When NOT to Use AI: An Installer's Honest List

By GetGood · July 16, 2026 · 5-minute read

Why is the installer telling you where to stop?

We set up AI for business owners for a living. You'd expect the list of things it can't do to be short around here, and you'd be right about our bias — read the rest of this accordingly. But here's what we've learned watching setups either stick or die: the fastest way to make an owner abandon AI completely is to aim it at a job it was never going to do well, let it fail in front of a customer, and burn their trust in the whole idea.

The setups that last are narrow ones. A few chores handled well, a person still approving anything that leaves the building, and a clear line around the work that stays yours. Knowing where that line sits isn't a footnote to installing AI. It's most of the skill.

What are the jobs to keep away from AI?

Six of them, in the order they tend to bite people.

1. Anything expensive, public, and permanent. A wrong quote a customer already accepted. A published claim about your business. A message to your whole list. AI is confidently wrong sometimes — not often enough to notice in testing, often enough to matter at scale. Cheap and reversible mistakes are survivable; the other kind isn't worth the minutes you saved.
2. The moments where the point is that you showed up. The apology after you dropped the ball. The condolence note. The thank-you to the client who's been with you eleven years. AI can produce those words in seconds, and the words are worthless — an apology means something because a person spent the time. Remove the person and you've removed the whole product.
3. Answers you have no way to check. This is the sneaky one. If you ask about something you know well, you'll catch a mistake instantly. If you ask about something you don't — a regulation, an unfamiliar market, a number you can't verify — you have no defense, and a fluent wrong answer looks exactly like a right one. The danger isn't AI's error rate. It's the gap between how sure it sounds and how sure you can be.
4. Decisions the rules require a licensed human to make. Medical, legal, financial, and compliance calls in your field may well require a qualified person — not because AI can't string together a plausible paragraph, but because someone has to be accountable for it. What applies to you specifically is a question for your own advisor, and we're not going to pretend otherwise.
5. Judgments about a specific person. Who to hire, who to let go, who gets the exception, who you believe. These deserve a human who can be asked "why?" and give a real answer — and who carries the weight of having decided. AI can organize the information. It shouldn't render the verdict.
6. One-time tasks. Not a risk, just a waste. If a job happens once, the time you spend explaining it well enough to get a usable result usually exceeds the time to do it. AI pays off on the thing you do every week, forever. Automation's return comes from repetition, and something you'll never repeat has none to give.

Where's the gray zone?

The list above is the easy part. The genuinely tricky cases are the ones where AI is almost right for the job.

The customer reply that's mostly routine. Nine in ten are "what are your hours" and AI handles them beautifully. The tenth is a furious regular, and the machine can't tell the difference — it just answers cheerfully. This isn't an argument against drafting replies; it's the argument for the approval rule. Let it draft, keep your eyes on the send button, and the tenth email stops being a hazard.

The proposal with a number in it. AI writes a fine proposal and gets the price wrong in a way that reads perfectly. Draft the prose, enter the numbers yourself, and check the total before it goes out.

The task that's obvious to you. Half of "AI can't do this" is really "AI doesn't know this" — it's guessing at what you never wrote down. Before you conclude the tool failed, check whether you gave it the context it needed to succeed. That's a fixable miss, and it's the most common one.

Notice the shape: none of these get solved by a cleverer prompt. They get solved by putting a person at the exit.

How do you test a task yourself?

Three questions, and you can run them in about ten seconds:

1. If this comes out wrong, what does it cost — and can I take it back? Cheap and reversible: go. Expensive and permanent: your hands.
2. Can I tell whether the answer is right? If you can't check it, you're not delegating — you're gambling with extra steps.
3. Does anyone need a human to have done this? If the answer is yes, no amount of quality in the output will fix it.

Three green lights and it's a fine candidate — likely one of the good first automations: the repetitive, low-stakes, customer-invisible chores. One red light and it stays with you, and that's not a failure of the tool. A hammer that isn't a screwdriver isn't a bad hammer.

What's the actual trade?

Here's the honest frame we'd want an owner to leave with. AI is very good at the work that drains you — the drafting, the sorting, the remembering, the second version of a thing you already wrote. It's not good at the work that defines you: the judgment, the relationship, the call only you can make.

That's not a consolation prize. That's the entire point. You maintain the relationships; the machine handles the rest. An owner whose AI writes their condolence notes has automated the wrong end of the business — and quietly given away the part customers actually stayed for.