Tokenmaxxing-Here's why heavy usage and good outcomes are not the same thing.

You open ChatGPT to debug a function or pull together some market research. Twenty minutes later, you have hit the limit. The model goes quiet. You either pay up or wait until midnight when your token window resets — and your AI-assisted productivity with it.

Now imagine this happening in an enterprise - A financial services firm approved a $400,000 annual AI budget in Q1. By Q3, they had consumed 70% of it. The head of technology presented the usage dashboard to the board: AI adoption was up 340%. Active users across every department. Thousands of requests per day.

The CFO asked one question: "Show me what we produced”. Nobody had that number.

Tokenmaxxing is the emergent behavior where employees — acting entirely rationally from their own perspective — over-engineer prompts, retry requests repeatedly, and consistently route tasks to the most capable (and expensive) model available, regardless of whether the task requires it. In the enterprise, it happens without any malicious intent. It is simply what happens when there are no guardrails and no feedback loop between cost and outcome.

How it happens — the behavioural mechanics

Tokenmaxxing is rational from the individual employee's perspective and irrational from the organization's perspective—let's understand the why.

1. The discovery loop. Employees learn through experimentation that elaborately structured prompts — with context, examples, role-playing instructions, and explicit output formatting — produce significantly better results than simple prompts. This is true. The behavior is rewarded. So they do it for everything, including tasks that do not require it.
2. The model selection problem. Given a choice between GPT-4o and GPT-4o-mini, a rational employee always picks the frontier model. The cost difference is invisible to them. The quality difference, even if marginal, is visible. Without routing policy, every task escalates to the most expensive option.
3. The retry cost. When an output is not quite right, the path of least resistance is to re-run the prompt with modifications rather than edit the output manually. Three retries on a task that costs $0.04 per run is $0.16 instead of $0.04. Multiply across a department running 2,000 tasks per week and the retry tax alone adds up to tens of thousands of dollars per year per department.
4. The metric trap. Organisations measure AI adoption by usage volume because it is the easiest thing to instrument. High usage gets praised. This creates an explicit incentive to use AI for everything, including tasks where it adds no value, because usage itself is the metric being optimised.

The governance vacuum — who owns this problem?

1. The employee had no policy defining appropriate AI usage, no visibility into what their usage costs, and no feedback connecting their behavior to outcomes. They were told to use AI and rewarded for using it heavily. They did exactly what they were asked to do.
2. The manager was given an adoption target. AI vendors, internal champions, and consultants all told them that adoption drives ROI — the more your team uses AI, the more value you get. Nobody told them that unconstrained adoption without outcome tracking is how you spend $400,000 and produce nothing measurable.
3. The CISO was focused on security and access controls — who can use which tools, whether data is leaving the perimeter, whether vendors are compliant. Usage governance — how AI is used, not just whether it is used — was outside the traditional CISO mandate.
4. The CFO approved the budget based on productivity projections that were never tied to a measurement framework. AI vendors sell on projected ROI that assumes appropriate usage. The CFO had no mechanism to track whether actual usage matched the model that justified the budget.
5. The CIO/CTO made the infrastructure and vendor decisions but typically did not build outcome attribution into the deployment. Instrumenting usage is table stakes. Instrumenting outcomes — commits, tickets closed, documents produced, decisions made — requires deliberate design that most deployments skip.

The result: a governance vacuum in the middle. Everyone owns a piece of the perimeter. Nobody owns the usage layer.

Why this is a governance problem, not a cost problem

The temptation is to treat tokenmaxxing as a procurement or budget problem — spend less, negotiate better contracts, set hard limits. That is the wrong frame and it creates the wrong solutions.

Hard spending limits without use-case policy create a different problem: rationing. Employees hit limits during legitimate high-value work. The engineering team cannot finish the sprint because they hit the token cap. Meanwhile, the marketing team spent 60% of their allocation rephrasing the same email forty times. Limits without policy punish the wrong behavior.

The right frame is governance: matching AI capability to task requirements, routing decisions to appropriate models, measuring outcomes rather than activity, and creating a feedback loop between cost and value. This is what distinguishes a mature AI program from an expensive experiment.

The three governance mechanisms that actually address tokenmaxxing:

1. Use-case classification. Tagging requests at the gateway level by task type — code generation, document summarisation, data analysis, creative writing, communication drafting. This creates the data layer needed to understand where spend is going and whether it is producing outcomes.
2. Model routing policy. Routing tasks to the minimum capable model for the task type, with escalation to frontier models only for explicitly complex tasks or by explicit user request with justification. This does not block employees — it sets a sensible default that can be overridden.
3. Outcome attribution. Connecting AI usage events to downstream system activity — code commits, tickets closed, documents finalised — to build a cost-per-outcome metric. This is the hardest piece to build and the most valuable. A department spending $12,000 per month with a cost-per-closed-ticket of $4.20 is performing well. A department spending $8,000 per month with no measurable outcome attribution is a governance failure, not a success story.

What good looks like — Four questions every AI governance programme should be able to answer:

1. What tasks is AI being used for, by which departments, and at what cost per task type?
2. What is the cost-per-outcome for each department — not cost-per-request?
3. Which departments are routing tasks to expensive models for work that cheaper models handle adequately?
4. What is the retry rate — what percentage of costs come from re-running requests rather than first-pass use?

Thoughts?

-Srikant