PRINCETON, N.J.--The rush to deploy artificial intelligence across the enterprise is creating an unexpected challenge: paying for it.
While AI has been promoted as a productivity engine capable of replacing routine work and reducing labor costs, BankInfoSecurity reports a growing number of technology and finance executives now believe the economics are becoming far more complicated. Gartner forecasts that by 2028, the cost of enterprise AI coding will exceed the average salary of the software developers the technology is designed to augment or replace.
The shift is being driven by exploding token consumption—the computational units large language models use to process requests—and a rapid move away from predictable subscription pricing toward consumption-based billing. As organizations deploy increasingly sophisticated AI agents capable of performing complex tasks with minimal human supervision, computing costs are climbing far faster than expected.
The trend is already reshaping how major technology companies manage AI. According to BankInfoSecurity, Uber, Microsoft and Meta have begun tightening oversight of AI usage as executives shift their attention from encouraging adoption to demonstrating measurable returns on investment.
GitHub recently moved its Copilot coding assistant from flat-fee pricing to usage-based billing, resulting in larger invoices for some enterprise customers and prompting some organizations to seek lower-cost alternatives. Amazon reportedly removed an internal leaderboard designed to encourage AI adoption after engineers aggressively competed to maximize AI use.
The phenomenon has earned its own nickname: "tokenmaxxing."
"It's super easy to burn tokens. Anybody can do that," said Andrea Malagodi, CTO at code quality and security platform Sonar. "Measuring tokens has the same effect that lines of code had, or story points. It's an output metric. It doesn't have anything to do with outcomes."
Malagodi said the industry has seen similar behavior before. Developers were once rewarded for writing more lines of code, while agile teams learned story points could become games instead of productivity measures. Token consumption, he said, risks becoming another misleading metric that rewards expensive computation rather than business value.
For chief financial officers, AI presents a budgeting challenge unlike traditional software.
"SaaS is deterministic in terms of the way it works and the costs," said Greg Henry, CFO at 1Password, a password management company. "Token-based pricing just flips that because consumption is unpredictable. It can spike in ways that traditional budgets were never built to absorb."
Unlike conventional software licenses, AI use spreads organically across organizations as employees discover new applications. At the same time, much of the detailed usage data finance teams need resides inside vendor dashboards rather than financial reporting systems.
BankInfoSecurity reports that technical architecture is also becoming a major cost driver. Malagodi said AI agents working through large legacy codebases must repeatedly reread the same software to rebuild context before completing tasks, effectively transforming technical debt into an ongoing infrastructure expense.
Even falling AI prices may not solve the problem.
According to Gartner Vice President Analyst Will Sommer, declining token prices will be overwhelmed by exploding demand and increasingly sophisticated models.
"Token demand and model complexity at the frontier of innovation are going to grow faster than token costs will fall," said Sommer.
A key driver is agentic AI. Gartner estimates autonomous AI agents consume five to 30 times more tokens per task than traditional chatbot interactions because they independently perform multiple steps while continually processing information.
BankInfoSecurity also reports enterprise buyers face another obstacle: pricing transparency.
Many AI vendors provide little visibility into how token consumption is measured or billed, making forecasting difficult.
Nikhil Mishra, co-founder and CTO of AI usage-billing platform Flexprice, believes vendors are creating customer frustration by delaying discussions about usage-based pricing until customers are already dependent on their platforms.
"They never built the expectation in the customer's mind that this consumption has a real cost behind it," Mishra stated in the Bank Info Security report.
He recommends vendors begin displaying detailed consumption metrics immediately—even under flat-rate subscriptions—so customers understand future pricing before billing models change.
Not every AI leader believes restricting usage is the answer.
Cars24, an online used vehicle marketplace, recently launched a $20-million AI initiative and deliberately chose not to impose token limits.
"What people are not understanding is the cost of intelligence has come down," said Jayesh Gupta, Cars24's head of AI, in the Bank Info Security report.
"You can't restrict on token budgets at all. That's what we believe inherently," Gupta said. "Instead of focusing on how to control the input, people should focus on how to control the output."
That debate reflects a growing divide inside executive suites. Chief AI officers generally want more computing capacity, CIOs worry about efficiency, and CFOs focus on financial discipline, Bank Info Security noted.
Henry believes organizations can achieve both goals.
"I land in the world of and's, not or's," Henry said. "I think we can do all of it. But you need a tool or something that gives you the visibility to help manage it. And when you can do that, then I think you can achieve AI-native status that CEOs want."
Henry cautioned that productivity gains alone do not guarantee business value. Developers may finish work faster, but organizations must ensure that freed-up time creates higher-value outcomes.
"It can't be just a finance-driven exercise," he said in the Bank Info Security report. "Without having this sort of single pane of glass, know where all the AI spend is happening, there's just no way to ultimately get that business outcome and ROI that you're really searching for."
According to BankInfoSecurity, the lesson echoes the early days of cloud computing: organizations should carefully meter AI usage, route workloads intelligently, maintain code quality to limit technical debt, and tie AI spending directly to measurable business results before committing to large long-term contracts.