The AI feature that only needed a lookup table

An operations lead at a contract manufacturer asked us for an AI feature. When a new part came in, someone had to assign it a handling category, which drove how it got stored, routed, and quoted. His people did this by hand, it took judgment, and he wanted a model to learn the judgment and do it automatically. He had budget, he had buy-in, and he was ready to build.

We asked him to walk us through how his best person actually made the call. He did, and as he talked, something became clear that surprised both of us. The judgment was not judgment. It was a rule. The category was fully determined by three fields that already existed on every part: material, dimensions, and destination market. Given those three, the answer was fixed. There was no ambiguity for a model to learn, because the mapping was deterministic. The “AI feature” was a lookup table wearing a costume.

The request said AI, the work said rule

This happens more than anyone admits. Someone experiences a task as hard and repetitive, reaches for the most powerful tool in the room, and asks for AI. The word describes their frustration, not the actual shape of the problem. When you slow down and trace how the decision is really made, you sometimes find there is no learned pattern in it at all. There is a rule the expert follows without noticing it is a rule, because they run it so fast it feels like intuition. Written down, the intuition is an if-then table.

Once you see that, the AI version loses on every axis to the boring version. A lookup table is deterministic, so it is right by construction, instead of right most of the time and confidently wrong on the rest. It is instant and nearly free to run. It has no training data problem, no drift, no evaluation suite, no monitoring, no hallucinations. You can read it, audit it, and explain any decision it made in one line. A model would have taken months to build, cost real money to run, needed constant babysitting, and at best matched the accuracy the rule gets for free. The manufacturer was about to spend a fortune to make a solved problem slightly worse.

Discovery’s job includes finding the non-AI answer

There is a quiet bias in a lot of AI work: the assumption that the answer will be AI. You get hired to build a model, so you build a model. Good discovery holds the opposite posture. It treats “does this even need AI” as a live question and genuinely tries to make the problem boring. It asks whether the decision is deterministic, whether a rule or a lookup or a simple bit of logic would do the job, whether the task is hard because it needs intelligence or hard only because a human is doing it manually and would stop being hard the moment you wrote the rule down.

McKinsey’s 2025 work found that around 70% of enterprise AI use cases are adequately served by off-the-shelf tools, and the same spirit applies one level down. A large share of “AI” requests are served by something even simpler than a tool, a rule you already know, once someone bothers to check. The check is cheap. Skipping it is what costs.

A logistics coordinator at a distributor told me about the mirror version. Her team assumed a scheduling problem needed a model, and discovery found a short set of rules handled ninety percent of cases perfectly, with only the last sliver needing anything smarter. They built the rules, shipped in days, and reserved the hard, genuinely ambiguous remainder for a real model later, if it ever proved worth it. The boring answer did most of the work, and it made the case for AI honest instead of assumed.

Try to make it boring first

The habit is simple and it saves the most money. Before you scope an AI feature, try to solve the problem without AI. Trace exactly how the decision is made. Look for the rule hiding inside the judgment. If a lookup table, a few conditionals, or an existing tool does the job, that is not a lesser outcome. That is the best outcome, and you found it in an afternoon instead of a quarter.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week engagement at $2,500, one of the first things we try to do is talk you out of the build. We trace the real decision and test whether a rule, a lookup, or an off-the-shelf tool already answers it. When the honest answer is “you do not need a model for this,” that is one of the most valuable things two weeks can tell you, and it is a lot cheaper to learn before the build than after.

Not every problem that feels hard needs intelligence. Some just need someone to write the rule down. Look for the lookup table before you reach for the model.