exactly what you think of their time.

The bottleneck was never production speed. It was having something worth saying.

Customer communication where tone carries the relationship.

Apologies, difficult conversations, anything where the person needs to feel heard. AI-drafted apologies read as AI-drafted apologies, and that reads as not caring.

Strategic decisions.

A model can structure your thinking usefully. It cannot know your market, your cash position or your risk tolerance. Asking it to decide is outsourcing judgement to something with no stake in the outcome.

Deterministic tasks.

If a rule can do it, a rule should. Using a model to reformat a date is slower, costlier and less reliable than code. I see this constantly in automations — an expensive, non-deterministic step doing work a regex would do perfectly.

Where it's genuinely strong

To be balanced:

• First drafts at volume — ad variations, subject lines, outlines
• Structured extraction from messy input
• Summarising long material
• Critique of your own work — reliably better than generation
• Explaining unfamiliar concepts while you learn
• Writing and debugging code with review

The pattern: strong where output is checkable and errors are cheap.

The trap I fell into

Early on I automated things because I could, not because they needed automating. I built a workflow that saved perhaps twenty minutes a month and took a day to build and several evenings to maintain.

The maintenance cost is the part people forget. Automations break. Platforms change. Every workflow is a small ongoing liability, and a workflow saving trivial time is a net loss forever.

A test before automating anything

1. Can I describe the process precisely? If not, I don't understand it well enough to automate it.
2. Can I verify the output? If not, don't.
3. What happens when it's wrong? If the answer is serious, keep a human in the loop.
4. Does the time saved exceed build plus maintenance? Be honest.
5. Would a rule work instead? Then use a rule.

What I'd say to someone starting

Use AI heavily for things you can check. Use it lightly for things you can't. Never use it for things you can't check and can't afford to get wrong.

That covers most of it, and it's a more useful frame than any tool list.