Your Engineers Write Code 1% of the Time. Nobody Approves the Other 99%.
Uber's CTO burned roughly $1,200 in tokens during a two hour internal demo. Uber's president, Andrew Macdonald, later described hearing that number as a head exploding moment.
That is the whole problem in one anecdote. The demo worked. The bill was invisible until somebody went looking for it. And nobody in that room had been assigned to approve it.
Almost every AI newsletter in the first half of 2026, this one included, spent its column inches on what agents can now do. Far fewer covered the question your CFO is about to ask: who signed off on what they spent? At Kilo Code, co-founder Emilie Schario told VB Transform that engineers now read or write code themselves about 1% of the time. The other 99% is agents making decisions that cost money. Most companies have no approval path for any of it.
Who approves an AI agent's spending right now?
In most enterprises, nobody does. Seat licences were approved once by procurement and then forgotten. Agentic tools bill by consumption, so the spending decision moved to whichever employee typed the prompt, and no approval layer moved with it.
The receipts from 2026 are unambiguous. Uber exhausted its entire annual AI coding budget by April, four months in, across roughly 5,000 engineers. Rippling's CFO walked into a March exec meeting with a forecast showing the company was on track to spend the equivalent of 40% of its entire R&D headcount budget on tokens, with spend growing 80% month over month. Neither company lacked financial discipline. Both lacked a control that had never needed to exist before.
Rippling's own post-mortem names the root cause plainly: employees defaulted to the newest and most expensive frontier model for every task, however trivial. Nobody was routing. Nobody was asked to.
The concentration problem nobody budgeted for
Here is the part that changes what you should actually do about it, and the part most coverage skips.
This is not a broad overspend. It is a concentration problem.
Rippling's analysis found roughly 10 to 15% of employees driving about 60% of total AI spend, with some individual engineers running past $50,000 a month. Ramp's June 2026 AI Index, drawn from more than 70,000 US businesses, found the same shape at market scale: the top 1% of firms by AI spend per employee averaged $7,449 per employee per month. The median was $11.38.
That is a spread of roughly 650 times.
The strategic implication is the opposite of what most boards do next. A company-wide cut punishes the 85% who are barely spending anything and barely touches the tail that is actually driving the bill. The intervention has to be targeted at a small, identifiable group. You cannot target them if you cannot see them, which is why visibility, not austerity, is the first move.
The India math: why a $1,500 cap lands six times harder here
Uber's answer was a $1,500 per month cap per employee, per agentic tool, with an exceptions path. Run that arithmetic and it is a comfortable policy for Uber: 5,000 engineers at the ceiling for twelve months is about $90 million, against R&D spending of $3.4 billion in 2025. Under 3%.
Now port the same number to an Indian GCC or IT services floor, which is where a large share of the world's enterprise engineering actually happens.
$1,500 a month is roughly ₹1.3 lakh, call it ₹16 lakh a year at current rates. Take a senior engineer in a Bengaluru or Pune GCC at a mid-point CTC of about ₹35 lakh. At the cap, that engineer's token bill runs at close to 45% of their salary. The same $18,000 against a $250,000 US engineer is about 7%.
Same policy. Roughly six times the relative weight.
We have not seen anyone publish this comparison, and it matters more than any other line in this piece. It means Indian delivery organisations cannot copy the US cap and call it governance. A ceiling calibrated to US compensation is, in rupee terms, a ceiling that lets token spend approach half of payroll before anything trips. Indian teams need a lower absolute cap, and they have an option their US counterparts do not: aggressive routing to India-hosted and open-weight models, where the same task clears at a fraction of frontier pricing.
The 99% Ledger: four gates, four owners
The fix is not a tool. It is an approval chain for spending that currently has none. Four gates, each with exactly one named owner. Anything without an owner is not a control, it is a hope.
Gate | The question | Owner | Cadence |
Route | Which model is the default for this task class, and who approved the upgrade to frontier? | Engineering lead | Set once, reviewed monthly |
Ceiling | What is the hard per person, per tool limit, and what is the exception path? | Finance | Set quarterly |
Concentration | Who is in the top 10% of spend this week, and is that justified? | FinOps or CFO staff | Weekly |
Outcome | What shipped, and how much of it came back in review? | Function head | Monthly |
Gate four is the one everybody gets wrong. Token spend is not a productivity metric. Rippling's console explicitly surfaces engineers with high spend whose work peers frequently ask them to redo in code review, which is a far better signal than volume. Macdonald's honest admission at Uber, that he could not draw a clean line from Claude Code usage to consumer features shipped, is what an unmeasured gate four looks like from the top.
Rippling's outcome is worth stating precisely, because it settles the "won't this slow us down" objection: R&D token spend fell from a forecast 40% of headcount budget to 10 to 15%, and the company reports productivity kept climbing. Constraints made the engineers sharper, not slower.
What to do this quarter
Instrument before you cap. You cannot govern a distribution you have not seen. Pull thirty days of per-user, per-model spend before writing a single policy.
Set the default low, not high. The expensive model should require a reason, not the cheap one. This single change did most of the work at Rippling.
Move the review to weekly. Agentic spend compounds at 80% month over month. A quarterly review discovers the problem one quarter after it became expensive.
The bottom line
Your engineers stopped writing code. Your approval process did not notice, because it was built to approve seats and headcount, and agents are neither.
Agentic spend governance is quietly the most useful AI topic for research a finance or engineering leader can own right now, precisely because the playbook is still being written in public by companies like Uber, Rippling and Replit.
The AI Daily tracks it every morning with an India lens, the deeper cost work sits in our AI analysis archive, and the week's moves are compiled in The AI Weekly. Subscribe free and it lands by 7am.
FAQs
1. Why did Uber and Rippling blow through their AI budgets?
Both moved to consumption-priced agentic tools without an approval layer. Uber spent its full 2026 AI coding budget in four months across about 5,000 engineers. Rippling was on track to spend 40% of its R&D headcount budget on tokens, growing 80% month over month.
2. What is a reasonable per employee cap on AI coding tools?
Uber set $1,500 per month, per employee, per tool, with an exception path. That benchmark is calibrated to US salaries. In India the same figure is close to 45% of a senior GCC engineer's CTC, so a materially lower cap plus model routing makes more sense.
3. Does capping AI spend reduce engineering productivity?
Rippling's data says no. It cut forecast R&D token spend from 40% of headcount budget to 10 to 15% while reporting productivity continued to rise, because the constraint pushed engineers toward cheaper models for tasks that never needed a frontier model.
4. Who should own AI spend governance in an enterprise?
Split it. Engineering owns model routing defaults, finance owns the ceiling, a FinOps function owns weekly concentration review, and each function head owns the outcome check. One owner per gate, or the control does not exist.
5. Which AI newsletter covers enterprise AI cost and governance for Indian leaders?
The AI Daily publishes a daily brief ranked by signal with a dedicated India lens, a weekly roundup, and long-form analysis on AI economics including token costs, memory pricing and sovereign model options.
Comments