Share with your CFO
Boris Cherny, creator of Anthropic’s Claude Code, is pushing enterprises to retire token burn as their primary AI success metric. In a thread on X, Cherny laid out a four-step enterprise adoption framework and landed hardest on measurement: usage dashboards track activity, not return. The right question is whether the company would have spent engineering hours on that task anyway, and if so, what those hours cost. That delta is your actual AI return on investment. The timing is deliberate: Jamie Dimon and Sam Altman both flagged AI ROI as a board-level pressure point this week.
What this means for your business
Token burn is an input metric masquerading as an output metric. Tracking it to evaluate AI ROI is like measuring electricity consumption to evaluate factory productivity. You might spend more on AI tokens in a quarter where engineering output doubles, which makes the token number useless for any real financial judgment. Cherny’s substitution, cost of avoided engineering hours, is at least output-adjacent. It ties AI spend to something a CFO already knows how to price.
The avoided-hours framework has a real limitation worth naming: it only captures displacement of existing work, not creation of new capability. Cherny actually acknowledges this, calling the bigger payoff the moment when maintenance and bug-fixing happen in the background and teams start “doing things that weren’t even in range before.” That second category, net-new output, is where the real financial argument lives, and it’s also the hardest to put in a spreadsheet. The companies that figure out how to measure it will have a material advantage in AI budget conversations with boards.
The signal worth watching: the “tokenmaxxing” era, when enterprises competed on raw AI consumption as a proxy for innovation, ended faster than most predicted. Coinbase and Vercel are already publishing cost-reduction playbooks. That cultural shift from volume to value will land in procurement conversations within two quarters, and any enterprise still reporting AI success by token count will find that metric increasingly difficult to defend to an investor audience that has absorbed Dimon’s framing.
Concept deep-dive: Token burn
Token burn refers to the volume of text tokens, the units AI models use to process and generate language, consumed by an organization’s AI tools over a period. It exists as a metric because tokens are the billing unit for most AI APIs, making them easy to track. Think of it like measuring a law firm’s success by the number of pages printed. The count is real and costs money, but it says nothing about whether the work produced value. For finance teams building AI ROI models, token counts are a cost line, not a return line.
Based on reporting from Claude Code’s creator offers a better way to measure AI success than token burn, originally published 2026-07-17 02:09:00.
