Where AI Doesn't Belong in Your Business (And What to Fix Instead)

Where doesn't AI belong in a business?

AI doesn't belong anywhere a decision hasn't been made yet: an unwritten judgment call, an exception nobody has authorised, a process that changes shape every time it runs. AI is excellent at producing output once the rule exists. Asked to guess the rule, it produces a confident answer that looks finished and quietly isn't.

Who is this for?

This is for the founder whose business already works and who feels the pressure to "add AI somewhere". You've tried a tool or two. One of them sort of works, if someone checks it. Another one everybody quietly routes around. And every week, somebody new tells you what you should automate next.

You don't need a longer list of places to add AI. You need the short list of places where it would make things worse, so the rest can actually hold.

Key takeaways

• AI performs well on tasks and badly on decisions nobody has made.
• A tool pointed at an undecided process runs the confusion faster, with a finished-looking output that hides the guess.
• The more a wrong answer costs, the more that decision needs a human owner before any tool touches it.
• Some judgment calls stay human on purpose, because that's what clients are paying for.
• Most processes hide one undecided piece inside an otherwise clear one. Find that piece, decide it, then automate the rest.

Table of contents

• Why nobody tells you where AI doesn't belong
• The difference between a task and a decision
• How to find where AI doesn't belong: 5 checks
• What this looks like in a real business
• Where AI genuinely earns its place
• Frequently asked questions
• Recommended tools and resources
• Final summary

Why nobody tells you where AI doesn't belong

Everyone selling a tool, a course or a template has a reason to tell you where to add AI. Almost nobody has a reason to tell you where to leave it out. So founders point AI at whatever hurts most this week, without checking what kind of problem it is.

The results show up at every size of company. MIT's Project NANDA looked at generative AI in business in 2025, through 52 executive interviews, a survey of 153 leaders and an analysis of 300 public deployments. It found that 95% of the pilots it studied had no measurable impact on profit and loss. Those were larger companies' pilots, and the methodology has its critics, so read it as a signal about the pattern, never as a forecast for your business. The pattern the researchers described is the one that matters here: the tools worked in the demo and broke inside the real workflow.

Here's a quick test you can run right now. Think of the last AI output someone in your business had to fix before it went out. Was the model wrong about the words, or was nobody sure what the right answer should have been in the first place?

That second case is the one this article is about.

The difference between a task and a decision

A task has a right answer that doesn't depend on who does it. A decision still needs someone to weigh context, exceptions or consequences.

Here's the same piece of work described both ways.

"Send the client update" as a task: take this week's delivered items, write them up in our usual format, send by Friday.

"Send the client update" as a hidden decision: work out what counts as worth mentioning, decide whether to flag the delay on the second deliverable, choose the tone for a client who's been unhappy.

AI does the first version well. Hand it the second, and it will pick a tone and decide what to mention, with total confidence, and nobody will know a choice was made until the client replies.

How to find where AI doesn't belong: 5 checks

1. Separate the task from the decision inside it.
2. Find who currently makes that call, and whether it's written anywhere.
3. Spot the process that's still being invented as it runs.
4. Price what a confidently wrong answer costs in that spot.
5. Name what stays human on purpose.

Check 1: Separate the task from the decision inside it

Write down what you're about to hand to AI. Underline every verb that involves choosing: decide, prioritise, judge, approve, flag. Whatever you underlined is decision territory. Whatever's left is a task, and a good candidate.

Check 2: Find who currently makes that call

If the answer is you, and nothing is written down, handing it to AI doesn't remove the judgment. It hides it inside an output that looks done. That's often worse than doing it by hand, because the guess becomes harder to spot.

Check 3: Spot the process that's still being invented

Any process that changes shape each time, because you're still working out what it should produce, isn't ready for AI. It's ready for a working session on what the process is for. Automating it early locks in this month's improvisation.

Check 4: Price the confidently wrong answer

Some mistakes cost one edit on a draft. Others cost a client, a price, a promise. The more expensive the wrong answer, the more that decision needs a named human owner first.

Check 5: Name what stays human on purpose

Some calls stay with a person because that's the product: taste, relationship, accountability, the read on a client's specific situation. Choosing those on purpose is a design decision. Avoiding automation out of habit is a different thing entirely.

✦ THE ONE THING THAT CHANGES EVERYTHING

The founders who get the most out of AI can say clearly where it doesn't belong yet, and why. That list is usually shorter than they expect. It's also what makes everything else trustworthy, because now the tools only run where a real decision already exists.

What this looks like in a real business

A client recently asked me whether moving his team onto a paid Claude plan would save the company money. It sounded like an AI question. We ran the actual numbers together, and the analysis showed there was no saving to be had. The useful work was the cost analysis, and the answer was to leave it alone.

Another client had been working around a constraint in the tool his team used every day, and assumed the tool was the limit. When we looked at it together, the real issue was that the order of data entry was the wrong way round. We corrected it live. No new tool, no AI layer, just a decision about how the work should flow.

And one from my own desk: AI still isn't reliable for precise formatting. When a document needs exact layout, the manual fix is often faster than the third prompt. Knowing that saves more frustration than any tool I could add.

In all three, the first question was never "which AI?". It was "what is actually undecided here?"

Where AI genuinely earns its place

Quick verdict: AI earns its place wherever the decision is already made and the work is producing output from it. It doesn't belong wherever the decision itself is still open, still political, or still needs a human owner.

It earns its place: drafting from an approved brief, summarising something already understood, applying a documented rule at speed, turning your written process into consistent first drafts.

It doesn't belong (yet): deciding what the rule should be, settling a disagreement about priorities, approving an exception nobody authorised, anything client-facing where a wrong call costs the relationship.

Frequently asked questions

Where does AI not belong in a business?

Wherever a decision hasn't actually been made yet: an unwritten judgment call, an exception nobody has authorised, a process that still changes shape every time it runs. AI performs well once the decision is clear, and badly, with confidence, when asked to guess one.

How do I know if something is a task or a decision?

Ask whether the same input would produce the same correct output every time, without anyone needing to think it through. If yes, it's a task. If someone still weighs context or exceptions, it's a decision, and it isn't ready for AI yet.

Isn't AI supposed to help me make decisions too?

It can help you think through options when you're the one deciding. It shouldn't be the one deciding, especially where a wrong answer is expensive: pricing, client exceptions, anything client-facing with real consequences.

What's the risk of automating an undecided process?

You encode a guess and call it a decision. It runs fast and consistently wrong, which is harder to catch than a slow manual mistake, because the output looks finished.

Does this mean I should avoid AI until everything is documented?

No. Document the specific piece you're about to hand to AI, never the whole business first. Most processes have one undecided piece hiding inside an otherwise clear one.

What should never be automated, even later?

The judgment calls your clients are paying for: taste, relationship, accountability, a human read on their specific situation. Naming those on purpose is a design choice.

Why do so many AI projects fail to deliver?

In MIT's 2025 research on larger companies, the pattern was tools that looked good in demos and broke inside real workflows. The organisational side, meaning who decides what and how the work flows, mattered more than the model.

Can a business systems audit show where AI doesn't belong?

Yes. That's most of what it looks for: the decisions hiding inside tasks, which ones are ready to hand off, and which ones still need a human owner.

What's the fastest way to test this on my own business?

Pick one place you already use or are considering AI. Ask who makes the judgment call inside it today, and whether that call has ever been written down. If it hasn't, decide it first.

Does having a team change where AI belongs?

It changes who owns the decision. A team without a written rule spreads the same undecided call across more people, and adding AI on top spreads it faster.

Recommended tools and resources

• One written document per process where the decision behind it is recorded once, before any tool touches it.
• The workflow system you already run, checked for whether it can hold a written rule, before you add an AI layer on top.
• A short standard operating procedure once a decision has held across a few real cycles, so it survives outside your head and outside any prompt.
• A "stays human" list, one page, naming the calls you keep on purpose.
• The Studio, if you'd rather separate the decision from the tool with someone in the room: one working session on your real business, €750, credited in full against L'Atelier if you go on.

Where to explore next

• What Is a Business Systems Audit? (And What It Actually Covers)
• Hire vs Automate: How to Tell Which One Your Business Actually Needs
• How to Delegate Without Losing Control
• How to organise a business that still depends on you
• The Studio: a working session on your real business

Final summary

AI doesn't belong where a decision is still open. It belongs where the rule exists and the work is producing output from it. Separate the task from the decision, find who owns the call, spot the process still being invented, price the wrong answer, and name what stays human on purpose.

Do that first, and every place you do use AI becomes something you can trust instead of something you have to check.

Want someone to look at where AI actually fits in your business?

If you can feel that something is being automated too early, but can't yet see which decision sits underneath it, that's the conversation to have before your next tool. Book a call.