Not Everything Needs AI

We've started reaching for the model the way we reach for a light switch — by reflex, before the thought has finished forming. A question lands; a chat window opens. Most days that reflex serves us well: the model is genuinely astonishing, and I build with it every day. But somewhere in the last two years “just use AI” became the default answer to almost any question about processing information — and a default is a decision you've stopped making.

A lot of what we now route through a model is work a much smaller, dumber tool does better — faster, cheaper, more reliably. Not because the model is bad. Because it's the wrong size for the job.

Here's the thing I'd rather you feel than take on faith. A support ticket comes in. Route it.

You just watched a language model reason its way to an answer a three-line list reached by matching one word. Both were right. One was instant, free, and explainable to an auditor; the other cost a little money, a little time, a little electricity — and you'd have to trust it. For this task the rule wins on every axis that matters. And most of the information-processing we hand to AI looks far more like this ticket than like writing a novel.

The simplest tool that works

This isn't an anti-AI position; it's an old engineering principle. Google's own playbook for machine learning opens with Rule #1: don't be afraid to launch without it — “a heuristic will get you 50% of the way.” Gartner publishes a note titled, plainly, When Not to Use Generative AI. In 2025, researchers put a plain regular expression up against an LLM on a structured extraction task across nearly eight thousand medical reports. Same accuracy — and the regex was about 28,000 times faster.

For structured, repeatable, well-defined work — classifying, extracting, validating, looking things up, scoring against fixed criteria — explicit rules aren't just cheaper. They're often the better tool: deterministic (same input, same output, every time), auditable (you can read exactly why), and incapable of making something up.

What a rubric actually is

A rubric is just an explicit, written-down set of criteria. Instead of asking a model “is this a good lead?”, you write down the three things that make a lead good, and you check them. Instead of asking a model to grade an answer, you state what a good answer contains. The discipline of writing the rubric is most of the value: it forces you to say what you actually mean by “good” — the part we usually skip and quietly outsource to a model that then invents its own definition.

And here's the counterintuitive part. As models have grown bigger and more eager to please, they've become less reliable on the easy stuff — likelier to hand you a confident, plausible, wrong answer than to admit they don't know. The simple tasks are exactly where a confident wrong answer slips by unnoticed. A rubric can't be confidently wrong. It can only be wrong in a way you can see — and fix.

None of this means rules beat models. Hand a regex a furious, sarcastic, misspelled complaint and it falls apart — that's the model's home turf, and it's genuinely magic there. The answer was never rules-versus-AI. It's a model for the fuzzy, open-ended, linguistic part, a rule for the structured part underneath, and the judgment to tell which is which.

The energy footnote, honestly

Here I have to be careful, because this is where the conversation usually derails. You've seen the headlines — every prompt drinks a bottle of water, AI is boiling the oceans. Most of those numbers are inflated. A short query is around a quarter of a watt-hour; Google measured one of its prompts at roughly the energy of running a microwave for one second. The viral “ChatGPT uses ten times a web search” line traces to a 2023 estimate built on hardware we've since retired. So no — skipping one AI call to route one ticket does not save the planet, and anyone who tells you it does is selling something.

The honest version sits at a different altitude. One needless query is nothing. A whole industry making that query its reflex for every trivial task, billions of times a day, is not nothing. Data centres were about 1.5% of the world's electricity in 2024; the IEA expects that to roughly double by 2030, and it climbed 17% in a single year — with AI the leading driver. And there's a trap underneath: each time the model gets cheaper and more efficient, we use it for more, so total demand keeps rising. You don't fix that one microwave-second at a time. But “default to the biggest tool for the smallest job” is a bad habit to scale to billions.

Reaching for the right tool

The skill here was never using AI — everyone can do that now; it's a light switch. The skill is the half-second of judgment before you flip it.

• Start with the problem, not the tool. Say what actually needs to happen before you decide what to use. “Default to AI” skips this step, which is precisely why it produces overkill.
• Ask whether the input is structured or fuzzy. If it's predictable and the rule for handling it is writable — write the rule. Save the model for the genuinely ambiguous, linguistic work it's uniquely good at.
• Write the rubric anyway. Even when you do reach for a model, naming the criteria first makes it more useful — and sometimes reveals you didn't need it at all.
• Notice the reflex. Next time your hand moves to open a chat window, pause one beat and ask whether a formula, a filter, or three lines of rules would do it. Often, they would.

We spent two years learning to use AI for everything. The next, more valuable skill is the opposite one: knowing what doesn't need it. That's not a smaller skill — it's the same muscle as knowing when to reach for AI at all, and the same one behind taste: judgment about what a situation actually calls for. The model will always say yes. Knowing when the answer is “we don't need you for this” is the part that's still yours.

Trying to work out where AI actually belongs in a process — and where it's quietly adding cost and risk? That's exactly the kind of thing worth talking through.