You might not need generative AI

Every week someone comes to us with the same sentence: "we want to put generative AI into process X". And, quite often, the honest answer is: don't. Not because AI doesn't work — it's what we do for a living — but because process X, examined up close, is a clear rule waiting to be automated the classic way: cheaper, more predictable, and 100% auditable.

This article is the criteria we use internally to decide when generative AI belongs in a project — and when it would just be inflating the bill.

01The question that starts out wrong

"Where do I put generative AI?" already starts from a decision that's been made: that the answer is an LLM. It's the hammer looking for a nail. The question that changes the outcome is a different one: what is the simplest thing that solves this problem? Sometimes the answer is an LLM. Often, it's a well-written if.

02When a rule solves it better

If your process follows a clear, stable rule — "if the invoice matches the order, reconcile it; otherwise, flag it" — deterministic automation beats the LLM at everything that matters: it's cheap (no per-token cost), predictable (the same input always produces the same output), and 100% auditable (every decision can be explained by reading the code).

We practice what we preach: one of our monitoring products is entirely deterministic, with no LLM — and out of honesty we don't call it AI. A rule that gets the job done doesn't need an AI badge to work well.

03When the LLM earns its place

There is a territory where the classic rule struggles and the LLM shines: language and ambiguity. Free-form customer text, unstructured documents, context interpretation, summaries, drafts. There, writing rules for every case becomes a losing battle — and a language model does in seconds what the team would take hours to do.

But even there it comes in wearing a seatbelt: in copilot mode (the AI proposes, a person approves), with a lock on sensitive data and an auditable trail of what was done. An LLM running loose in production isn't innovation — it's a liability.

04The cost nobody puts on the slide

Generative AI has three costs the demo doesn't show. Tokens: every call costs money, and the cost grows with volume — what was pennies in the pilot becomes an invoice at rollout. Variance: the same question can produce different answers, which demands human review or automated evaluation. And governance: someone has to monitor, measure, and answer for it when the model gets it wrong.

None of this is a reason not to use an LLM — it's a reason to use it where it pays for its own cost. Don't throw an LLM at a spreadsheet problem.

05The simplest-thing test

Our internal criteria fits in three lines. Clear, stable rule? Deterministic automation — and be honest about it. Language, ambiguity, exceptions that change every day? An LLM in copilot mode, with a lock and a trail. In doubt? Start with the deterministic option, measure, and only move up to AI when the evidence calls for it — because the reverse path (shutting down an expensive LLM that never needed to exist) is far more embarrassing.

And in either case: measure, don't estimate. Token costs on one side, hours won back on the other. What isn't measured becomes opinion — and opinion doesn't sustain a budget.

06Why would a vendor say this

A vendor that sells LLMs for everything has an obvious incentive: an AI project bills more than a business rule. We prefer the long-term incentive: 98% of our clients come back, and they come back because the recommendation was honest the first time. Where AI belongs, we point it out. Where a rule solves it better, we say so — and we don't charge you an LLM for it.

If the answers point to a rule, celebrate: you just saved yourself an AI project. If they point to an LLM, great too — now it comes in for the right reason, with the right controls.