You Are Paying Frontier Prices for Commodity Work

Here is the pattern we find in almost every AI estate we audit. The business pays for a frontier model, the best reasoning engine money can buy, and then sends it everything. Meeting summaries. Subject lines. Formatting a table. Classifying a support ticket into one of four buckets.

Every request goes to the most expensive model in the building, because nobody decided otherwise.

That sentence is the whole problem. Not the model, not the vendor, not the pricing. The absence of a decision. Routing is a decision, and in most businesses nobody owns it.

How the blowout actually happens

Nobody plans to overspend on AI. The blowout is structural, and it arrives in three quiet steps.

**Defaults do the deciding.** When a developer wires up an agent, they pick the model that made the demo impressive. That default ships to production and nobody revisits it, because it works. Working and worth the price are different questions, and only one of them gets asked.

**The invoice hides the shape of the spend.** Your AI bill arrives as one line per vendor. Behind that line are tens of thousands of calls doing wildly different work: a few hundred genuinely hard reasoning tasks, and a long tail of mechanical transformations any capable small model handles identically. The invoice cannot tell you which is which, so the tail rides at frontier prices.

**Nobody is measuring cost per conversation.** Teams track total spend, which only tells you the number went up. The unit that matters is the conversation: what did this workflow cost, in tokens, tool calls and model time, and what did it produce? Without that resolution, optimisation is guesswork and budget reviews are vibes.

McKinsey's State of AI research found that while 78 percent of organisations now use AI regularly, more than 80 percent report no material impact on earnings. Unmanaged spend is a meaningful part of that gap: the costs are real and rising, while the value is untracked.

The surgeon stapling paper

A frontier model summarising a meeting transcript is a surgeon stapling paperwork. It will do the job flawlessly. You are still paying surgeon rates for stapling.

The uncomfortable truth about most enterprise AI workloads is how much of the volume is stapling. Classification, extraction, reformatting, first-draft boilerplate, routing decisions themselves. On our clients' estates, the genuinely hard calls, the ones that need frontier reasoning, are a minority of traffic. The rest runs identically on models that cost a fraction per token.

The inverse mistake also happens: a cheap model grinding on a task above its weight, failing quietly, and burning retries. Cost discipline is not about always going cheap. It is about matching the model to the work, per request.

Routing is a decision. We made it a feature.

Semantic model routing does one thing: it reads each request and sends it to the model that handles it best for the lowest cost. Simple turns go to fast, cheap models. Hard reasoning goes to the frontier. Nobody on your team thinks about it, because the routing layer owns the decision that previously nobody owned.

On the [Sentry Agent Management Platform](/agent-management-platform), routing is one of five levers we run across an estate:

1. **Semantic model routing.** Usually the single largest line-item saving.

2. **Context re-orchestration.** Most workflows stuff entire histories into every prompt. Restructuring retrieval so each call carries only what it needs cuts tokens before the model even starts.

3. **API orchestration tuning.** Tool calls that run one after another, parallelised. A workflow that took twenty seconds takes two, and latency is a cost too.

4. **Knowledge base restructuring.** Fixing layout, chunking and embeddings so retrieval answers in one hop instead of five.

5. **Conversation-level token tracking.** Every interaction recorded with its tokens, tool calls and latency, so a cost regression is visible the day it appears, not the month the invoice lands.

Run together, these routinely take 20 to 40 percent off the same workload. Not by doing less AI. By doing the same work deliberately instead of by default.

Why this needs a control plane, not a memo

You could email the team and ask everyone to pick cheaper models. It will not survive the quarter, for the same reason cost policies never survive contact with defaults: the decision has to live in infrastructure, not in intentions.

That is the argument for running every agent through one control plane. When every conversation across every vendor is measured on the same axes, the expensive patterns become visible, the routing rules apply everywhere at once, and the saving is evidenced rather than asserted. When each team's agents live in their own console, none of that is possible, and the defaults win again.

We wrote up how the whole system fits together in the [AIOS whitepaper](/aios-whitepaper), and the platform that runs it is in free beta now.

What to do this month

1. **Get cost per conversation, not cost per invoice.** If you cannot say what your top three AI workflows cost per run, start there.

2. **Find your stapling.** Sample a day of traffic and sort it: how much genuinely needed the frontier model? The honest answer is usually the business case for routing.

3. **Put the decision somewhere.** Routing policy, owned, in infrastructure. Not a guideline. A layer.

If you want the shortcut, start with our [AI Opportunity Audit](/ai-opportunity-audit). The browser step is free and takes two minutes: map your tools and teams, see where the AI spend concentrates, and get your top three opportunities. The full week ships a working fix in five days.

You are already paying for the intelligence. Stop paying frontier prices for the stapling.