An owner wanted to move her whole website onto a new platform. I told her to revert to the old one first.
Her reason for wanting the migration, as she gave it to me: she'd heard everyone talking about that platform on her coach's group calls. That's a fair reason to be curious about a tool. It is not a diagnosis of her website.
I build AI systems for coaches and consultants. It's the whole business. And the more of it I do, the more I think the skill that actually separates people is the invisible one: knowing when not to use AI, and when not to use the new thing at all.
TLDR
- "Everyone is talking about it" is a reason to be curious. It is not a reason to rebuild anything.
- AI is a very good editor and a very bad strategist. Most people use it the other way round.
- The day my own system started inventing rules is the day I closed its autonomy for good.
- There are 4 places I keep AI out of on purpose, and 1 of them is my own marketing budget.
- Restraint doesn't come with a product attached, which is why nobody pitches it.
The Owner Who Wanted a New Platform
What I proposed to her instead of a full rebuild: our web person builds just the homepage on the new platform and hosts it on her domain. The rest of the site was functional and stays as is. Then I shared a custom image tool I'd built so she could run her own photo pass, and I'll make her a short hero video after.
So the scope went from "migrate everything" to 1 page plus an image pass. That's the entire intervention.
I may have drawn the line in the wrong place. I have no measurement of what her old site was costing her, so I picked the smallest change I could defend and left the working parts alone. That's a preference, not a finding. But I'd rather defend a small change than sell a big one I can't justify, and I think that's the whole job more often than people admit.
AI Is a Good Editor and a Bad Strategist
Here's the split I use, and I use it on myself first.
Give a model a draft and a standard and it will tell you, reliably, what's off. Too long, wrong register, a claim that isn't supported, a hook that's table stakes. It's a good editor. Given a reference library of your own work it becomes a very good one. That's most of what I build.
Ask it what your business should do next and you get the statistical average of every business article it has read. Confident. Well-formatted. Generic. I've written about why agents have no opinion about quality, and strategy is the extreme case. The model doesn't know which of your rules is load-bearing. It doesn't know that your regulars come for the thing you're about to cut.
Most people run it backwards. They ask AI to decide and then edit the decision by hand. Decide by hand. Let it edit.
The Day My System Invented Its Own Rules
This one cost me a weekend, so I'll be specific.
I rebuilt a report system to pull from a bigger set of proven levers. It worked, mostly. Then the coach who reviews the reports caught something I'd missed. The system had started writing new rules that were never ours. It suggested killing birthday and anniversary comps. It suggested flat pricing. Both sound reasonable. Both quietly destroy the exact value the method protects, and either one could hurt a real client.
I'd been arguing for giving the model a little room to reason from research instead of just repeating our tested set. He shut it down cleanly: it cannot have autonomy, because then it's guessing. The whole point of the method is that we already know what works, so we only do those things.
He was right. A model given room to reason will reason its way past your guardrails. Every lever in our set earned its place by being tested on a real business. A clever new idea has not. So the lever set stays closed, and research is a manual trigger, not a standing permission. I built the thing that went off the rails, and the fix was less autonomy, not more.
4 Places I Keep AI Out on Purpose
Concretely, for the coaches and consultants who want a list:
- Strategy. Which offer, which audience, which price, which thing to cut. AI can lay out options. The decision is yours, and the model should never be allowed to make it by default.
- Anything a client can't control after you leave. I overrode 2 levers in my own library this month. Both were server scripts, lines the staff should say. They'd produced results. They came out because a menu change holds after we're gone and a line a server says doesn't. If the output depends on someone else's behaviour, don't automate the recommendation.
- The first draft of anything new. A format you've never made. Write the first one yourself, badly. Then it becomes the reference. Skip that and you're running a template with your name on it.
- My own marketing budget. I've never spent on paid ads. Not once. My prospecting tools are all free tier. I'll add AI tools when something is working organically and needs to go faster, not before. The tool is not the strategy.
That last one is uncomfortable to admit publicly, and the reason is the same one that keeps the best marketers I know from marketing themselves: it sits last on the list. But the principle holds: don't buy the system before you need it.
The Restraint Nobody Sells You
Nobody pitches "do less." There's no product in it. Every AI tool on earth is sold on the premise that you should be using more of it, and the people selling it are not wrong about their incentives.
But the owner whose site I didn't rebuild got a working site and 1 better page. The report system I locked down produces reports a coach will sign. My own business runs on a cloud subscription and a few odds and ends. None of that is a system I bought. It's a series of things I decided not to do, and I think it's most of why the things I did do actually work.
So the skill is invisible because it looks like nothing happened. That's fine. Nothing happening was the point.
Look, you already know the tool you're being pitched isn't a diagnosis. You've known that for a while. The part you're stuck on is what the smallest defensible change would be, and that's a 30-minute conversation, not a rebuild.
That's the conversation I have with every client before we build anything. Book a free call and I'll tell you honestly which parts of your setup should be left alone. Or, if content is the part that actually needs the work, send me 1 piece of it and see what comes back in 48 hours.
FAQs about when not to use AI
When should a coach or consultant not use AI?
For strategy decisions, for recommendations that depend on someone else's behaviour after you leave, for the first draft of a format you have never made, and for anything you have not measured the problem behind. AI edits reliably against a standard. It decides badly, because it defaults to the average of everything it has read.
Is AI good at business strategy?
No. It is good at laying out options and very good at editing a decision you have already made against a standard. Asked to choose, it produces the generic choice with confidence. Decide by hand. Let it edit.
Should AI systems be allowed to reason beyond their rules?
Not by default. A model given room to reason will reason past your guardrails, because it cannot tell which rules are load-bearing. Keep the tested set closed and make research a manual trigger you fire on purpose, not a standing permission.
How do I know if I actually need a new tool?
Ask what the current setup is costing you, in a number. If you cannot name one, you do not have a diagnosis, you have curiosity. Curiosity is fine. Pick the smallest change you can defend, leave the working parts alone, and measure before you buy the rest.