Is one premium model eating your AI budget? The case for routing

Most companies running AI in production made one model decision, once. Legal or security approved a premium model, and every workflow since has landed on it by default. Classification, extraction, summarization, drafting: all of it runs on the most expensive model available, because that's the approved one.

The symptom shows up on the invoice. The bill grows faster than usage, because every new workflow inherits frontier pricing whether it needs frontier quality or not. You're not paying for quality. You're paying for a procurement shortcut.

A quality bar per task, not a model per company

"Which model" is a per-task question, and most organizations answered it once for the whole company.

The fix starts with defining what good output looks like for each task. That's your quality bar. A three-way classification with an obvious rubric has a different bar than a nuanced summary going in front of a regulator. Once the bar is explicit, the routing rule is simple: use the cheapest model that clears it.

Paying frontier prices for a task a mid-tier model passes doesn't buy you anything. The output was already good enough. The extra spend bought margin you can't see and won't ever use.

How we measure it

Not benchmarks. Your work.

We take real samples from the actual workflow, redacted, and run them through each candidate model. Every model gets a pass rate against the quality bar for that task and a cost per run. Then we sort the passers by cost.

The result is a routing table, not an opinion. Task by task: which models clear the bar, what each costs per run, and which one you should be on. When the cheapest passer is a model at a tenth of the price, the data says so. When it isn't, the data says that too.

What stays premium

Some tasks only the frontier model passes. Complex reasoning over messy inputs, drafting where tone carries legal weight, anything where the mid-tier pass rate falls off a cliff.

Keep those on the premium model without guilt. That's what it's for. Routing isn't about being cheap everywhere. It's about knowing which minority of your tasks actually need the expensive model, and paying for exactly those.

The number

The savings math is one line: the cost gap between your current model and the cheapest passing model, times monthly volume, summed across tasks.

Illustrative numbers, so you can run your own. A task costs $0.08 per run on the premium model. A mid-tier model clears the bar at $0.01. At 200,000 runs a month, that one task is worth about $14,000 a month in routing savings. Across a typical production portfolio we estimate 40 to 60% of model spend is routable, with a downside of 20 to 30% if more of your tasks turn out to genuinely need frontier quality than the samples suggest.

Either way, you end the exercise knowing the number instead of suspecting it.

Find out what your routing table says

We run this as a Routing Assessment: $4,500 fixed, one week. We measure pass rates on your real redacted samples, price every passer, and hand you the routing table with the savings math attached. The fee credits toward the build if you move ahead.

Worth a conversation if your AI bill is growing faster than your usage and nobody can say exactly why.