AI Token Budgets Are Here. Usage Was the Wrong Metric.

Companies are putting employees on AI budgets because they measured adoption by how much people used AI, and that number cannot tell them which spend paid off. A cap brings the bill down. It leaves the measurement problem untouched. The better move is to measure cost per accepted outcome for each workflow, route every task to the cheapest model that can pass a check, and set budgets at the workflow level.

The evidence arrived in one week.

What happened to corporate AI budgets this week?

Apple cut the daily AI spending cap for some engineers from $300 to $200 in October, according to Bloomberg, and is considering $100 for some engineering teams. Some non-engineering groups that already sit at $100 a day could drop to $25. Staff who need more must ask senior executives for an exemption, and some teams are being pointed toward cheaper or on-device models.

Apple is not alone. The same report notes that Microsoft set division-level AI budgets in July, Tesla caps staff at $200 a week, Uber limits each AI coding tool to $1,500 a month, and Accenture caps Claude at $29 a month.

Goldman Sachs went further. In October it limited Claude Code and GitHub's coding tools mainly to engineers after a review of productivity and risk, removing access for most traders and bankers who had been using them to prototype tools. Quantitative staff known as strats were exempted.

Then there is Proton. CEO Andy Yen planned company-wide "re-interviews" asking employees to account for their AI use, and cancelled them on October 30 after 404 Media reported on the push. One employee told 404 Media that the items the re-interviews would have measured are expected to move into performance reviews instead.

These look like two opposite moves. Apple and Goldman are restricting AI use while Proton was pressing employees to use more. Both responses come from the same blind spot.

Why was AI usage the wrong thing to measure?

Usage counts activity. Tokens consumed, prompts sent and share of code "written by AI" all go up when people work harder with AI. They also go up when people loop, retry, paste whole repositories into context, or run agents that wander. The metric cannot tell these apart.

The Apple story shows the problem from the other side. Some Apple employees told Bloomberg that few people used their full $300 allowance, so the savings from the first cut may be small. A cap that most people never reach has little effect on spend. It does give leadership a number to hold, which is useful when the bill is the only data you have.

Once a usage number becomes a target, people optimize for the number. Leaderboards that rank staff by token consumption pay people to burn tokens. Re-interviews that judge people on AI adoption invite performative use. Then the bill arrives, and the response is a cap, which treats a measurement problem as a spending problem.

What does the spending data actually show?

Ramp's AI Index, built from card and bill-pay data across tens of thousands of US businesses, is one of the few direct spending datasets available. Three findings matter here.

Spend is highly concentrated. In its October 2026 update, Ramp estimates that roughly 80% of OpenAI and Anthropic spend on its platform comes from 1% of firms.

Heavy spenders are getting more for less. In the same update, token volumes rose while token spend showed virtually no growth in August or September. Top-1% firms spent $9,568 per employee per month in September, 3.1% below their July high. Ramp also reports the effective price per million tokens fell 41%, from $1.15 in March to $0.68.

The model mix is shifting down. Frontier models' share of tokens fell from 53% in early August to 45% by early September as standard and lite models picked up share.

My inference: the most sophisticated AI buyers are already doing model routing. They send routine work to cheaper tiers and save frontier models for the tasks that need them. Their bills flatten while their output grows. That is the pattern the rest of the market should copy, and per-person caps work against it.

There are caveats. Ramp's customers lean toward venture-backed and smaller firms, Ramp itself flags its top-1% estimate as volatile, and it has seen seasonal dips in AI spend around November and December before. Treat the direction as the signal and the exact numbers as noisy.

Why don't per-employee AI caps fix the problem?

A cap answers "how much may this person spend?" The question finance needs answered is "what did this spend produce?" A cap cannot answer that, for three reasons.

1. It aims at the wrong people. Concentration means a small group of heavy users drives most of the cost. Some are your highest-leverage engineers running multi-agent workflows. Others are wasting money. A flat cap hits both the same way.
2. It adds friction where value may be highest. Agentic coding and research use far more tokens than chat. Sending your best builders to an executive for an exemption slows exactly the work you adopted AI for.
3. It leaves the blind spot in place. After the cut, leadership still cannot say which $200 a day was worth it. The next budget review starts from the same guesswork.

Goldman's version is closer to a real policy, because it scopes access by role and risk instead of a dollar ceiling. It still decides who gets tools without measuring what those tools produced.

What should you measure instead of AI usage?

Measure cost per accepted outcome, workflow by workflow. An accepted outcome is a finished unit of work that passes a quality bar someone other than the producer checks.

Workflow	Accepted outcome	Cost metric
Agentic coding	Pull request merged and not reverted within 30 days	Token + tool cost per surviving merged PR
Customer support	Ticket resolved with no reopen within 7 days	Cost per clean resolution
Sales research	Account brief used in a logged meeting	Cost per used brief
Document processing	Record extracted with zero human corrections	Cost per clean record

This metric does three things usage cannot. It shows whether quality held when cost dropped. It makes workflows comparable, so you can move budget toward the ones with the best return. And it gives the heavy user a defense: if your $300 a day produces merged PRs at a lower unit cost than the team average, you have the case for your exemption already written.

How do you put outcome-based AI budgeting into practice?

You can start this in a quarter without new vendors.

1. Tag tokens to tasks. Route model calls through a gateway or proxy that records a workflow and task ID on every request. Without this, nothing below works.
2. Define the accepted outcome for your top three workflows by spend. Keep the definition boring and checkable. Reverts, reopens and corrections are good signals because systems already log them.
3. Set a default model tier and an escalation rule. Start each task on a cheaper model. Move to a frontier model when a verifier fails: tests break, a schema check fails, or a grader scores below threshold. Log every escalation.
4. Budget per workflow, not per person. Give each workflow a target cost per accepted outcome and a monthly envelope. Individuals spend inside it without asking permission.
5. Review monthly. Look for workflows where unit cost rises while acceptance holds flat. That is where to tune prompts, trim context, or change the default tier.

Who should skip this? If your monthly AI bill is small and spread across chat subscriptions, a simple per-seat plan is fine. Outcome budgeting pays off once you run agents or API workloads where one person can spend more in a day than a seat costs in a month.

Where does this approach break?

Some valuable work has no clean outcome. Exploration, learning and early prototyping often end in "we decided not to build it," which is a good result that no counter records. Give exploratory work a small, separate budget and judge it quarterly, not per task.

Attribution is also messy. A merged PR may reflect two hours of AI work and twenty minutes of a senior engineer's judgment. Cost per outcome will overstate AI's share in some workflows. That is acceptable as long as you compare a workflow against its own trend instead of claiming precise ROI.

Finally, verifiers cost money too. A grader model or a full test run adds to the cost of every task. Include it in the unit cost so routing savings are not overstated.

What happens next with enterprise AI budgets?

This is a forecast based on the current trend, so treat it as speculation. The 2027 planning cycle is likely to be the first where large companies budget AI at the workflow level rather than per seat or per person, because finance teams now have a year of bills and little proof of what they bought. Teams that arrive with cost-per-outcome data will keep their budgets. Teams that arrive with usage dashboards will get caps.

The vendors are moving too. Falling token prices and the drift toward lighter models mean the frontier labs need customers who can justify premium tiers on results. Expect more tooling for routing, attribution and evaluation built into the platforms themselves.

The takeaway

The caps this week show that companies measured AI by how much it was used and never checked what it produced. Fixing that takes one change: pick your most expensive workflow, define what a finished, accepted unit of work looks like, and calculate what one costs today. Once you have that number, budget conversations get much easier.