> ## Content Index > Fetch the complete content index at: https://varops.com/llms.txt > Use this file to discover other available public pages before exploring further. # We budget AI tools per seat while the vendors meter them per token - URL: https://varops.com/we-budget-ai-tools-per-seat-while-the-vendors-meter-them-per-token/ - Published: 2026-08-31T12:03:53.000Z - Updated: 2026-08-31T12:03:53.000Z - Description: GitHub meters Copilot in tokens and still bills by the seat. Uber's AI budget went in four months. The unit in our planning documents stopped measuring the work. - Author: Ran Aroussi - Tags: Overhyped *Two days in June, one problem. GitHub started metering Copilot in tokens and kept billing it by the seat. Uber capped its people at $1,500 a month each after the AI budget went in a third of a year.* [***Nix***](https://varops.com/columnist/nix/) *reads both as the same failure, and it isn’t a pricing failure - the seat used to measure two things at once, access and work, and AI pulled them apart while every planning document kept treating them as one number. His Founder Mode argument: the assumptions came bundled with the tools, and nobody re-denominates them for you. There is an audit at the end. —* [*Muximus*](https://varops.com/columnist/muximus/) --- A seat is not a unit of work. It used to be. That is the whole problem, and it is sitting in three documents that have nothing to do with software procurement: the license count, the headcount plan, and the capacity forecast. Two days in June made it visible. On 1 June 2026, GitHub moved Copilot to usage-based billing. Premium request units became GitHub AI Credits, consumed against token usage - input, output and cached - at each model’s published API rate. (Subscribers on annual Pro and Pro+ plans stay on premium-request pricing until their plan expires.) GitHub’s reason, from its own [27 April announcement](https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/?ref=varops.com), was that Copilot had become “an agentic platform capable of running long, multi-step coding sessions,” and that under the old model “a quick chat question and a multi-hour autonomous coding session can cost the user the same amount.” The next day, [Bloomberg reported](https://www.bloomberg.com/news/articles/2026-06-02/uber-caps-usage-of-ai-tools-like-claude-code-to-cut-costs?ref=varops.com) that Uber had capped employee spending on agentic coding tools at $1,500 per person per month, per tool - Claude Code and Cursor among them - after blowing through its 2026 AI budget. The burn rate behind that cap came out earlier, from Uber’s own CTO, reported in April by The Information and [carried by TechCrunch](https://techcrunch.com/2026/06/02/uber-caps-employee-ai-spending-after-blowing-through-budget-in-four-months/?ref=varops.com): the full-year budget went in roughly four months. The vendor changed what it charges for. The buyer’s budget was still counting something else. ## A seat used to be two things at once Per-seat pricing worked because it measured two different things with one number, and for most of the SaaS era those two things moved together. The first is access: who is allowed to open the software. The second is work: how much gets done inside it. Business software made a named person more productive, so the count of people with access and the volume of work produced rose and fell together. One number served both, and it served both well enough that nobody had to notice it was doing two jobs. Bain & Company put the same observation in an [October 2025 analysis](https://www.bain.com/insights/per-seat-software-pricing-isnt-dead-but-new-models-are-gaining-steam/?ref=varops.com): per-seat pricing “historically worked well because most business software provided value to customers by making employees more productive,” and “it was easy to understand, forecast, and budget around.” AI separated the two. How much work gets done now depends on how much compute a person directs, not on whether they hold a license. Bain’s version of the consequence is blunt: “If a business customer needs fewer humans to operate the software, the economics of pricing based on headcount disintegrates.” The word did not change. The two things it names came apart underneath it. ## Read GitHub’s own fix as the admission GitHub’s June change is the cleanest illustration available, because GitHub changed the unit of consumption and kept the unit of account. This is the vendor describing its own price change, so the terms are authoritative and the framing is theirs. Consumption is metered in tokens now. Base plan pricing did not move: Copilot Business is still $19 per user per month, Copilot Enterprise still $39\. The included credits are denominated to match the seat - $19 of monthly AI Credits for a Business seat, $39 for an Enterprise seat. Existing Business and Enterprise customers got promotional included usage of $30 and $70 a month across June, July and August. The revealing part is the mechanism GitHub added to make it work. Credits can be pooled across an organization, which GitHub says “helps eliminate stranded capacity” - each user’s unused allocation otherwise sitting isolated in their own account. Pooling is a fix for allocating a shared, uneven resource in per-person portions. The fix exists because the allocation no longer fits the resource. Alongside pooling, GitHub added budget controls at enterprise, cost-center and user level, so admins can cap spend once the pool runs out. The meter changed and the ledger did not. Tokens are what gets consumed. Seats are still what gets counted. ## What the mismatch costs Uber is the case with a number on it. A full-year AI budget consumed in roughly four months. Then a cap of $1,500 per employee per month, per tool. That is a per-head instrument applied to a spend that scales per token, which is why it works as a cap and not as a forecast. A cap stops the overrun. It does not tell anyone what the correct number was. Worth being precise about what is Uber’s and what is mine. Uber confirmed the cap to Bloomberg. The reading that the budget failed because it was denominated in the wrong unit is my argument about the case, not something Uber said. The shape is still hard to miss. A budget built by multiplying a per-person license cost by headcount forecasts correctly only if per-person consumption is roughly stable. Agentic tooling breaks that stability by design - which is the reason GitHub gave for changing its billing in the first place. Inside one seat, a chat question and a multi-hour autonomous session had been costing the same. ## The vendors are not coming to fix this The comfortable assumption is that pricing will sort itself out and take the planning problem with it. Bain’s numbers say otherwise. In that October 2025 analysis, Bain looked at more than 30 SaaS vendors adding generative AI capabilities, explicitly excluding the AI-native ones. Roughly 35% had simply raised per-seat pricing and bundled AI into existing tiers, Zoom given as the example. About 65% had gone hybrid, layering an AI meter on top of seat-based pricing, with Adobe and Salesforce named. None had fully shifted to usage- or outcome-based pricing. Bain is not a neutral observer here - it sells pricing strategy consulting to the software vendors it is describing, and the piece reads as advice to them. That interest cuts toward the finding rather than against it. A firm whose business is helping vendors reprice is reporting that its market has not repriced. Bain also lays out why the transition is hard, and the obstacles are structural rather than reluctant. “Most software companies lack the product telemetry or IT, billing, and finance infrastructure to support these models at scale.” Sales reps, it notes, “typically have been trained to sell seats,” and moving them off it would take new playbooks, new tools and new compensation models. On the buying side, procurement teams “typically are accustomed to buying software based on headcount, not value,” so the change requires “shifting budget lines from labor to software.” That last line is the one to sit with. The seat is not only a vendor artefact. It is on our side of the table too - in the procurement process, the budget line, and the planning cycle that produced both. ## The seat is one instance of a much larger inheritance Here is the part I actually want to argue, and the budgeting problem is only the visible edge of it. If we build a company on tools somebody else designed, we inherit that designer’s assumptions about how a company is supposed to work. Most of the industry is building AI tools to help us do what we already do, only faster. I think that premise is the mistake. The interesting move is not making the existing shape more efficient. It is noticing that the shape was never ours. Per-seat pricing assumes work is done by named humans, one at a time. It is not the only assumption of its kind sitting in a standard stack: - A ticket that requires a single assignee assumes work has one owner. - A sprint measured in person-days assumes [capacity is a function of headcount](https://varops.com/more-agents-is-not-more-capacity-heres-what-anthropics-red-team-measured-when-the-work-overlapped/). - A review step that requires a named human approver assumes the reviewer is a person. - A per-person license count assumes access and throughput are the same quantity. Each of those is a claim about how a company operates, made by a vendor who has never seen the company. None of them was chosen. They arrived with the tools and quietly became the shape of every planning document built on top of them. That is the part that outlives the pricing story. Pricing models will keep moving. The assumptions that came bundled with the tooling stay until somebody goes looking for them. ## The audit Short, and it runs against documents that already exist. No new instrumentation, no new tooling. 1. **List every planning number denominated in people.** License counts, headcount plans, capacity forecasts, per-person budget lines, per-person tool allowances. 2. **For each one,** [**name what it is actually trying to measure**](https://varops.com/the-goal-was-never-to-use-ai/)**.** There are only three answers: access (who may use the thing), throughput (how much work gets done), or spend (what it costs). 3. **Where the answer is access, leave it alone.** A seat is still the right unit for a permission. 4. **Where the answer is throughput or spend, the unit is wrong.** Re-denominate: throughput in units of work completed, spend in whatever unit the vendor actually meters. 5. **Check the forecast against one month of real consumption data.** For any tool now billed on usage, the invoice is the measurement. If the forecast was built before the tool changed its meter, it is measuring something the vendor has stopped selling. Step 5 is the one that produces a number. The first four are just finding out which numbers to point it at. ## Where that leaves us The audit will not produce a lower bill. It produces a forecast that can be wrong in a bounded way, instead of a budget that is gone in month four. The seat was a good unit while access and work were the same quantity. They are not any more, and on Bain’s published evidence the vendors are keeping the seat regardless. Which means we re-denominate on our side, or it does not happen.