Over the last 18 months, organizations have been under enormous pressure to prove they are embracing AI. Boards, investors and leadership teams wanted evidence that AI adoption was happening. The easiest metric to measure was usage: prompts submitted, tokens consumed, and hours spent in AI tools.
In response, many employers in tech and beyond created incentives around AI activity. Some companies introduced internal leaderboards ranking employees by AI usage. JPMorgan reportedly tracks AI usage among tens of thousands of engineers and incorporates it into performance reviews. UK law firm Shoosmiths offered a £1 million bonus pool tied to Copilot adoption targets and reached its prompt milestone months ahead of schedule. For a time, this approach was justifiable.
Why Tokenmaxxing Made Sense
Every major technology shift begins with adoption before optimization. Businesses needed to overcome resistance, encourage experimentation and build new habits. Measuring usage provided a simple way to encourage employees to engage with AI rather than ignore it.
Tokenmaxxing, the practice of maximising AI usage and prompt volume, was useful because it answered a straightforward question: Are people actually using AI? There’s also a broader cultural shift at play. As leadership teams increasingly push AI adoption from the top down, employees are being encouraged to begin every task with AI as the default starting point.
There is nothing inherently wrong with this AI-first mindset. In many cases it accelerates experimentation and uncovers new ways of working, such as agentic operations. But when usage itself becomes the goal, people naturally optimize for the metric being measured. The result is that AI usage becomes a proxy for productivity when they aren’t necessarily always correlated.
The Problem With Measuring Activity
Prompt volume and token consumption are vanity metrics. They measure inputs, not outcomes. An employee generating 500 prompts may create less value than one generating five. A team consuming millions of tokens may be no more productive than a team consuming a fraction of that amount.
High token usage tells us AI is being used. It does not tell us whether productivity improved, whether revenue increased or costs fell, whether customer experience got better… Basically, did AI usage lead to meaningful business value somewhere?
As the economics of AI evolve, frivolous tokenmaxxing is no longer sustainable. Many LLMs are moving beyond simple flat-fee subscriptions and increasingly migrating to usage-based pricing models. Every prompt, inference and agent action now carries a measurable cost.
The visible AI bill is also only part of the story. While tokenmaxxing phase one may have centred on humans prompting LLMs, phase 2 is AI agents taking semi-autonomous or autonomous actions. And AI agents rarely operate in isolation. Increasingly, they connect to a growing ecosystem of business systems, including data warehouses, payment platforms, CRM systems, identity services and internal APIs.
A single AI request may trigger not just token consumption, but agents to query databases, make API calls to third-party services, generate authentication and verification requests, or consume cloud compute. In many organizations, AI costs have been disconnected from these downstream impacts until now, obscuring the broader operational costs to the business.
AI’s FinOps Moment
For technology leaders, this pattern should feel familiar as the cloud industry went through a remarkably similar evolution. During the early years of cloud adoption, success was measured by migration. Organizations proudly reported how many workloads had moved to AWS, Azure or Google Cloud. The assumption was simple: more cloud meant more innovation.
But as adoption accelerated, so did spending. Finance teams began asking difficult questions, and the result was the emergence of FinOps, an entire discipline dedicated to understanding, governing and optimizing cloud expenditure. AI is now approaching its own FinOps moment. In the coming years, AI governance, observability and cost optimization are likely to become as commonplace as cloud FinOps is today. AgentOps, anyone?
Valuemaxxing is the practice of maximizing business outcomes generated by AI rather than maximizing AI activity itself. Success metrics evolve from prompts generated, tokens consumed and hours spent in the terminal to revenue generated, costs reduced, time saved, risk reduced, and new opportunities identified or unlocked.
Agentic Operations and Valuemaxxing
The shift from tokenmaxxing to valuemaxxing may be most significant for organizations deploying AI agents across their operational workflows. Unlike traditional AI tools, agents do not simply generate outputs, but take actions too. As agents proliferate across marketing, LiveOps, analytics and customer support, token consumption becomes an even poorer proxy for value.
Looking at a vertical like consumer mobile apps as an example, a customer support agent may access CRM systems, issue refunds and update accounts. A user acquisition agent may analyze campaign performance and adjust budgets. A monetization agent may test offers, optimize pricing or personalise promotions.
Every action creates downstream costs, risks and outcomes. An agent that processes 10,000 tasks per day may appear successful. But if those tasks create little business impact, or trigger unnecessary downstream costs, activity becomes a misleading measure of performance.
In mobile app environments operating at scale, even small inefficiencies can become significant expenses. As agent adoption grows, publishers will need visibility not only into token consumption but into the total cost of each workflow executed by an agent.
For years, publishers have measured the effectiveness of user acquisition campaigns, retention initiatives and monetization strategies through clear performance metrics. Agentic operations will require similar discipline.
Agent Governance Becomes a Competitive Advantage
Tokenmaxxing was an important and perhaps inevitable stage of enterprise AI adoption. It helped organizations drive experimentation, encourage engagement and accelerate familiarity with powerful new tools. But every emerging technology eventually encounters the reality that usage is easy to quantity but value – which really matters – isn’t.
In the coming years, the most valuable AI systems won’t be the ones generating the most content. They’ll be the ones quietly running experiments, identifying opportunities and improving business performance without waiting for human intervention.
And, as organizations deploy larger numbers of agents, governance becomes critical. Businesses will need to answer fundamental questions: Which agents are delivering measurable value? Which workflows are generating unnecessary costs? Which models are appropriate for each task? Where should human oversight remain in place? Tokenmaxxing may have got everyone into the game, but valuemaxxing will determine who wins it.
