Tokenmaxxing is dead. Finance teams hold the measuring stick.

By COGScontrol Team · July 9, 2026

Tokenmaxxing — the practice of incentivizing employees to consume AI tokens at maximum volume as a proxy for AI productivity — is over. Companies ran this experiment for two years. In 2026, the bills arrived without the ROI. Uber burned its entire 2026 AI coding-tools budget by April. Meta quietly dismantled its internal token-consumption leaderboard. Finance teams are now being asked to account for the spend — and token count is not a sufficient answer.

What was tokenmaxxing, and why did companies do it?

The logic seemed defensible at the time. AI productivity tools were new, adoption was uncertain, and the organizations that moved fastest appeared to win. The simplest signal that employees were actually using AI was the token bill: if consumption was high, people were engaged. Token volume became a proxy for adoption, which became a proxy for productivity, which became a proxy for competitive positioning.

Some companies made that logic explicit. Meta built an internal system called "Claudeonomics" that ranked roughly 85,000 employees by AI token usage and handed out gamified titles — "Token Legend," "Cache Wizard" — to the heaviest consumers. The message was unambiguous: use more AI. The question of what that usage was producing for the business was, in the meantime, deferred.

Why did tokenmaxxing fail as a business metric?

Tokens measure inputs. Business results live on the other side of a causal chain that tokens cannot see. When Andrew Macdonald, Uber's President and COO, was asked whether the company's AI spend had translated into more product shipped, his answer was telling: "Maybe implicitly there's more that is getting shipped, but it's very hard to draw a line between one of those stats and 'Okay now we're actually producing like 25% more useful consumer features.'" He also described the spend as becoming "harder to justify" — this from a company that had exhausted its entire 2026 AI coding-tools budget four months into the year.

The pattern repeated across the industry. In spring 2026, J.R. Storment, executive director of the FinOps Foundation, began hearing the same alarm from company after company: "In April and May, I started hearing from companies: 'Oh my god, we are 3x over our entire 2026 token budget and it's only April.'"

The survey data points the same direction. A January 2026 analysis found 56% of CEOs reported no revenue growth or cost reduction from AI despite significant investment. In both cases, the problem was not that AI failed to generate tokens — usage was high. It was that no one had connected those tokens to a business outcome anyone could defend.

The corporate pullback: what companies are doing now

The response has been swift and structural. Uber capped employee AI tool spend at $1,500 per month per tool after the budget overrun. Microsoft cancelled Claude Code subscriptions for employees in several key product divisions — at roughly $2,000 per engineer per month, the per-seat cost had become difficult to justify without productivity data. Meta took down the Claudeonomics leaderboard.

Even executives who remain committed to AI investment are shifting their framing. Marc Benioff, Salesforce's CEO, noted that his Anthropic bill would reach roughly $300 million this year and said he wished there were a "smart router" to direct queries to the right model tier based on what each task actually required — a value question, not a volume question.

Jeff Henry, President of Consulting at Highspring, reported that clients are pulling back until they "can really start to prove an ROI," with some waiting 12 to 18 months before making significant new AI spending decisions. The era of AI spend justified by enthusiasm is ending. The era justified by measurement is what comes next.

What comes after tokenmaxxing?

IBM has named the successor: "valuemaxxing." Where tokenmaxxing tracked how much AI was consumed, valuemaxxing tracks how many tasks were completed, how much developer time was saved, how many vulnerabilities were resolved, how much rework was avoided. The unit of account shifts from input to outcome.

This is not a new discipline. It is the same unit-economics logic that governs every other line of the P&L, applied at last to AI. Finance teams know how to do this. What they have lacked is the tooling to execute it against an AI cost base that spans multiple providers, sits inside cloud invoices, and arrives monthly with no breakdown by project or business outcome.

The attribution problem: one bill, every engineer

The first problem tokenmaxxing obscured is structural: the token bill that arrives from Anthropic or OpenAI does not name a team, a project, or a business outcome. One invoice covers every engineer's Claude Code sessions, every internal AI tool, every proof-of-concept that got abandoned in February. Finance receives a total and nothing else. Capping spend at $1,500 per seat per month is a cost control, not a measurement. It tells you how much you didn't spend; it still tells you nothing about what each team's AI investment produced.

The fix is a gateway. Route all LLM traffic through a gateway — LiteLLM is the standard open-source option — and issue each engineering team or project its own virtual key. The gateway is the only point in the chain where "which team made this call" is actually known. When the platform team holds a key named platform-team and the growth team holds one named growth-team, every token each team consumes is recorded against that name — during the build phase, and every sprint since.

How to turn token counts into cost per feature shipped

A gateway records token counts, not dollars. Negotiated discounts, committed-use tiers, credits, and taxes all live on the provider invoice — not in the gateway, and not in any public price list. Multiply gateway tokens by list prices and you fabricate a per-team cost that will not reconcile with what the company actually paid. Finance cannot defend a number that doesn't tie to the bill.

The reliable method inverts the naive one: allocate the real invoice across teams in proportion to each team's share of tokens.

The gateway supplies the weights; the invoice supplies the dollars. No price is invented; a real one is distributed. Attributed costs sum back to the invoice to the cent — a figure finance can take to close, or an auditor can check.

With per-team cost in hand, the value denominator does the rest. For engineering AI the right unit is cost per feature shipped — or cost per PR merged, cost per story point delivered, whatever the engineering org tracks at the cadence it reports. Each team's attributed spend divided by its output metric produces a unit cost that can be tracked quarter over quarter, compared across teams, and benchmarked against the pre-AI baseline.

One discipline keeps the value side honest: measure output; don't survey for it. In a 2025 randomized trial by METR, experienced developers took 19% longer on tasks with AI coding tools while believing the tools had made them 20% faster. Self-reported time savings overstate the gain by a wide margin. The unit cost that survives a board question is derived from what actually shipped, not from what engineers estimated.

For a deeper treatment of the attribution mechanic, see measuring the ROI of internal AI initiatives. For the cost-side definition, see what is AI COGS. For the broader ROI framework, see how to measure the ROI of AI initiatives.

How COGScontrol makes this measurement possible

The arithmetic above is straightforward. The operational challenge is executing it continuously: provider invoices arrive monthly without team-level breakdowns, gateway token counts need to be joined to those invoices and allocated daily, and engineering output metrics live in Jira, Linear, or GitHub — not in the finance stack.

COGScontrol connects to a LiteLLM gateway with a read-only spend key — it can see usage and nothing else — and turns the real provider invoice into a per-team ledger, reconciled to the cent. Every dollar is attributed or explicitly unattributed; nothing is estimated from a list price. Engineering output metrics arrive via CSV upload, API, direct entry, or Google Sheets, and the platform divides attributed cost by each metric to produce cost-per-outcome views: cost per feature shipped, cost per PR merged, cost per sprint.

The output is not a token dashboard. It is the AI value management view finance teams can take to a board: what each team's AI investment cost, what it returned, and whether the unit economics are improving. Run your own figures through the AI unit economics calculator, or see the full system on the features page.