When AI is not worth implementing: six signs to stop

Most writing about AI in business explains why you should start. This explains when you should not. Not because AI does not work, but because even working technology does not pay off when applied in the wrong place. The six signs below are the ones I see most often. Any one of them is enough to stop and think again.

1. The process happens less than once a week

Automation pays off through repetition. If a process happens a few times a year, a person will do it faster than you can explain to a system what to do. Count the hours per month first. If the number is small, stop here.

2. Nobody can describe how the process should run

If three people do the process three different ways and all three are “correct”, the system has nothing to automate. It will automate one of the three versions, and the other two people will work around it. Agree how the process should run first, write it down, and only then talk about automation. That agreement alone often saves more than any system.

3. The data lives on paper, in photos, or in people's heads

An AI system needs data it can read. If invoices arrive as phone photos and orders are written in a notebook, the first step is not AI but collecting the data in one place. It is dull, necessary work, and it has to happen before any pilot.

4. The cost of an error exceeds the time saved

Every system makes mistakes. The question is what one mistake costs. If a system sorts emails and gets one wrong, someone will notice and fix it. If a system sends priced quotes to customers on its own and gets one wrong, a single error can exceed a year's savings. In such cases the answer is not to give up on automation but to keep a person approving every decision. If that is not possible, do not do it.

5. A simple rule would do

A good share of “AI projects” are rules that fit in one sentence: if the amount is above X, send to the manager; if the email contains the word “invoice”, move it to a folder. Any automation tool implements such rules without AI, faster, cheaper, and without errors. AI is needed when the rule cannot be written down because the decision depends on context.

6. Nobody can maintain the system

A system nobody maintains stops working within months: an email format changes, a program updates, a vendor changes its terms. If there is no one in the company who can give a few hours a month and understands what the system does, it becomes one more thing that “used to work”. Handover and training are not an add-on to the project but the condition for it making sense.

What to do instead?

• Count the hours and the cost of an error first. It takes one afternoon and answers half the questions.
• Agree on a written description of one process. It is often the most useful result of the whole project.
• Try a simple rule without AI. If it handles 80 percent of cases, a person can handle the remaining 20.
• If it still looks like AI is needed, test on real data before building a system. A pilot is the cheapest way to learn the truth.

That is what the assessment stage is for: answering whether it is worth doing before spending money on how.