Token budgets are the new phone bill.
First companies told staff to use as much AI as possible. Then the invoices arrived. Now everyone’s writing token budgets — mostly badly.
The cycle ran fast. Early this year, big tech firms openly encouraged “tokenmaxxing” — consume as much AI as you can, productivity will follow. Within months, one large tech company had burned its entire annual AI-coding budget in four, with individual engineers running up $500–$2,000 monthly token bills before per-engineer caps arrived. Budgets are now the norm; there’s just no consensus on the number. Surveyed companies set them anywhere from $250 a month to tens of thousands, with one pattern recurring: juniors capped around $800, seniors and data scientists at $1,600–$4,000.
Meanwhile the spread across companies is enormous. Payment-platform data shows a 99th-percentile customer spending nearly $90,000 a year per employee on AI, while the median customer spends about $136 — and the median Fortune 500 is still below $100 per employee. Whatever you’re spending, someone in your industry is spending a hundred times more, or a hundred times less.
Figures above: payment-platform data, surveyed company caps — checked 2026-07-06
Companies in the top quintile of productive token use are reporting ~16.5% revenue growth. The bottom quintile: ~5%. The gap isn’t spend. It’s spend with a return.
A cap is not a strategy
Most token budgets we see are written like phone-bill caps: a number picked to stop the pain, applied evenly to everyone. That’s how you get the pathologies — staff downgraded to weaker default models, premium tiers switched off, employees gaming a second tool’s allowance to stretch their own. The spend goes down; so does the return. Nobody measures the second part.
A budget that works starts from the other end: which workflows demonstrably pay, and what do they cost to run? Fund those without friction — a senior engineer’s $2,000 token month is trivial against their loaded cost if it ships work faster. Cut the spend that returns nothing, which is rarely the biggest line and usually the least examined one. And give the cap an owner who sees both the invoice and the output, not just the invoice.
The blunt version
If you can’t say what a token earns you, a token budget is just a smaller version of the same mystery. Baseline first, budget second. It’s the same discipline as every technology wave before this one — and it’s precisely what a two-week Reality Check establishes.
- Token budgets are now near-universal — but most are set to a pain threshold, not to a return threshold.
- The spread is extreme: from $136 median to $90,000 per employee at the 99th percentile. Your peers’ benchmarks are almost certainly wrong for your context.
- Top-quintile productive token users report ~16.5 % revenue growth; bottom-quintile report ~5 %. The variable is return-on-token, not token volume.
- Flat caps applied uniformly across roles create perverse incentives: weaker models, tool-hopping, and invisible productivity loss.
- Baseline before you budget. If you can’t say what a token earns, the cap is just a smaller mystery.