Classic AI Is Making a Comeback
I have a colleague who has been doing machine learning for several years. Random forests, XGBoost, logistic regression, the kind of model that doesn’t make conversation, doesn’t “hallucinate,” answers in a few milliseconds, and can be audited with a confusion matrix.
Ever since we got in the habit of pointing an LLM at nearly everything, he’s been repeating the same line in meetings: before you send a request to a large model, ask whether a classifier would do the job for a fraction of the price. For two years, we nodded politely and kept going the way we were.
Right, sure, your colleague who makes a living off classic ML is obviously going to lobby for more classic ML. That’s just business for him.
Fair. Except he was already pushing this before it was profitable for him to say so, back when nobody wanted to hear about anything but the latest model release. What changed isn’t his pitch. It’s that people are listening now.
That’s a signal. And I think it’s going to repeat itself across a lot of organizations before the year is out.
The pendulum is swinging back
The last two years have been a rush toward a single tool. An LLM to classify emails, an LLM to route tickets, an LLM to extract a date from an invoice. It worked, in the sense that it produced an answer. Nobody worried too much about price, latency, or predictability, because the pilot cost next to nothing and the new toy was fascinating.
You see it in the bill before you see it anywhere else. I’ve already written about what a token actually costs: it’s never the sticker price that sinks an organization, it’s the volume, and the volume only goes up. Once the “AI” line on the monthly bill gets big enough that someone actually looks at it, the question changes. It’s no longer “which model should we use,” it’s “what actually needs a model at all.”
That’s exactly where classic ML resurfaces, as an alternative. A large language model is built to produce text and reason over ambiguity. Asking it to sort six categories of email is heavy artillery aimed at a fly, and heavy artillery gets billed like heavy artillery.
What I see changing isn’t organizations abandoning LLMs. It’s that they’re starting to stop treating them as a universal hammer. The pendulum isn’t swinging against generative AI. It’s swinging toward a more honest division of labor between two tool families that each have their own job.
This isn’t a step backward. It’s a division of labor.
Classic ML never stopped working. It just stopped being an interesting thing to talk about in 2023. A model that flags fraud, scores a lead, or routes a support ticket doesn’t make a good demo in front of a board. A conversational agent answering in flawless prose does.
But a board doesn’t pay the API bill at the end of the month. Someone else does, and that person starts asking unglamorous questions: why does this six-category classifier need a model billed at $25 per million output tokens? Why can’t anyone explain, in verifiable terms, why this file got categorized the way it did?
Classic ML answers both questions in the same breath. It’s predictable, it’s fast, it’s cheap, and it’s measurable. On a closed classification task, an LLM is competing against a tool that has done the same job for fifteen years, at a fraction of the cost, with an identical answer for the same input every time. That’s not a fight it wins often.
What’s coming back isn’t nostalgia for 2015-era data scientists. It’s the belated recognition that a good chunk of what’s been routed through a large model for the past two years never needed one.
What this actually changes
Three effects, I think, will show up within a year.
Hiring rebalances. Organizations that put all their 2024-2025 energy into LLM-facing roles (prompt engineering, agent orchestration) are going to rediscover they need people who know how to train a supervised model, label a dataset, and read an ROC curve. These aren’t exotic skills. They’re skills that got sidelined for two years because they seemed less glamorous.
Architecture becomes the real subject, not the model. The question that matters is no longer “which LLM should we use,” it’s “what actually needs an LLM in the first place.” That’s close to the question I ask in single agent vs. multi-agent: the right answer depends on the problem, not on this month’s leaderboard.
Model vendors will push the other way. They sell tokens, not classifiers. Expect the marketing to keep insisting on “one model for everything,” because that’s their revenue line. The friction between that message and your own interest is going to get more visible, not less.
I don’t think LLMs are losing ground. They’ll keep doing what they do better than anything else: reasoning over ambiguity, producing text, holding a conversation. What’s changing is that they’re going to stop also doing the work classic ML has always done better.
If your organization has someone who’s been doing classic ML since before ChatGPT existed, this might be a good time to call them. Have a good week!