For much of the past year, many companies pushed employees to use more AI, especially in software development. The idea became known as tokenmaxxing, meaning maximizing AI usage with the assumption that more model consumption would eventually translate into more productivity, faster delivery, and better outcomes.
That experiment is now running into the finance office, and enterprise AI has entered its first cost-accountability phase.
The New York Times reported Meta plans to limit AI use after seeing an “exponential increase” in costs, while Uber said in May it had used its projected AI spending for the year in just four months and placed monthly limits on some AI coding tools. Walmart also set limits for different AI tools, and Amazon and Meta removed internal token-use leaderboards.
The Next Web reported a similar pullback, saying companies that once ranked employees by how much AI they used are now capping usage, steering teams toward cheaper models, or removing incentives that rewarded token consumption over useful work.
IBM argued on June 25 the counterreaction should not be simple token minimization. Cutting token use can lower visible spend, but it can also remove the business context, technical constraints, and task detail that AI systems need to perform well. The result can be more retries, tool calls, validation work, and human rework.
The real issue is not whether enterprises use more AI or less AI, but whether they can prove what the AI produced.
Analysis
What this means: AI’s enthusiasm is being replaced by adoption skepticism. The first wave rewarded employees and teams for using AI because usage was easy to measure. The next wave will ask whether that usage reduced delivery time, improved code quality, resolved vulnerabilities, accelerated modernization, or simply shifted cost into a new line item.
Usage Was the Easy Metric
Tokenmaxxing took hold because companies needed a fast way to signal AI adoption.
Token counts gives leaders something visible. They can see who is using tools, which teams are active, and how quickly AI is spreading through the workforce. In the early experimentation phase, that was useful. It encouraged employees to try new tools and pushed AI out of innovation labs into daily work.
But usage is a poor substitute for value.
IBM cited Neil Dhar, SVP for IBM Consulting, who said organizations created usage leaderboards in the absence of true metrics, and employees quickly learned to game them. That is the core problem with tokenmaxxing. It treats volume as maturity.
The flaw becomes starker in agentic development. AI agents do not just answer prompts. They inspect repositories, plan tasks, call tools, generate code, test solutions, revise work, and run multi-step workflows. That can be powerful, but it also means a single task can burn through far more compute than a chatbot interaction.
When companies reward consumption without tying it to outcomes, AI adoption can look successful right up until the bills arrive.
The Bill Came Due
Per the New York Times, OpenAI and Anthropic sell individual subscriptions, but major enterprise revenue comes from companies paying for the tokens used by large workforces. A simple AI task may use a few hundred tokens, while complex coding work can use tens of thousands.
That explains why AI budgets are tightening first around engineering work. Developers can generate meaningful productivity gains with AI, but coding agents also consume a lot of tokens because they work across codebases, reason through dependencies, and iterate on solutions.
Uber COO Andrew Macdonald said the company needed a clearer line between AI spending and useful features or functionality shipped. Salesforce CEO Marc Benioff said Salesforce now tracks “agentic work units” rather than tokens, shifting attention toward output rather than raw consumption.
That is the direction enterprise AI measurement is likely to move. Token counts may remain useful for cost monitoring, but they do not answer the question CFOs and CIOs now care about most: What did the business get for the spend?
Analysis
What this means: Token counts are not AI ROI. They can show where money is going, but they cannot show whether a process improved, a backlog shrank, a release moved faster, or a finance team closed the books with fewer exceptions. Enterprise buyers will need better AI metrics than usage dashboards if they want to defend adoption after the experimentation budget runs out.
More Value, Not More Volume
IBM’s answer is “valuemaxxing,” meaning a shift from measuring AI consumption to measuring AI outcomes.
That means asking how many tasks were completed, how much developer time was saved, how much modernization work moved forward, how many vulnerabilities were resolved, and how much rework was avoided. Those are the measures that connect AI activity to business value.
For ERP teams, this is where the conversation becomes practical. AI assistants and agents are being added to finance, supply chain, procurement, HR, customer service, IT, and software delivery workflows. Each one creates a cost curve. Each one also needs a business case that goes beyond whether employees used it.
A finance agent should reduce exceptions, speed reconciliation, improve forecast accuracy, or lower manual close effort. A supply chain agent should improve decision speed, reduce planning friction, or help teams respond faster to disruptions. A developer agent should help modernize legacy code, improve release quality, or reduce support debt.
The metric cannot be “more AI.” The metric has to be better work.
Model Routing, Not Model Worship
The cost reset is also changing how companies think about models. The New York Times cited AT&T Chief AI Officer Andy Markus saying companies can save as much as 90% by using less advanced AI models for tasks that do not require frontier capabilities. The latest and most powerful model is not needed for most use cases.
That points to a more mature AI operating model. Enterprises will not stop using frontier models. They will reserve them for the work that needs the most reasoning, context, or precision. Simpler tasks will move to smaller, cheaper, or open-source models.
IBM makes a similar argument around model orchestration. As models become infrastructure, competitive advantage shifts toward the systems around them: context management, routing, memory, governance, evaluation, workflow orchestration, and cost optimization.
ERP and enterprise software vendors are embedding AI across business processes. Customers will need to understand not only which model powers a capability, but how the system decides which model to use, what data is sent, how context is managed, and how cost is controlled.
Analysis
What this means: The model decision is cost-control architecture. Companies that default every task to the most expensive model will struggle to scale AI economically. The more durable strategy is routing the right task to the right model with enough context to succeed and enough governance to keep spending visible.
Sponsor Industry-Grade Research
The ERP Cost Question
The tokenmaxxing backlash should sound familiar to ERP leaders.
Enterprise technology has gone through this cycle before. Cloud adoption created new flexibility, then cloud bills forced stronger FinOps discipline. SaaS sprawl made teams more agile, then procurement and IT had to rationalize overlapping tools. AI is now entering the same stage.
The lesson is not to starve AI usage. Overcorrecting into token minimization can create its own problems if employees are pushed to use less context, cheaper models, or shorter prompts in ways that reduce quality and increase rework.
The better approach is to make AI accountable at the workflow level. That means connecting consumption to outcomes, setting budgets by use case, defining when frontier models are justified, and measuring whether AI changes business performance rather than just user behavior.
This will become part of the product conversation. Customers will ask how AI usage is priced, what is included, where consumption charges apply, how agents are monitored, and whether the vendor can show value by workflow. For customers, the buying question will shift from whether AI is available to whether it is economically sustainable in daily operations.
Analysis
What this means: AI pricing should be part of ERP due diligence. As agents move into finance, HR, procurement, supply chain, and IT workflows, buyers will need to know what happens when usage grows. The uncomfortable question is no longer whether a vendor has AI, but whether the customer can afford to make that AI part of everyday work.
