When AI is not worth it (and what to do instead)

There is a question almost no technology supplier asks the client: “are you sure you need artificial intelligence?”. We do not ask it out of altruism — we ask because an AI project that should not exist always ends the same way: expensive, frustrating and abandoned. And whoever paid for it starts distrusting the whole technology.

This text is the opposite of a sales pitch. It lists the cases where AI is the wrong tool — and says what usually works better instead.

011. When the process already has clear rules

If a person can write down, on a sheet of paper, every condition that leads to each outcome, the process is deterministic. And a deterministic process is solved with traditional automation: a rule, a flow, an ordinary system.

Putting a language model in charge of deciding something an “if… then” would settle is trading predictability for probability. You start paying per token, depending on an external provider and living with the possibility that the answer varies — all to obtain a result that was already guaranteed.

022. When there is no reliable data

AI learns from what exists, or answers from what it is given. If the base is incomplete, out of date or full of contradictions, the model does not fix that — it reproduces the problem with a confidence the data does not have. And then the error gets worse: before it was visible in a spreadsheet, now it arrives wrapped in a well-written sentence.

In those cases, the first project is not an AI project. It is a data organization project. And it has to happen either way — with or without AI afterwards.

033. When the volume is small

Automating has a cost: development, integration, maintenance and, in AI's case, a recurring usage cost that a traditional system does not have. That cost is diluted by volume. If the task happens ten times a month and takes three minutes, the maths simply does not work — and no amount of technical elegance changes it.

• How many times a month does the task happen?
• How long does it actually take, each time?
• What would it cost to build, integrate and maintain the automation?
• What would it cost to keep it running, month after month?

If you cannot answer all four, it is not yet time to hire AI — it is time to measure.

044. When the decision demands absolute accuracy

There are decisions that admit no “almost right”: tax calculation, payroll, dosage, regulatory compliance. A language model is, by construction, probabilistic — it produces the most likely answer, not the only correct one. Where an error is unacceptable and traceability is mandatory, the answer has to come from a deterministic rule, auditable line by line.

AI can help prepare the decision. It should not make it alone when the cost of being wrong is high.

That does not mean excluding AI from those areas — it means putting it in the right place: suggesting, summarizing, reviewing, flagging. Never signing off.

055. When the problem is not one of knowledge

This is the most common case, and the most expensive. The company adopts AI to solve a symptom whose cause lies elsewhere: a badly defined process, an unsuitable tool, a decision the leadership has not taken. The technology goes in, the symptom moves, the problem remains.

Automating a broken process does not fix it. It makes it break faster, and with more confidence.

06So when is AI worth it?

When the inverse of everything above is true. When there is volume, when the data is accessible and reliable, when the task involves language, images or classification that no fixed rule can cover, when a degree of uncertainty is acceptable — and when there is human review at the point where an error would be costly.

• Reading and classifying documents that arrive in different formats.
• Service that answers the same questions all day, from controlled sources.
• Extracting information from free text — which a rule never fully covers.
• Decision support: summarize, compare, flag what deserves human attention.

07What to do before hiring anything

1. Write the business problem in one sentence, without naming any technology.
2. Measure: how often it happens, how much time it consumes, what it costs today.
3. Check whether the necessary data exists, and in what state.
4. Ask whether a fixed rule would solve it. If it would, stop here.
5. Only then evaluate AI — and require the supplier to tell you where it does NOT apply.

A supplier who never says “do not use AI here” is not helping you decide. They are selling to you.