An AI token budget per employee is a spending cap on model usage, set by role or team, with alerts and an approval path for going over. In 2026 it stopped being a theory. Business Insider reported on November 17 that a Salesforce memo gives every role a monthly token budget, from $100 for non-technical staff to $1,000 for technical sales roles. Uber was reported in June to cap coding-agent spend at $1,500 a month per tool. Coinbase went the other way in June and cut its AI bill nearly in half without hard caps, then was reported in the same November story to run weekly caps of $500 to $5,000.
Salesforce disputes part of the reporting. Below you get what each company did, with dates and sources, where the accounts conflict, and a policy you can copy: how to size budgets, where to cap, what to route before you cap, and which controls Anthropic's own admin tools give you today.
What did Salesforce actually do with AI token budgets?
According to an internal email from COO Miguel Milano, reported by Business Insider on November 17, 2026, Salesforce now sets monthly token budgets by role. The memo says some employees "spent thousands of dollars per month on tokens last quarter" and states: "Starting this month, every role has a monthly token budget, so we can focus our AI spend on the work that matters most for our customers and business" (MLQ.ai, Let's Data Science, both citing Business Insider).
The reported budget levels:
| Group (as reported) | Monthly token budget |
|---|---|
| Technical roles in the sales org | $1,000 |
| Technical roles in Customer Success, Professional Services and Forward Deployed Engineering | $500 |
| Non-technical roles, including sellers | $100 |
The same reports describe how it is enforced:
- When someone nears 100% of their limit, they and their manager get a Slack alert from an internal app called DevBar.
- Going over follows a ladder. A one-time overage of up to 20% needs the second-line manager. Between 20% and 3x the limit needs the third-line manager plus an ROI justification, approved for 90 days. Above 3x goes to the org's executive and finance, reviewed monthly.
- Org leaders manage the budgets as a pool, similar to travel and expenses, so one person's spike can be offset by others.
- The memo says a pilot team "was operating well under budget by month two with no drop in output."
What Salesforce says
A Salesforce spokesperson told Business Insider the company does not have companywide individual token budgets and that team leaders manage AI usage. The spokesperson's quoted line: "Our focus is always to ensure that our AI investments deliver the greatest impact for our business and customers" (as quoted by MLQ.ai and Let's Data Science).
So the two accounts differ. The memo, as reported, says "every role has a monthly token budget." The company says there is no companywide individual budget. One way both can be true is the pooled design: budgets set per role but owned and balanced by org leaders, not enforced as a hard per-person cap across the whole company. That reading is ours, not Salesforce's. The original article is Business Insider, November 17, 2026 (paywalled; we worked from the outlets above that cite it).
Why the number matters: in May 2026, Marc Benioff said on the All-In podcast that he expected Salesforce to spend about $300 million on Anthropic tokens in 2026, almost entirely on coding (The Next Web, May 16, 2026). At that scale, a $100 cap on non-technical roles saves little in dollars. What it does is keep the spend pointed at the work that pays back.
How did Uber cap AI spending?
Uber caps each employee at $1,500 a month per agentic coding tool, such as Claude Code or Cursor, Bloomberg reported on June 2, 2026 (TechCrunch, Investing.com via Yahoo Finance). Three details from that coverage:
- The cap applies per tool, so spend on one tool does not count against another. Investing.com's write-up says the limits cover only agentic coding software.
- Each employee can see their usage on an internal dashboard.
- The caps can be exceeded with permission.
The trigger: CTO Praveen Neppalli Naga told The Information in April that Uber had already used its full-year 2026 AI budget, about four months into the year (as reported by TechCrunch and Investing.com).
What happened next is worth more than the cap itself. Startup Fortune, citing a Fortune report from August 7, 2026, says Uber added prompt caching and smarter model defaults, and that the number of Uber employees using frontier AI tools has since quadrupled from January (Startup Fortune, August 8, 2026). We could not read the Fortune piece directly, so treat that figure as second-hand. The cap and the engineering fixes landed in the same months, and usage kept growing; none of the coverage we read says which mattered more.
What did Coinbase do differently?
Coinbase is the company where the reports conflict most, so here are both, in date order.
June 2026: cut spend, no hard caps. In a post on X, CEO Brian Armstrong said Coinbase had "cut our AI spend nearly in half, while our token usage continues to grow" (Analytics India Magazine, June 29, 2026). The levers he described:
- On model defaults: "We're experimenting with defaulting to open weight GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task" (Let's Data Science, June 29, 2026).
- On caching: "Our cache hit rate went from 5% to 60% in LibreChat once properly implemented" (Analytics India Magazine). That figure is for LibreChat only.
- On limits: Analytics India Magazine reports that Coinbase replaced hard spending limits with visibility into usage, and that 91% of employees had never reached their old usage caps.
November 2026: weekly caps reported. The Business Insider story on Salesforce also reported that Coinbase runs weekly AI spending caps of $500 to $5,000, depending on role and seniority (per MLQ.ai and Let's Data Science, citing Business Insider).
We found no public Coinbase statement that reconciles the two. Caps that 91% of staff never hit can sit alongside a visibility-first policy, so they may describe the same system from two angles. What is clear from Armstrong's own words is the order of operations: route and cache first, then decide whether you still need a hard limit.
The three models side by side
| Salesforce (reported Nov 2026) | Uber (reported Jun 2026) | Coinbase (Jun 2026 post, Nov 2026 report) | |
|---|---|---|---|
| Unit | Per role, monthly, pooled by org leader | Per employee, per coding tool, monthly | Per role and seniority, weekly (Nov report) |
| Amount | $100 to $1,000 | $1,500 per tool | $500 to $5,000 (Nov report) |
| Over the limit | Manager approval ladder, ROI note above 20% | Exceed with permission | Not described in the reports we read |
| Main lever besides the cap | Default to cheaper tools for drafting and research | Prompt caching, model defaults (per Fortune) | Gateway routing, open-weight defaults, caching |
| Visibility | Slack alerts to employee and manager | Personal usage dashboard | Visibility into AI usage instead of hard limits (per AIM) |
| Company comment | Disputes companywide individual budgets | Spokesperson: policy is meant "to responsibly encourage agentic AI adoption and experimentation at scale" (per Investing.com) | CEO post on X (June) |
How to set AI token budgets for your team
None of the three started with a number. Each started because a bill surprised finance. You can skip the surprise. This is the policy we would put in place for a company of 50 to 5,000 people, built only from the patterns above and the admin tools that exist today.
1. Measure 30 days before you cap anything
Pull spend by user, team, model and tool for a full month. Without that, a budget is a guess, and a guess set too low pushes people to personal accounts you cannot see. If you are on the Claude API, the Usage and Cost Analytics API groups cost by workspace, and we walk through the queries in our Claude usage and cost analytics API guide.
What you are looking for: the spread. You will likely find a long tail, which is what Coinbase's 91% figure and Salesforce's "thousands of dollars per month" for some people both point to. Budgets should be set for the middle of each role, with an exception path for the tail, not for the tail itself.
2. Budget by role and pool by team
Copy the Salesforce structure and bring your own numbers from step 1.
| Field | What to fill in |
|---|---|
| Role group | e.g. engineering, technical sales, support, everyone else |
| Monthly budget per person | Your measured median for that group, plus headroom |
| Budget owner | The team lead who can move headroom between people |
| Soft alert | Employee notified at a share of the limit you choose |
| Hard alert | Employee and manager at 100% |
| Exception step 1 | Small one-time overage, approved by the manager |
| Exception step 2 | Larger or repeating overage, with a written ROI note and an expiry date |
| Review | Monthly, against output, not only against spend |
Pooling matters. A per-person hard cap punishes the engineer who spends $900 in the week of a migration. A pooled team budget lets the lead absorb that and catch the person who spends $900 every week on work that does not ship.
3. Cap coding agents per tool
Coding agents are where the large individual bills come from: Benioff said Salesforce's Anthropic spend is almost entirely coding, and Uber's cap targets only agentic coding tools. A per-tool cap, as Uber did, stops one runaway agent loop from eating the whole budget and makes tool comparisons honest. Pair it with Claude Code rate limits so you know which ceiling a developer is hitting, the plan's or yours.
4. Route and cache before you tighten
Coinbase got its result from routing and caching while usage kept growing. Before you lower a budget, check these:
- Does drafting, summarizing and classification go to a cheaper model by default? Salesforce's memo reportedly steers staff to Slackbot and Gemini for that work. Our AI API cost optimization guide shows which routes to move and the model pricing table gives current per-token prices.
- Do long system prompts and repeated context hit the prompt cache? Coinbase's 5% to 60% jump in one internal app is a sign of how much is usually left on the table.
- Does every call go through one internal gateway that tags requests by team and project? That is what turns a monthly invoice into a chargeback report.
5. Use the controls your vendor already ships
For Anthropic, these are documented today (Workspaces docs, Spend Limits API, read November 20, 2026):
| Control | What it does | Where |
|---|---|---|
| Workspace monthly spend limit | Caps a workspace's monthly spend, below the org limit | Claude Console, Spend limits page |
| Email notifications | Alerts when a workspace reaches an amount you choose (if Anthropic bills you directly) | Same page |
| Claude Code workspace per-user limit | Monthly spend limit per person; the only workspace that supports per-user limits | Claude Console |
| Per-user spend limits (Claude Enterprise) | Daily, weekly or monthly limit per member via the Admin API (beta); seat-tier and group defaults in claude.ai (Admin API writes may need to be enabled for your org) | Spend Limits API |
| Usage and Cost Analytics API | Cost by workspace for chargeback | Admin API |
One workspace per team gives you team-level caps and team-level reporting with no extra tooling. Per-person caps beyond Claude Code need Claude Enterprise or your own gateway.
6. Review against output, not spend
A budget review that only asks "who went over" teaches people to stop using AI. Ask what the top spenders shipped. Salesforce's pilot claim, "no drop in output," is the test that matters, and it is the one you should be able to answer with your own data after the first two months.
Where budgets go wrong: a cap with no exception path. When the limit blocks real work and there is no fast approval, people move to personal accounts, and you lose both the spend data and control over what company data goes into which model. Ship the approval ladder on the same day as the cap.
When you need outside help
You can run this policy in-house if you already have someone who owns AI spend and can read a usage export. Where teams usually get stuck is the middle: deciding which workloads move to which model, setting up a gateway, and writing a policy that finance and engineering both accept. That is the work our fractional CAIO team does. If your problem is that developers are spending a lot on coding agents without getting much out of them, Claude Code training is often the cheaper fix than a lower cap.
FAQ
What is an AI token budget per employee?
It is a cap on how much an employee or role can spend on AI model usage in a set period, usually monthly or weekly. Salesforce reportedly sets them by role from $100 to $1,000 a month, and Uber caps coding-agent spend at $1,500 a month per tool. Good budgets come with alerts and an approval path for going over.
Did Salesforce really cap employee AI spending?
Business Insider reported on November 17, 2026 that a memo from COO Miguel Milano gives every role a monthly token budget. A Salesforce spokesperson said the company does not have companywide individual token budgets and that team leaders manage AI usage. Both statements are on the record. In our reading, the memo's design, with budgets pooled by org leaders, may explain the gap.
How much does Uber let employees spend on AI coding tools?
Uber caps each employee at $1,500 a month per agentic coding tool, such as Claude Code or Cursor, according to Bloomberg's June 2, 2026 report. Spend on one tool does not count against another, employees can see usage on a dashboard, and they can exceed the cap with permission.
How did Coinbase cut its AI spend in half?
CEO Brian Armstrong said in June 2026 that Coinbase cut AI spend nearly in half while token usage kept growing. The levers were routing through an internal LLM gateway, experimenting with open-weight models as defaults, and prompt caching. A November 2026 Business Insider report separately said Coinbase runs weekly caps of $500 to $5,000 by role and seniority.
How do I set an AI budget for my team?
Measure spend by person and team for 30 days, set budgets per role from the median, and let team leads pool the headroom. Add alerts at 100% to the employee and manager, a two-step exception path, and a monthly review against output. Our fractional CAIO team can set this up with you.
Can I set per-user spend limits in Claude?
Yes, in two places. The Claude Code workspace in the Claude Console supports per-user monthly spend limits, and Claude Enterprise organizations can set daily, weekly or monthly per-user limits through the beta Spend Limits API. Other Console workspaces support workspace-level monthly limits.
Should I cap AI spend or optimize it first?
Optimize first, then cap what is left. Coinbase's results came from routing and caching, and Uber's usage grew after it added caching and better model defaults. A cap without those fixes mostly limits useful work.
Sources: Business Insider, Nov 17, 2026 (paywalled), MLQ.ai, Let's Data Science on Salesforce, Startup Fortune on Salesforce, The Next Web, May 16, 2026, TechCrunch, Jun 2, 2026, Investing.com via Yahoo Finance, Jun 2, 2026, Startup Fortune on Uber, Aug 8, 2026, Analytics India Magazine, Jun 29, 2026, Let's Data Science on Coinbase, Jun 29, 2026, Anthropic Workspaces docs, Anthropic Spend Limits API
Book a Free Strategy Call
Building this in production?
Walid runs a 30-min call to map your AI engineering team. Free, no slides.

Taha builds and ships custom AI agents and workflow automations for AY Automate clients across SaaS, finance, and professional services.
