New format, starting this week: one idea every Tuesday, five minutes to read, one free tool to use on Monday.

Someone asked your team for an AI tool this week. There is a fair chance what they need is a rule.

Use the least clever tool that does the job.

Every idea on an AI wish list belongs in one of three piles.

Rules: the logic is known and stable. Matching, thresholds, reconciliations, approval routing.

Models: the pattern hides in historical data. Late payment risk, anomaly flags, cash forecasts.

Language: the input is text nobody structured. Contract clauses, supplier emails, first drafts of commentary.

Only the third pile needs generative AI. The first pile does not need AI at all.

The wording gives it away.

"Can AI flag invoices above the approval limit?" is a rule. "Can AI explain why a supplier disputes an invoice?" is language.

The clever option costs more after go-live, not before.

A rule gives the same answer every time, and an auditor can read it on one page.

When a rule is wrong, it is wrong the same way every time. That makes it easier to find and fix (deterministic).

A model needs monitoring, because the data it learnt from keeps changing.

A generative tool needs a reviewer, because the same question can get two different answers (probabilistic).

Rules are also cheaper to run. No licence per user, no fee per question, no model to revalidate when the vendor ships a new version.

None of that is a reason to avoid AI. It is a reason not to pay for it where a rule would do.

There is a second cost. Every rule dressed up as AI takes a slot on the roadmap.

The use case that really needs a model waits behind it.

Monday: put every idea on your list into one of the three piles.

Pull the list of AI requests your team has raised this year. The informal ones count too.

Tag each one: rule, model or language.

If you cannot tell, write down the decision it supports and the data it uses. The pile usually becomes obvious.

Send the rules to whoever configures your ERP or workflow tools. They are often your quickest wins, and nobody needs an AI budget to ship them.

Do not be surprised if the rules pile turns out to be the biggest.

Score the other two piles before anyone buys a licence.

Free this week: the Use Case Selection Scorecard. Rank what is left on value, feasibility and regulatory exposure, and see which idea deserves a pilot. Score your use cases

Want the longer version? Chapter 6 of my new book, Generative AI for Finance Leaders, covers when to choose deterministic, generative or agentic approaches. Out now in paperback and PDF.

Christophe

P.S. How many ideas on your list turned out to be rules? Reply with a number.