When NOT to Use AI: An Honest Guide

One rule covers most of it

If you cannot quickly tell whether the output is right, do not use AI for it.

That single test resolves the majority of cases. These systems produce confident, fluent, plausible material regardless of whether it is correct, and the fluency is the danger. Wrong output does not arrive looking wrong. It arrives looking finished. So the whole question becomes whether there is a fast, reliable check standing between the output and the customer. Where there is, automation is usually a good trade. Where there is not, you have moved risk rather than work.

Situations where it is simply the wrong tool

Advice with professional consequences

Legal, medical, tax and financial advice given to a customer. Not because the output is always wrong, but because being wrong has consequences you cannot absorb and no model carries liability. Use it to prepare questions for your accountant. Do not use it as the accountant.

Anything requiring guaranteed accuracy

Regulatory filings, safety documentation, dosage or dilution instructions, structural or electrical specifications, allergen information. If a single wrong figure causes real harm, the check has to be a qualified person, and once you have added that check the automation saved you nothing.

Moments where a customer needs to feel heard

Complaints, bereavement, cancellations, apologies, redundancies. The message is not the words, it is the fact that a person spent their attention on you. A generated apology is worse than a clumsy human one, and being caught sending one is worse still.

Work you do rarely

A task you perform twice a year does not repay the time spent building and testing an automation. Frequency is the strongest predictor of whether automating pays back, and low-frequency tasks almost never do.

Anything you cannot evaluate

This is the sharpest one. If you do not know good from bad in a domain, you cannot supervise output in it, and you will accept confident nonsense because it reads well. Contracts, technical specifications and specialist copy in an unfamiliar field all sit here. Expertise is what makes generated output safe, so the tool is least trustworthy exactly where you most want it.

Small data

Twelve rows in a spreadsheet do not need a model. Neither do four customer reviews. Automating small things adds tooling, cost and a failure mode in exchange for a saving you could have made by looking.

Where the human touch is the product

If people pay you specifically because you are personally involved, automating the contact removes what they are buying. A bespoke tailor sending generated follow-ups has misunderstood their own business.

The cost nobody counts

Every automated task creates a verification task. That work is real, it lands on somebody, and it is missing from almost every ROI calculation.

There is also a subtler cost. Checking output is more tedious than producing it, and human attention degrades quickly on repetitive checking when the output is usually fine. Three weeks of correct drafts train the reviewer to skim, and the fourth week is when something wrong goes out. If your process depends on a person catching a rare error in a stream of good material, assume they will eventually miss it and design accordingly.

The honest grey areas

Not everything is clear-cut, and pretending otherwise is its own kind of dishonesty.

• First-draft writing you will heavily edit. Usually fine, provided the substance comes from you and you check every figure.
• Summarising documents. Good, with the caveat that summaries drop the exception buried in paragraph nine, which is often the part that mattered.
• Routine customer replies. Fine for factual questions with stable answers. Not fine the moment the customer is upset.
• Translation. Strong for understanding, risky for publication without a native speaker reading it.

A test to run before automating anything

Four questions, answered honestly, in about two minutes.

• How often does this task happen? Rarely means stop here.
• What happens if the output is wrong and nobody catches it? If the answer involves harm, money you cannot lose or a customer you cannot replace, stop.
• Can someone verify it in a fraction of the time it took to do it? If not, you have not saved anything.
• Would you be comfortable if the customer knew? If the honest answer is no, that is the answer to the whole question.

Saying no to AI in the right places is what makes saying yes elsewhere credible. If you want an outside view on which parts of your own operation are worth automating and which are not, our Stamford AI consulting page explains how we approach that assessment.

Automating work they cannot check. The output is fluent and confident whether or not it is correct, so errors do not look like errors. When nobody in the business has the expertise or the time to verify what is produced, mistakes reach customers looking entirely professional, which makes them harder to spot and more damaging when found.

Use it to help you prepare, not to reply. Summarising the history or listing the points that need addressing is fine. The reply itself should be written and sent by a person, because the customer's actual complaint is usually that nobody paid attention, and a generated response confirms exactly that suspicion.

Not inherently, provided the content is accurate, genuinely useful, and reflects your real knowledge and prices. The useful test is whether you would be comfortable if a customer found out. If disclosure would embarrass you, the problem is usually the quality or honesty of the content rather than the disclosure itself.

Ask what happens if the output is wrong and nobody notices. If the answer involves physical harm, legal exposure, money you cannot afford to lose, or a customer relationship you cannot replace, keep a qualified human in the loop. Low-consequence, high-frequency, easily checked work is where automation genuinely pays.