The first two weeks of any job with us are spent working out what you actually need. If that ends in do not build this, it has done its job. I say so on the first call and people usually assume it is a line. Here is what it looks like in practice.
The process has never been written down
Someone in operations has been making a judgement call for years and they are good at it. The rule has never been written down, partly because it has exceptions and partly because writing it down means somebody has to approve it.
You cannot automate a rule that does not exist yet. Defining it is a management task rather than an engineering one, and once it exists you may well find you do not need a model at all.
The volume is too low to matter
A task that happens forty times a month and takes six minutes is buying you back four hours. Four hours is real. It is not worth a programme, a set of connections into your other systems, and a piece of software that somebody has to look after for the next five years.
Automate it with a form and a rule if it irritates people. Spend the AI budget on the thing that happens four thousand times.
The data exists and nobody can release it
This one is painful because the use case is often excellent. The records are there, the value is obvious, and two departments cannot agree who owns them. The disagreement predates your project and will comfortably outlive it.
That is an argument about who owns what, wearing an AI costume. Settle it in that order and the technical work becomes straightforward later.
A simple set of rules would do it
Plenty of decisions follow from the facts every time, with no judgement involved. If you can write it out as a simple list of if-this-then-that rules, write the list. Faster, cheaper, you can show anyone exactly why it decided what it decided, and it will not surprise you in eighteen months.
I have replaced a proposed model with about two hundred lines of ordinary code more than once. Nobody has ever complained afterwards.
The organisation is not ready to change how it works
The system works. Nobody uses it. This is the most expensive failure on the list precisely because everything technically succeeded.
If the people whose daily work changes have not been asked before the build, the adoption problem already exists, and no amount of accuracy will fix it.
None of this is an argument against AI. We build a great deal of it. It is an argument for finding out which of these you are looking at before the budget is committed, and that takes about two weeks.
Working through this on a live programme?
A 45-minute call with the engineer who would run the work. We will tell you whether AI is the answer, including when it is not.