We build AI systems, and a meaningful share of our scoping calls end with us recommending something simpler. That is not modesty; it is the fastest route to a working solution. Here are the cases where reaching for a model makes the problem worse.
When the rules are already known
If a competent person can write down the logic — if invoice over ₹50,000 then two approvals — implement the logic. It is deterministic, testable, instant, free to run, and explainable to an auditor. Using a model to infer rules you already possess adds cost, latency and uncertainty in exchange for nothing.
If you can write the rule down, write the rule down.
When the answer must be exactly right, every time
Payroll calculations, tax computation, ledger balancing, dosage arithmetic. These are domains where 99.5% is a failure, and where deterministic code with tests is the appropriate tool. AI belongs around such systems — extracting the inputs, explaining the output — not inside the calculation.
When the data does not exist
No amount of modelling compensates for absent or unlabelled data. If the historical record needed to learn the pattern was never captured, the honest sequence is to start capturing it, solve the problem another way meanwhile, and revisit in a year with something to train on.
When the volume does not justify it
Twelve documents a month does not warrant an extraction pipeline with review tooling and monitoring. The build cost, the maintenance and the attention exceed the manual effort many times over. Volume is what converts a task into an engineering problem, and it is worth calculating before building.
When a form would fix it
A surprising number of extraction problems exist because the data was collected as free text when it could have been collected as structured fields. If you control the input, changing the form is a permanent fix costing a day. Building AI to parse text you chose to accept unstructured is solving a self-inflicted problem.
Why we say this out loud
Recommending a smaller solution costs us revenue on that engagement and earns the next one. It also avoids the outcome that damages everyone: an AI project that consumed a year, delivered little and made the organisation sceptical of the next proposal — including the one that would have worked.
