When AI Is the Wrong Answer

Saying no to AI work

We build AI systems and we turn down AI projects regularly. Not out of caution — because a cheaper, more reliable answer exists and the client will be better served by it.

The six situations below cover almost every AI project we have declined.

1. The answer must be exact, every time

Tax calculations, payroll, invoice totals, regulatory determinations, medication dosages. Anywhere a wrong answer is unacceptable rather than inconvenient.

These need deterministic code with tests. A model that is right 99.5% of the time sounds excellent until you multiply by ten thousand transactions and find fifty errors in a ledger.

2. Ordinary code already does it

We are sometimes asked to replace working rule-based logic with AI because AI is the current word. If the existing rules produce the right answer, replacing them trades reliability for novelty.

The right question is not “could AI do this?” but “is anything currently going wrong?” If nothing is, the project is a rewrite with extra failure modes.

The exception is a rule set that has grown to hundreds of conditions and become unmaintainable. That is a genuine case for a model — but it is a maintainability argument, not an accuracy one.

3. There is no way to check the output

If nobody can tell whether an answer is right, you cannot evaluate the system, you cannot detect regressions and you cannot improve it. You will have a black box that everybody stops trusting after the first visible mistake.

Before any AI project we ask how we will know whether it is working. If there is no answer, that is the project — establish a ground truth first.

4. The data does not exist

“Predict which customers will churn” requires historical churn data with the signals that preceded it. If nobody recorded why customers left, no model can learn it.

• Prediction needs history, labelled with outcomes
• Personalisation needs behaviour recorded over time
• Forecasting needs several cycles of clean data
• Anomaly detection needs a definition of normal

The right first project in these cases is instrumentation. It is less exciting and it is the only thing that makes the exciting version possible in a year.

5. The volume is too low

AI projects have a fixed cost regardless of how often they run. Automating something that happens four times a week saves a couple of hours a month and costs £12,000.

Our rough threshold is around five hours a week of human time on the task. Below that, a better form, a template or a checklist usually delivers more for less.

6. The real problem is the process

A distributor asked us to build extraction for supplier documents arriving in fourteen formats. The better answer was a supplier portal with one form, which removed the extraction problem entirely for two thirds of the volume.

Before automating the reading of a badly-structured input, always ask whether the input can be structured at source. It is frequently cheaper, always more reliable, and nobody thinks of it because the manual work has been normal for years.

Frequently asked questions