From tokenmaxxing to ROI-maxxing: Why enterprises are finally putting a price on AI

A particular news sent companies to take a step back. It was about an employee at a large company who managed to rack up an eye-watering bill of hundreds of thousands of dollars on Anthropic's Claude. This incident became folklore in enterprise technology circles. But it is also a sign of something larger - the bill for two years of AI experimentation and running pilots has arrived, and most companies were not ready for it.

The budget that vanished in four months

Uber had rolled out Anthropic's Claude Code to its engineering team in December 2025, with usage doubling by February. By April, the company had exhausted its entire annual AI allocation, which, as per reports, is roughly equivalent to its $3.4 billion R&D budget in four months! About 95% of Uber's engineers were using AI tools monthly. Around 11% of the company's live backend code was being written by AI agents. In June, Bloomberg reported that Uber had instituted a $1,500 monthly cap per employee per agentic coding tool.

Microsoft followed a month later. As per reports, in May, the company's Experiences and Devices division, the group responsible for Windows, Microsoft 365, Outlook, Teams, and Surface, was told to stop using Claude Code by June 30, 2026, which is also the last day of Microsoft's fiscal year. Engineers were directed to shift to GitHub Copilot CLI, Microsoft's own command-line coding tool. Per-engineer costs for Claude Code across the industry were running at between $500 and $2,000 a month. So, for a division the size of Experiences and Devices, that math compounds quickly.

These are not isolated incidents, but show the reality of what happens when pilots become full-scale.

A governance gap, not a pricing problem

The temptation to frame this as a cost story where tokens are expensive, models need to get cheaper, misses the point. "Reports of enterprises exhausting AI budgets are primarily a governance issue rather than a pricing issue," says Mahesh Makhija, Partner and Leader, Technology Consulting, EY India. "Most organisations today have mature controls around cloud infrastructure, software licences, and labour costs, but AI consumption currently occurs outside those governance frameworks."

In Deloitte's 2026 State of AI in the Enterprise report, based on a survey of 3,235 senior leaders across 24 countries, found that only 21% of respondents said their organisations had a mature governance model in place for agentic AI. This is happening even as agentic deployments scale at speed: by 2027, 74% of respondents in the same survey expect their companies to be using AI agents actively. The gap between governance and deployments is massive.

Gartner reports that by 2027, it predicts 40% of enterprises will demote or decommission autonomous AI agents because governance gaps are discovered only after production incidents, not before. Separately, Gartner also forecasts AI governance platform spending will reach $492 million in 2026, more than double what it was two years ago, and cross $1 billion by 2030 as AI regulation extends to 75% of the world's economies.

"A wide swathe of providers and fragmented point use case implementations compounds the issue," Makhija says. In some advanced deployments, he adds, AI costs are approaching, and in select cases exceeding the salaries of the teams doing the same work. This is on course to become a significant line item on enterprise profits and losses. 

The hidden layers nobody priced in

Model and token costs are visible, but it is only on the surface. Srikara Rao, CTO of R Systems, describes what he calls "silent cost blowups”. These are expenses that accumulate well below the line of executive visibility. When an AI agent runs autonomously, it continuously interacts with systems, APIs, databases, and other agents. It retries failed tasks, maintains context, invokes tools, and triggers downstream actions. "While users see a single outcome, hundreds of underlying actions may have been executed to produce it," he says. "By the time organisations notice the impact, the economics may no longer justify the deployment."

In one enterprise engagement, R Systems found AI costs distributed invisibly across orchestration, monitoring, and multiple disconnected systems, with no unified view of actual spend. Consolidating onto a single platform with built-in cost analytics improved cost visibility by 30–35% and reduced deployment effort by 40%.

Arun Shetty, CTO and Senior Director, Solutions Engineering, Cisco India and South Asia, calls this the hidden infrastructure layer. Cisco’s AI Readiness Index shows that only 32% of organisations in India are integrating AI into their security frameworks, and only 45% are fully equipped to control and secure AI agents. The costs of properly instrumenting agent interactions, enforcing identity controls for non-human actors, and building real-time anomaly detection are high and largely unforeseen. "Building observability infrastructure from the start is more cost-effective than retrofitting it later," he says.

On top of this, IDC's Worldwide AI Spending Guide projects global enterprise AI spending at $407 billion in 2026, up 34.8% from 2025. Gartner separately forecasts that AI governance platform spending alone will reach $492 million this year. These are not incremental additions to IT budgets. They are structural cost lines that most finance teams are still learning to model.

From tokenmaxxing to efficiency-maxxing

The framing that has gained traction in enterprise technology conversations this year is the shift from "tokenmaxxing”, which means consuming as much AI as possible and measuring success by adoption, to something harder and more consequential.

"Token consumption is becoming a poor proxy for value creation," says Sambhav Jain, Managing Director and Partner at BCG. "In the early AI era, more tokens often meant better outputs. Increasingly, enterprises care about outcomes per dollar spent, not tokens consumed." Jain describes the direction of travel as "efficiency-maxxing”, where advances in models, retrieval, caching, fine-tuning, and agent orchestration are reducing the tokens required to achieve the same outcome.

Critically, he argues that excessive token consumption often signals something else entirely. "Many agent architectures compensate for poor tooling, memory, or process design by throwing more context at the model. Mature deployments focus on better system design before increasing context windows or agent loops."

Therefore, this bill's arrival is not just the price of success, but rather a price of architectural shortcuts. The question of when to use an agent, as opposed to simpler automation, is one that most organisations have not yet answered well.

K. A. Prabhakaran, Senior Vice President and CTO at Cyient, uses a principle his team calls matching technology to the characteristics of the task. "Highly deterministic workflows may benefit more from conventional automation, while workflows requiring contextual reasoning across multiple systems may justify the additional autonomy and cost associated with agents," he says. "Agentic AI is a powerful capability, but it is not the destination. The destination is business value."

Rao puts it more directly. "The most expensive AI decision an enterprise can make is deploying an agent on a workflow that doesn't need one. Scripted automation is deterministic, auditable, and a fraction of the cost."

The economics are being rebuilt from scratch

The budget question is also reshaping where money flows inside enterprises. Jain breaks it down to a typical AI-plus-human deployment. In a customer experience process, roughly 60% of spend currently goes to systems integrators and BPO players that provide human-in-the-loop oversight, 15–20% to software providers, 12–15% to advisory and implementation services, and only 4–5% to hyperscalers or foundation model providers. A few years ago, the human element represented 80% of the cost. That shift is continuing.

Pricing models are unsettled, too. Outcome-based pricing, where vendors take a share of the value generated rather than charging per token, is attracting early interest. However, it remains at the edges of most enterprise conversations. "Over time, revenue sharing and outcome-based pricing are the ones that are expected to grow in popularity," says Jain. "We are seeing a lot of traction in early conversations on this one."

For companies like Nazara Technologies, the AI economics question plays out differently but arrives at the same conclusion: the value is in the revenue line, not the cost line. The company's Spanish acquisition, Brutal Games, cut development cycles from six months to two using AI-native workflows. CEO Nitish Mittersain describes the Kiddopia model, with 250,000 paying subscribers in the US at $10 a month, and AI is allowing the same team to deliver four times the content frequency. "When people think of AI-driven margins, usually what they think is: reduce headcount," he says. The better play, in his view, is to use the same team to generate more output and more subscriber retention. The margin impact follows from the revenue expansion, not the cost reduction.

The unit economics that matter are not token costs but cost per outcome, per workflow, per decision delivered. "The biggest mistake enterprises make today is treating AI cost management as a technology problem," says Makhija. "AI agents are becoming a new factor of production. Enterprises will eventually manage them much as they manage human labour today: through budgets, performance metrics, accountability, and governance."

McKinsey's most recent global AI survey puts a sharp number on where the industry sits—94% of respondents report not seeing "significant" value from their AI investments, even as nearly nine in ten companies have deployed AI in at least one business function. The scaling gap is real, and it is widening.

The next phase of enterprise AI will be defined not by who deploys the most, but by who can demonstrate what every token is actually worth.