The unit economics problem in enterprise AI

Ask an enterprise what it spends on AI and you’ll get a number. Ask what a resolved support ticket costs with AI in the loop, or a processed document, or a completed agent task, and the room goes quiet. That silence is the unit economics problem. The bill is visible. The cost is not. And in my work leading enterprise AI delivery, that gap is where more programs die than anywhere in the model layer.

A bill you can’t decompose is a bill you can’t forecast, and a bill you can’t forecast is a budget you can’t defend. When the finance conversation finally happens, and it always happens, engineering shows up with usage graphs and finance shows up with an invoice, and neither can say what the organization actually bought. CFOs respond to unexplainable spend the way CFOs always have: they freeze it. I’ve watched programs that engineering considered successful get paused not because they failed, but because nobody could say what success cost per unit.

Agents broke the chat-era cost model

The first generation of enterprise AI budgeting was built on chat assumptions. A user asks a question, the model answers, the interaction costs a few thousand tokens, and usage scales roughly with headcount. You could budget it like seat licenses and be approximately right.

Agentic workloads do not behave like that. An agent working a task plans, calls tools, reads the results, recovers from failures, and iterates. Every one of those turns carries context forward. A task a chat interaction would have handled in one exchange can consume an order of magnitude more tokens as an agent workflow, and the consumption scales with task complexity, not with the number of humans logged in. Budget models built on per-seat intuitions get run over by per-task reality.

This is also why total bills climb even as per-token prices fall. Prices per token have dropped steadily, and every price drop makes new workloads economical, which means more workloads run, with more autonomy, over more context. Cheaper units, more units, bigger bill. Anyone who lived through cloud adoption has seen this movie. Falling instance prices never made the AWS bill go down.

The split market

The result is a market that has split in a revealing way. Some organizations are cutting AI budgets because spend feels out of control and nobody can prove value per dollar. Others are planning major growth in AI investment on the conviction that the returns are there. Industry surveys keep finding both camps at once, sometimes inside the same company.

What strikes me is that both camps usually lack the same thing: instrumentation. The cutters can’t show that spend is wasted, and the growers can’t show that it isn’t. They are making opposite decisions from the same missing data. The disagreement isn’t really about AI. It’s about the absence of unit economics, resolved by temperament instead of measurement.

What FinOps already figured out

We have run this experiment before. Early cloud adoption produced the same shape of crisis: bills that grew faster than understanding, engineering teams who provisioned freely, finance teams staring at an invoice with no allocation behind it. The response, over about a decade, was FinOps. Cloud cost went from bill-watching to an operating discipline with named practices: allocation and tagging so every dollar maps to a team and a workload, rightsizing so capacity matches need, forecasting tied to business drivers, chargeback and showback so consumers of a resource feel its cost.

None of that was a tool. Tools existed, and mattered, but the change was organizational. Cloud spend became something teams operated, with owners, reviews, and feedback loops, instead of something they discovered at month end.

AI spend is at the bill-watching stage, and it needs the same maturation, faster, because agentic workloads scale spend faster than humans provisioning VMs ever did.

What a mature practice looks like

The practices translate more directly than people expect. A mature AI cost discipline, in my experience, has a few load-bearing pieces.

Per-workflow attribution. Every model call carries metadata tying it to a workflow and a unit of work, so you can state cost per resolved ticket or per processed document, not just cost per month. This has to be wired in from the first deployment. Retrofitting attribution onto an untagged history is miserable, and mostly doesn’t happen.

Model right-sizing. Not every step of a workflow needs the frontier model. Classification, extraction, and routing steps often run fine on smaller, cheaper models, with the expensive model reserved for the reasoning that actually needs it. Right-sizing is the AI equivalent of not running every batch job on the largest instance type, and the savings compound because agentic workflows repeat.

Caching and routing policy. Which prompts are cached, which requests route to which tier, and who is allowed to change those defaults should be decisions made once, as policy, not per-developer improvisation. The gap between a tuned and an untuned configuration on the same workload is large enough to decide whether the unit economics work at all.

Pre-production cost review. New agentic workloads get a cost model before they ship, the same way they get a security review. Expected tokens per task, expected task volume, worst-case loop behavior, and the resulting monthly range. The point is not precision. The point is that nobody is surprised, and that a workload whose economics don’t close gets redesigned before production, not after the invoice.

Tools instrument. Disciplines decide.

There is a strong temptation to treat all of this as a dashboard problem, buy an AI cost observability product, and declare it handled. The dashboard is necessary and insufficient. A dashboard nobody is obligated to act on is a screensaver.

Cost management is an operating practice problem. Somebody owns cost per unit of work for each workflow. Reviews happen on a cadence. Forecasts are compared against actuals and the misses are investigated. Chargeback makes consuming teams feel the spend they generate. Tools instrument all of this. They don’t decide any of it. The organizations that internalize that distinction will be the ones whose AI budgets survive contact with their CFOs, and the ones that get to keep building.

Related: The business case for AI has two numbers. Most people only produce one.