Use the Stepladder When You Can
September 15th, 2026 | by Lane Nelson
An interlude – the system of rules
You probably need a rule, not a model.
Someone at Spiese Fluid Power had a good idea in a meeting. "We should use AI to catch orders that are about to blow through a customer's credit limit." Heads nodded. It sounds like exactly the kind of thing AI is for.
It is not an AI problem. The credit limit is a number. The order total is a number. If the second exceeds the first, route it to a manager. That is an IF statement, and your ERP has been able to do it since the Clinton administration. Spending a probabilistic language model on it would be like hiring a Cirque du Soleil acrobat to change a lightbulb – impressive, expensive, and not obviously better than the stepladder you already own.
This happens constantly. And most of the time, the thing being called an "AI opportunity" is something far less glamorous in disguise: a backlog of integrations and business rules nobody ever got around to building. Name it for what it is, and a good part of your AI roadmap turns back into the automation project it always was – cheaper, faster, and due years ago.
The box nobody counts
In the launch post we drew two machines. Your ERP is a system of record – deterministic, arithmetic, auditable. A language model is a system of inference – probabilistic, linguistic, fast. The whole argument was about the handoff between them.
There is a third box, and it sits quietly in front of both: the system of rules. It is where anything belongs whose input is already structured and whose logic is stable. It is also the box that gets skipped in almost every AI conversation, because it is the least exciting thing in the room.
A rule is an integration, a workflow trigger, a lookup, a report. It is not glamorous. Nobody gives a keynote about a well-placed IF statement. But when a task fits the box, a rule beats a model on every axis that matters: it is cheaper, it is faster, it is fully auditable, it does not need a GPU or a monthly API bill, and – the one that should end most arguments – it does not hallucinate. When a rule is wrong, it is wrong the same way every time, in a spot you can find and fix once.
A two-question test
You can sort most "AI projects" in about ten seconds. Ask two questions:
Is the input structured? Fields and codes, not free text, images, or a human being's phrasing.
Is the logic stable? The same conditions produce the same action, every time, and you could write the action down.
If both answers are yes, you do not have an AI project. You have a rule you have been meaning to write. A few that people routinely mislabel:
"Use AI to read our recurring EDI orders." The customer sends the same layout every week. That is an integration, not a document model.
"Use AI to apply the right contract discount." The customer's tier is a field; the discount is a table. That is pricing logic.
"Use AI to route warranty claims to the right queue." If the queue is a function of product line, that is a lookup.
"Use AI to flag POs with no receipt after ten days." That is a report with a WHERE clause.
Inference earns its keep on the other side of those questions – when the input is genuinely messy (the PDF that arrives on the customer's letterhead, the handwritten note in the margin) or the judgment genuinely varies case by case. That is real work for a model. The credit-limit check is not.
Why the third box gets skipped
If rules are so obviously better when they fit, why does everyone reach past them? Three reasons, and only one of them is good.
Rules are boring. No one demos a workflow trigger at a conference. "We deployed AI" makes the board deck; "we finally wrote the business rule" does not, even when the second sentence describes the better decision.
Rules don't carry an invoice. A model you subscribe to has a vendor behind it who would like you to subscribe. A rule you could build yourself has no such champion. Follow the enthusiasm and it tends to point at the thing someone is selling.
And the honest one: sometimes the rule genuinely wasn't built. A lot of what looks like an "AI opportunity" is really a backlog of integrations and workflows nobody got to – years of small automations that should exist and don't. That is a real problem. But it is technical debt wearing a costume, and the fix is to write the rule, not to wrap the gap in a language model and pay by the token to paper over it.
That last case is the expensive one. Reach for inference where a rule would do, and you take a deterministic, auditable, effectively-free process and replace it with a probabilistic, opaque, metered one – while adding a failure mode it never had, the quiet plausible wrong answer. That is not modernization. It is a downgrade with a bigger bill.
The honest sequence
The system of rules where you can. The system of inference where you must. The system of record for everything that has to be true.
Run the two-question test before the next AI project leaves the meeting. Most of the time the structured, stable ones are not waiting on a model at all – they are waiting on an afternoon of integration work someone keeps deferring because it isn't exciting. Do that work. It is cheaper than anything with "AI" in the invoice, and it frees the inference budget for the work that genuinely needs a machine that can read, weigh, and guess.
Reach for the stepladder first. Use the model on the problems that actually need one.
Spiese Fluid Power is fictional – a composite built from published industry benchmarks and anonymized operating ratios. Its full operating profile, with sources, is published separately.