INDUSTRY
The AI token bill reached the org chart. Budgets belong on the key, not in a memo
6 min read
We built Qevron around an unglamorous conviction: every token metered, every key budgeted, every call attributed — because one day AI spend would stop being a rounding error. On 18 November that day got a memo: Salesforce capped employee AI spending at $100 to $1,000 a month by role, with a Slack nag at 100% and a 20% overage you can apply for — after staff burned thousands a month. Meanwhile AT&T reports some 45 billion tokens a day across hundreds of autonomous workflows. The unit of AI spend is no longer the person. Here is why the budget has to live in the path.
We built our gateway around an unglamorous conviction: every token metered, every key budgeted, every call attributed to the team, app or agent that made it — because one day AI spend would stop being a rounding error and start being a line item someone has to defend. This week that day arrived with a memo attached. On 18 November, Salesforce — reportedly after employees ran up thousands of dollars a month on AI tools — capped monthly AI spending per employee: $1,000 for technical sales roles, $500 for technical customer-success roles, $100 for non-technical sellers, with a Slack notification when you near your cap and the right to apply, in advance, for a 20% overage. Read that mechanism twice: one of the world’s most sophisticated software companies is rationing intelligence through the org chart, one apology at a time. Meanwhile AT&T, by its own account, runs hundreds of autonomous workflows through some 45 billion tokens a day — spend no per-human cap will ever see. Last week Wall Street repriced the buildout’s debt; this week the enterprise met its own token bill. The question both weeks ask is the same: where does the budget actually live?
The memo is the symptom
Role-based caps with manager notifications are not a cost policy; they are what you improvise when the infrastructure cannot enforce one. Three things give it away. First, the enforcement point is social: a Slack nag and an approval workflow police the spend after the fact, at the layer of people, because the layer of calls is a blur of invoices from a dozen vendors. Second, per-employee caps do not map to value: the $100 seller whose one live deal genuinely needed $300 of research tokens is now filing a request, while a $1,000 allowance can evaporate into retries against a failing prompt without anyone asking whether the task was worth running at all. Third — and decisively — the unit of spend has already left the building: agents, not employees, are where token volume lives now, as AT&T’s numbers show, and an autonomous workflow neither reads memos nor fears its manager. The per-human budget is a 2023 abstraction applied to a 2026 cost structure. The unit of AI spend is the key, the task and the route — and that is the layer where a budget can actually bind.
Budgets belong in the path
Here is what the enforceable version looks like, and none of it is exotic. Metering at the gateway: when every call crosses one governed endpoint, every token is counted once, with attribution — this team, this app, this agent, this key — so the finance question “what did we spend, on what, and was it worth it” is a query, not an archaeology project. Budgets as properties of keys: a key carries an owner, a scope, a budget and an expiry, and when the budget is gone the key stops — no nag, no apology workflow, and revocation is one click, the same discipline we argued for when the threat was stolen keys rather than runaway spend. Routing as cost policy: the cheap, in-house model carries the routine load, the frontier call is an explicit route you price and permit, and caching stops you paying twice for the same answer. And for the agent fleets specifically, the discipline our control plane insists on: capacity, concurrency, priority and cost are four different things — squash them into one word and nobody ever knows which limit they hit — while completion-by-evidence stops an agent from re-running a confident failure at machine tempo, which is the most expensive token pattern there is.
The ad-funded middle
The metered gateway
Where we stand
Qevron is that path, productised: one OpenAI-compatible endpoint over our five in-house model families and 43+ external providers, with routing, caching, per-token metering and cost attribution built into the same surface that enforces keys, scopes and budgets. The point was never frugality for its own sake — it is that visibility precedes judgment: you cannot decide what AI work is worth until you can see what each unit of it costs, attributed to the team, the feature or the agent that incurred it. The same conviction runs through Kimene on the workforce side, where provider capacity, concurrency policy, leases and priority are kept as four separate, visible things and no task gets to declare itself done without evidence. A memo asks people to behave. A metered path makes the economics legible — and then lets you decide, with numbers, where generosity is the right policy.
- Every token metered at one governed endpoint — our models plus 43+ providers
- Budgets on keys, not people: owner, scope, budget, expiry — and the key stops when it is spent
- Cost attribution per team, app and agent — “was it worth it” becomes a query
- Routing and caching as cost policy: in-house for the routine, frontier by explicit route
- Agent spend disciplined at the control plane: capacity, concurrency, priority and cost kept distinct
What we will not claim
Three honest limits. First, the memo is not stupid. Salesforce acted on real waste, and a cap in a memo beats no cap anywhere; our argument is about where enforcement belongs, not whether restraint does — and a company that publishes its limits is being more honest about AI economics than most of its critics. Second, the numbers deserve their caveats: the Salesforce figures come from reporting on an internal memo, and AT&T’s token volumes are the company’s own account — read both as the shape of the trend, not audited fact. Third, a meter does not make a call worth making. Attribution shows you the waste; deciding what is worth running remains a human judgment, and no gateway feature substitutes for asking whether the task should exist. What the meter buys is the ability to have that argument with numbers instead of anecdotes — which is, in the end, all a budget ever was.
A budget in a memo begs for compliance. A budget on the key enforces itself — metered where the tokens flow, attributed to the work, revocable in one click.
For teams in Türkiye this pairs with last week’s story like two halves of one invoice: the buildout’s financing is being repriced above you, and the token bill is being rationed below you — and in a country where AI costs arrive denominated in foreign currency, visibility is not a luxury but a necessity on every CFO’s desk. Our answer is the same boring architecture at both scales: our own production-sized models for the routine load, one metered gateway over 43+ providers for reach, budgets that live on keys, and a cost ledger that makes unit economics visible before the memo becomes necessary — from Arpanet Bilişim A.Ş., engineered for the KVKK from the first line. The token bill is now an org-chart problem at some of the world’s biggest companies. Put it back in the path, where it can be a config instead.