The main case
On November 3 Menlo Ventures published its annual report on AI in the enterprise.
Deployment spend reached 177 billion dollars in September against 33 billion a year earlier (Source: menlovc.com).
That is 5.4 times more in twelve months and the pace is accelerating. In 2025 growth was 3.9 times.
The money is not spread evenly either, since Anthropic, OpenAI and Google take 122.7 billion, 69% of the total when a year ago it was 50%.
Anthropic goes from 40% to 55% of enterprise API spend while Google falls from 21% to 9%.
On OpenRouter agents already generate 80% of the tokens, about 130.9 trillion per week (Source: menlovc.com).
Consumption stopped being measured by people and is measured by work executed.
If the cost goes per task, the number of employees stops predicting the invoice.
55% of that spend is new budget, not money that came out of software or payroll.
That is, AI does not replace a budget line, it only adds one.
Spend also concentrates where the result can be checked.
Programming and chatbots add up to 76% of an applications market of 77.3 billion (Source: menlovc.com).
It is no accident: those are the places where an agent can act and where someone can verify whether it got it right.
For whoever approves support budgets, the change is direct, since a variable invoice gets signed for the first time and the metric stops being how many people there are.
The other side
The same week, Deloitte published its annual survey and set the counterpoint.
- 81% of leaders say AI improved decision speed
- 25% is truly transforming the business
(Source: deloitte.com).
The headline of its second finding is the phrase of the month.
"The bill is visible, the unit cost is not."
You approve by model capacity and you review by the feeling of speed, almost never by what a resolved task costs.
Menlo closes with a figure that works as a mirror, pay attention. Every dollar invested in training models generated 1.70 dollars of revenue from deployment (Source: menlovc.com).
The industry already knows how to measure its aggregate return, but from all this long talk I give you my conclusion is that what it has not solved is the cost of the concrete task, the one approved on Monday and reviewed on Friday, for example.
My own reading
The problem is not the invoice, it is the unit of measurement.
You buy capacity and you review feeling, and between those two things there is no number that says how much it cost to resolve the customer's case.
An agent that fails and retries costs twice and does not warn you, it happens a lot, nobody measures that rate, the budget goes into work that never got delivered.
Deloitte's 25% is what has been transformed with all that invoice and the rest is purchased haste, they used to call it FOMO, right?
What to look at?
- Cost per task, not per seat. If the provider bills by tokens, ask for the spend of one real case from start to finish. A three-week pilot gives you that number and a demo does not.
- Who pays for the retries. An agent that fails and tries again costs twice and does not warn you. If nobody measures that rate, the budget goes into work that never got delivered.
- Whether the 25% transforms or only accelerates. Deloitte's 81% is speed. Before renewing, compare the cost per unit delivered with last year's. If latency drops and unit cost rises, you have bought haste.
How I would test it
I have this half built in BENCHMARKS-LLM, a Streamlit app that reads my reports and runs batteries against local GGUF models.
I measure speed in tokens per second, latency and a 0 to 5 rubric scored by an optional cloud judge. What I do not have is the cost column, since they are local LLMs, but without that column I repeat the report's mistake.
I would add three fields to the report writer, input token price, output token price and number of retries, and I would divide the run's spend by the answers the judge approved.
Originally published at https://sammideblas.com/notas/the-ai-bill-is-visible-and-the-cost-per-task-is-not
Si has leído hasta aquí... reacciona y comparte -> "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?
Top comments (4)
The retry point is the one I'd underline twice. The failure mode we kept hitting wasn't a loud crash — it was an agent that quietly re-ran a tool three times, "succeeded," and billed us for all four attempts while shipping nothing. Per-seat accounting never sees that; per-task only sees it if you carry a correlation/run ID through every model and tool call and sum against it, not against the invoice line.
What finally gave us honest numbers was treating a task like a trace: one ID from the user request down to the last tool result, with token cost tagged at each hop. Then the "cost of a resolved case" and the "cost of work that never delivered" fall out as two separate totals — and the second one was consistently bigger than anyone guessed.
Curious how you'd attribute cost when one task fans out into sub-agents — do you roll child spend up to the parent task, or keep a separate line for the orchestration overhead?
The Deloitte finding hits the exact friction of moving from seat licenses to consumption billing. When 55 percent of inference spend is net-new budget rather than reallocated software spend, finance treats the invoice as an open-ended R&D line item. The audit crunch arrives at annual renewal when procurement tries to map aggregate model spend back to specific workflow outcomes. Without run-level cost attribution to separate successful completions from retries, teams default to dividing the total bill across headcount, which recreates the seat-based accounting they thought they left behind.
Cost per task is the right unit once agents generate 80 percent of the tokens, since headcount stops predicting the invoice. The point that 55 percent of the spend is new budget deserves more attention than it gets. Did the Menlo report offer any way to measure cost per task, or is that still something each team has to build from its own logs?
An invoice cannot show the cost per task. In an agent loop the re-read input can be most of the total, so a per-token rate alone hides where the money goes.