Tier Your Models Like You Tier Your Customers

You don't put every customer on the same service tier. You don't put every model on the same tier either, and treating them the same is how AI budgets explode.

The default setup is one model for everything. The team picks the smartest model they can afford — or the one that demos best — and routes every request to it. The trivial request, the important request, the request that needs deep reasoning, the request that needs a quick answer. All to the same model. All at the same price.

That's like flying every customer first class because first class exists. It works. It's also wildly expensive, and it's the single easiest AI cost to cut without touching quality.

The tiering frame

Every AI request has two properties you should know before you route it: how hard it is, and how much it matters.

How hard it is determines which model can handle it. A lot of AI work is easy: classification, extraction, summarization, form-filling, routine drafting. The frontier model handles it well, but so does a small, fast, cheap model. The hard work, the complex reasoning, multi-step analysis, and subtle judgment, needs the frontier model.

How much it matters determines what you're willing to spend on it. A wrong answer on an internal draft costs an hour. A wrong answer on a customer quote costs an account. A wrong answer on a compliance decision costs a regulator. The cost of error is the number that sets your tolerance.

Put the two together and you get a routing matrix. Easy and low-stakes: cheapest model, no review. Easy and high-stakes: cheap model, human review. Hard and low-stakes: mid-tier model, spot checks. Hard and high-stakes: frontier model, human review, full escalation path.

The model you use is a business decision about cost and risk, not a technical decision about capability. The cheapest model that's good enough, reviewed at the level the stakes demand, is the right model.

Why this feels wrong at first

The resistance to tiering is usually the same argument: the frontier model is better, so why would I ever use a worse one?

Because "better" isn't the goal. The goal is good enough at the right price. The frontier model is better in the same way first class is better — it's nicer, and it's not what most trips need. A classification task that a small model gets right 99.2% of the time and a frontier model gets right 99.4% of the time is not a task that needs the frontier model. That 0.2% is expensive, and you're paying for it on every single request.

The other resistance is engineering comfort. One model is simple — one integration, one vendor, one prompt convention, one bill. Tiering means multiple integrations, routing logic, and evaluation per tier. That's real work. But the routing layer is the same shape as the routing you already do for traffic, for queues, for service levels. And the savings pay for the complexity many times over.

How to build the matrix

You don't need a platform to start. You need a table and a couple of weeks of logs.

Step one: list your AI tasks. Every place your system calls a model, write down the task. Step two: for each task, estimate the cost of a wrong answer. High, medium, low. Step three: run the task on the cheap model for a week, and have the humans who review the outputs rate them. How often does the cheap model's output need fixing?

The results are usually predictable. The easy tasks — extraction, formatting, routing, routine classification — pass on the cheap model with a fix rate so low nobody noticed the difference. The hard tasks fail. You now have your matrix, built on evidence instead of opinion.

Then set the routing rules: this task always goes to the small model, this one to the mid, this one to the frontier, and here's the review level for each. Start with the 80/20 — the top few high-volume, low-stakes tasks routed to the cheap model will capture most of the savings.

The tiering that compounds

Once the matrix exists, it gets sharper over time instead of stale. The cheap models improve every few months, which means tasks migrate down a tier. The evaluation layer you built to build the matrix is the same evaluation layer that catches a cheap model's regression when it happens. The routing rules become a living document instead of a decision you made once and forgot.

And the savings compound in a second way. The money you're not spending on frontier-model requests is the budget for the requests that actually need it — and for the human review layer that makes the high-stakes outputs safe. Tiering doesn't just cut cost. It redirects cost to where it matters.

The question that opens the conversation

Ask your team one question: what percentage of our model calls actually need the model we're using?

The honest answer is usually under thirty percent. The frontier model is doing classification, extraction, and drafting that a model a tenth of the price handles fine. Nobody set that up deliberately. It's just the default, and defaults are expensive.

Tier your models the way you tier your customers. Cheap for routine, premium for what matters, and a routing rule that knows which is which. The quality doesn't drop — the budget does, and the money moves to where the stakes are. That's not optimization. That's just not paying first-class fares for coach trips.

Think this argument fits your event? Tell me about the room — the calendar is selective.