
New Delhi, Nov. 6 -- For a few months this year, many large enterprises treated AI token consumption as a stand-in for productivity. Between January and April, as tools such as Claude Code, Codex and Cursor grew more agentic, internal leaderboards began ranking employees by the tokens they burned.
Then the invoices arrived. Leaderboards came down, spending caps went up, and by mid-year the business press was calling time on usage as a stand-in for success. The question had moved from how much AI people were using to what that use was returning.
Tokens are the units of text, code and data that large language models process, and most providers bill by them, often alongside subscriptions or per-seat licenses. Business leaders who understand token economics early will build leaner, more intelligent organisations. Those who dismiss it as an engineering detail risk losing control of one of their fastest-growing cost centres.
From License Seats to Token Flows For decades, enterprise software was priced by the seat. You paid for the number of employees who could access a tool, regardless of how much or how little they used it. Cloud computing started chipping away at this status quo with consumption-based pricing, AI has sped that up in three ways.
Providers increasingly charge by usage, so two developers on the same team can run up very different bills depending on how they work. Subsidised flat-rate plans are also giving way. A heavy user on a fixed subscription could consume far more than the price implied, and the caps and rate limits now appearing across the market suggest that subsidy is being withdrawn. The biggest factor is agents. An agentic workflow can consume many times the tokens of a chat prompt, and per-developer consumption has multiplied without a matching rise in output.
This forces finance and operations teams to look at procurement, budgeting, and vendor evaluation through a new lens. The question is less about "How many seats do we need?" and more about "How many tokens will this workflow consume, and what is the return on each?"
When Usage Became the Target The early-2026 playbook measured something easy to tally and easy to game, and people gamed it. This is Goodhart's Law at work. Once a metric becomes a target, it stops being a good metric. Tokens have become the new lines of code, a number that records activity and says little about whether the work was any good.
Faros.ai's 2026 AI Engineering Report, built on multi-year telemetry from thousands of developers, found teams completing more tasks and closing more epics under high AI adoption. Quality moved the other way. Bugs per developer climbed, review times stretched, more pull requests merged with no review at all, and code churn rose sharply. Throughput tells you what went out the door. It says nothing about what survived.
Putting Every Token Where It Earns the Most The instinct after a spending scare is to cap everything. That carries its own cost. Starve an agent of context or force it onto an underpowered model and you get more retries, more tool calls and more human rework. The savings on the invoice leak back out through the workflow.
The better approach sits between the two extremes. Spend where the model earns its return, route the rest to something cheaper, and judge each task by what it produced.
The companies doing this well share a few habits. Smaller tasks go to smaller models, and code quality, bugs fixed and reverted pull requests get tracked next to spend. Deterministic steps stay deterministic, with a model brought in only where a task calls for judgement. These companies also tolerate a handful of power users who spend well above the average.
A reasonable default sends most work to cheaper or open models and saves frontier models for the tasks complex enough to warrant them.
Cost Per Outcome Is the Number to Watch If tokens are the unit of AI consumption, then cost per outcome becomes the clearest proxy for how intelligently an organisation uses AI.
Consider two companies in the same industry. with the same AI spend. One resolves far more support tickets and produces far more internal analysis per token. That company has a stronger AI operating model because it converts consumption into business output more effectively.
The boardroom question should move from "Are we using AI?" to "how much value are we extracting per unit of AI consumed?" Answering it starts with a handful of outcome metrics per team, such as cycle time, rework, defects, vulnerabilities closed and hours redeployed. Token spend sits next to them as a cost line.
Leaders also need visibility into spend by team, workflow and model before the invoice arrives. Per-engineer budgets with one-click overrides, circuit breakers and human review of the heaviest spenders keep costs in view without choking useful work. Efficiency is an engineering skill too, and context engineering, task decomposition and knowing when not to use an agent belong in developer training.
The correction is giving rise to AI Financial Operations (AI FinOps), as organisations manage token consumption with the same financial rigour they already apply to cloud spending. The FinOps Foundation's open billing standard, known as FOCUS, is set to add native token tracking in its next release, a sign the discipline is formalising. For many organisations, AI FinOps is becoming the practical layer that turns AI from an experiment into something that can be budgeted, governed and scaled responsibly.
The Bottom Line Tokens give enterprises a unit for machine work, and strategy will have to adapt to it. The first question for any leadership team is what it rewards today. If usage leaderboards or "share of code written by AI" targets are still running, the incentives point the wrong way. The next question is where frontier-model spend earns its premium and where it has become the default by habit.
Organisations that build discipline around AI spend now will be better placed to capture long-term value. Markets rarely wait for consensus; they tend to reward those who act before the change becomes obvious.
Published by HT Digital Content Services with permission from MINT CIO CIRCLE.