How much are you really spending on AI per employee?

Ask a finance leader what their company spends on software per employee and you'll get a confident, defensible number. Ask them what they spend on AI per employee and the room goes quiet. The invoices are real: seat licenses for a dozen copilots, API bills that arrive with no line-item detail, departmental subscriptions expensed on personal cards. The clarity is not. As the monthly totals climb, the visibility into where the money goes shrinks in inverse proportion.

Based on what we see across mid-market deployments, most firms spend somewhere between $1,300 and $2,000 per employee per year on AI. For companies that have gone all-in, the "AI-pilled" ones stacking frontier subscriptions across every team, that figure climbs toward $7,000. These are not small numbers at scale. And yet the two questions that should be trivial to answer, what is our spend per employee and do we have rate limits on any of it, come back blank almost every time. That blankness is the debt. It is AI-driven operational debt, and it compounds quietly on the balance sheet of every process the organization runs.

The invisible leak in your workflow architecture

The largest single source of waste is not fraud or unused seats. It is routing. Most organizations run every AI task through a frontier model by default, because the frontier model is the one everyone knows and the one that was easiest to sign up for. But the majority of AI work inside a business is commodity work: classification, extraction, summarization, formatting, routing tickets, drafting boilerplate. Running commodity tasks through frontier models means overpaying by as much as 10x for output that a cheaper, faster model would have produced at parity.

This happens because there is no control plane. Without a headless control plane sitting between your workflows and the models, there is nowhere to make a routing decision. There is no place to say "this task goes to the cheap model, that task goes to the frontier model, this one gets cached." There are no rate limits, because rate limits require a chokepoint and there is no chokepoint. Every application talks directly to every provider, and the spend leaks out of a hundred small holes that no dashboard can see because no dashboard sits in the path.

Why the model is a commodity and your architecture is the asset

The uncomfortable truth for anyone anchoring their AI strategy to a single vendor is that the frontier models are now within a fraction of a percent of each other on the enterprise benchmarks that matter. The gap that existed in 2023, where one lab was clearly ahead on reasoning or coding or instruction-following, has collapsed. On the tasks a real business runs, the leading models are functionally interchangeable. The model is a commodity.

This is not a criticism of the models. It is a statement about where the value has moved. When the intelligence itself becomes a fungible, metered utility, the durable asset is no longer which model you rent. It is the architecture layer: the routing, the governance, the observability, the data controls, the fallback logic that sits between your business and whichever model happens to be cheapest and best this quarter. Companies that understand this stop shopping for models and start building the layer that lets them swap models without rewriting a thing. The model is the commodity. The architecture is the asset, and it is the only part you actually own.

"The model is a commodity. The architecture is the asset, and it is the only part you actually own."

The high cost of shadow AI

While finance debates a formal AI budget, the organization has already built one, department by department, without telling anyone. Marketing has three content tools. Sales has an AI notetaker and an enrichment copilot. Engineering pays for two coding assistants. Support runs a chatbot on a separate contract. None of it went through procurement, because each individual subscription was small enough to slip under the approval threshold and urgent enough to justify a personal card.

This is shadow AI, and its cost structure is death by a thousand seats. No single line item is large enough to trigger scrutiny, so nothing gets scrutinized, and the aggregate quietly becomes one of the larger uncontrolled expenses in the business. Worse than the money is the exposure: sensitive data flowing into tools nobody vetted, no record of what was sent where, no ability to answer a customer or a regulator who asks where their information went. The spend is only the visible symptom. The governance gap underneath it is the actual liability.

Governing tokens without killing productivity

The instinct, once leadership sees the number, is to clamp down: consolidate to one vendor, kill the shadow subscriptions, lock everything behind an approval queue. This almost always backfires, because the reason shadow AI exists is that the official channels were too slow and the tools were genuinely useful. Kill the productivity and you have traded an unmanaged asset for a managed liability.

The alternative is to govern tokens the way mature engineering organizations govern any shared resource: at the infrastructure layer, invisibly, without asking people to fill out a form every time they want to think. That means rate limiting so no single team or runaway process can blow through the budget in a weekend. It means model routing so commodity work lands on commodity models automatically. It means data residency and intelligence sovereignty so you control where information flows and which providers ever see it. And it means audit logs so every prompt, response, cost, and destination is recorded and searchable. Done right, none of this is visible to the person doing the work. The productivity stays; only the leak closes.

The Krastor Method: a framework for recovery

Recovering from AI-driven operational debt follows the same sequence every time. First, make the spend visible: instrument the actual flows so you can see, for the first time, what runs through which model at what cost. You cannot govern what you cannot measure, and most organizations have never measured. Second, insert the control plane: a single point through which the traffic routes, so that a decision can be made about each request rather than defaulting every request to the most expensive path.

Third, route by task, not by habit: commodity work to commodity models, frontier reasoning reserved for the problems that actually need it, caching wherever the same question gets asked twice. Fourth, close the governance gap: bring shadow AI into the light without killing it, apply the rate limits and audit logs, and where the data is sensitive enough to warrant it, move inference on-premise for full sovereignty. The result is not less AI. It is the same or more AI at a fraction of the cost, with a paper trail you can defend. The architecture, not the vendor, becomes the thing you own and the thing that compounds.

The Token Governance Audit

This is exactly what the Token Governance Audit is built to do. There is no upfront fee. We instrument your flows, find the leaks, and implement the routing and governance layer that closes them, and we are paid as a percentage of the savings we produce. If we don't reduce your spend, there is nothing to pay. The engagement is designed for firms doing $10M or more in revenue, where the AI spend is large enough that a percentage of the savings funds the work several times over.

The premise is simple. Your AI spend per employee is a number you should be able to state and defend, the way you can state your spend on any other category of infrastructure. Today, for most firms, it is a number nobody knows and nobody controls. The audit turns it into a governed line item, and turns the leak into the budget for fixing it. For the underlying philosophy on why the architecture layer is the asset worth owning, see the Krastor Blueprint.

"Your AI spend per employee should be a number you can state and defend. For most firms, it is a number nobody knows and nobody controls."