Where AI doesn't belong in your operation

Every operation we walk into has been told it needs AI. Most of the time, the most useful thing we can say is where it doesn't.

That's not skepticism. Models are genuinely good at some jobs — reading documents no two of which look alike, drafting replies, sorting a pile of free text. But a model is the most expensive, least predictable tool in the box, and most of the work in a business doesn't need what only a model can do.

So we route every job before we build anything. Four answers cover almost everything.

Rule: if it can be written down, write it down

If the person who does the job can explain the decision in two sentences, it's a rule. Approve a reorder if it matches contract pricing. Set the deadline from the date the regulation names. Warn when a window's shade is below the legal limit.

Rules written as code are exact, instant, free to run and never make things up. A model asked to do the same job will get it right most of the time — which, for a rule, means it's wrong.

Model: messy input, at volume

Models earn their place where the input won't fit a form: scanned invoices in a hundred layouts, customer emails, photos, notes typed in a hurry. They read, extract, sort and draft far better than any template.

Two conditions make it work. The volume has to be real — automation pays back on repetition. And there has to be a way to check the output: evaluations before launch, sampling after, and a person who signs off where it matters.

Software: when the work needs a place to live

A surprising amount of "we need AI" turns out to be "we need a system of record." The quote logic that lives in someone's head. The customer that exists in four tools with four spellings. The spreadsheet that secretly runs the business.

A model reasoning over records that disagree gives confident wrong answers. Settle the source of truth first — then decide what, if anything, a model should do with it.

Person: judgement, relationships and costly calls

Some decisions should stay human: the warranty claim that could go either way, the conversation with an unhappy customer, anything where a wrong answer is expensive and hard to notice. The right move is to do the prep work for that person — gather the history, draft the options — and leave the decision with them.

And some work happens so rarely that the best automation is a clear checklist.

The field guide

The calls we make most often, in one place:

None of these is a law. They're where we start, and the map of your operation decides. But if a proposal puts a model on a job a rule could do, it's worth asking why.

The whole field guide→