Every company just hired a third workforce. Nobody’s managing its budget.
AI is the fastest-growing, most under-managed spend category in business. It’s initiated by software, billed by the token, and invisible to every control built for people and vendors. Seeing it was chapter one.
Govern is a working concept for chapter two: policy, budget instruments, and enforcement for intelligence.
The problem
growth in average monthly AI token spend since January 20251
share of U.S. businesses now paying for AI services2
year-over-year growth in AI charges surfacing as employee reimbursements3
gap between the median company and the top decile in monthly AI spend per employee3
Every system a company uses to control spending makes the same three assumptions: a human decides to spend, at human speed, at a price agreed in advance. Payroll works this way. Vendor contracts work this way. The corporate card works this way, too. Swipe, policy check, approval.
Token billing breaks all three assumptions at once. Spend is initiated by software, at machine speed, at prices that float with usage. A prompt-template change can triple a bill overnight. An agent stuck in a retry loop can burn a quarter’s budget before Monday standup. And on the invoice, a junior engineer’s Friday-night experiment looks identical to production inference.
Finance teams know it. AI charges are leaking onto personal cards and coming back as reimbursements, up 3× in a year.3 The most sophisticated control publicly deployed at scale is a flat monthly cap per employee.4 That is a seat-era instrument aimed at a usage-era problem. A cap rations spend, punishes the teams whose AI use is actually working, and governs nothing.
The market’s answer so far is visibility: dashboards that show where the tokens went. Visibility is necessary. It is also where every product on the market stops. A dashboard reports the fire. Policy decides which fires can’t start. For people and for vendors, that policy layer took decades to build. For intelligence, it doesn’t exist yet.
What would it look like if it did? Generate one and find out. ↓
- Ramp Economics Lab, Business Spending Report, Spring 2026.
- Ramp AI Index, ramp.com/data, May 2026.
- Ramp AI Index dataset, May 2026: median monthly AI spend per employee $11.38; top-decile median $610.61.
- Reported enterprise per-employee AI budget cap. TechCrunch, June 2026.
STEP 1 · PROFILE
Pick a company
Choose a preset or set your own. Each one generates a full policy calibrated to published industry benchmarks.
INSTRUMENTS
Budgets for spend that thinks
A cap is not an instrument. It is a surrender: one number, set once, blind to whether the spend it blocks was waste or the best money the company spent that month. Volatile, machine-initiated spend needs instruments the way portfolios need them: tools that price variance instead of forbidding it.
Govern’s policies carry three. Variance bands hold each team to a tolerance around its envelope, so a 96% overnight jump in unit cost gets caught while a growing team’s healthy ramp does not. Response ladders make enforcement proportional: alert, then throttle, then block, each step reversible until the last. And routing economics make the default cheap: efficient models by default, frontier models by exception, because the spread between tiers is the single largest lever in any AI budget.
One number these policies never use: cost per token. Unit costs only mean something operational. Cost per resolved ticket. Cost per merged PR. Cost per qualified lead. A budget denominated in work can be governed; a budget denominated in tokens can only be watched. Strictness, in this scheme, never changes the size of a budget. It changes the controls around it.
Routing economics
The spread between model tiers is the lever. Drag it.
saved per year by routing
Prices are illustrative and editable. The ratio is the point.
The other side of the market is already moving. Model providers are building budget APIs, programmatic limits, and real-time cost telemetry into their platforms, because unbounded spend produces unhappy buyers and churn. At Stripe Sessions this year, Anthropic’s monetization platform lead described a near future where agents run thousands of tasks against cost envelopes they can query and tune, and called manual limit-setting impossible at that scale. The supply side is building the meter. The demand side, the policy engine that decides what the meter is allowed to record, is still unclaimed. Governing token costs and pricing agentic work are the same curve read from opposite ends. You cannot budget what you cannot meter, and you cannot price what you cannot govern.