Token-maxing: the AI cost trap hitting enterprise budgets

Key Takeaways

• Boomi's CEO reports 10x increase in Claude spending year-over-year, calling current token consumption unsustainable
• Token-maxing, where professionals maximize AI usage for personal productivity, is emerging as a major enterprise cost sink
• Business leaders recommend guardrails with flexibility rather than outright restrictions on agent exploration

AI agent costs are spiraling out of control. Boomi CEO Steve Lucas spent 10 times more on Claude in the past year than the year before, and he says the trajectory is unsustainable. The culprit: token-maxing, a phenomenon where professionals burn through computational tokens to squeeze every drop of productivity from agentic AI.

The shift caught enterprise leaders off guard. A year ago, organizations celebrated token leaderboards showing who used AI most heavily. Today, those leaderboards are relics. No one can afford to waste tokens when agents consume orders of magnitude more than simple chatbot queries.

"Whether you work inside or outside a company, it feels like your core hustle is to use AI and you token max the heck out of that technology for your own job," Lucas told ZDNET. The problem isn't individual productivity gains. It's the math. When every employee optimizes their personal AI usage without considering aggregate costs, the enterprise bill explodes.

Why agentic AI multiplies token consumption

Tokens are the fundamental unit AI providers charge for. Each word, each reasoning step, each tool call consumes them. Traditional chatbot interactions might use a few hundred tokens per query. Agentic AI, which chains multiple reasoning steps and tool uses into autonomous workflows, can consume thousands of tokens for a single task.

The difference is structural. When you ask ChatGPT a question, you get a response. When you deploy an agent to research a topic, synthesize findings, and draft a report, that agent might make dozens of API calls, each generating tokens. Multiply that across an organization where everyone is trying to maximize their AI advantage, and costs scale faster than any productivity gains.

Lucas coined the term "tokenomics" to describe the discipline enterprises now need. A year ago, Boomi wasn't tracking token economics at all. Now it's a core business activity. "Last year, I personally spent at Boomi 10 times the amount on Claude that I did the previous year, 10 times; that's not sustainable," he said. "I can't do that every year."

The ROI question executives must answer

The fundamental question, according to Lucas, is whether organizations can operate AI at a return. Agentic AI will transform how businesses function. That's not in dispute. The dispute is whether the transformation pays for itself, or whether enterprises are subsidizing individual productivity gains with organizational losses.

"Most organizations will look to AI as the enterprise engine of the future," Lucas said. "So, what matters now is, 'Can I operate AI at a return?' That is the fundamental question."

The answer isn't obvious. AI inference costs are dropping, but agent usage is rising faster. And unlike traditional software licensing, AI costs scale with usage intensity. A power user might cost 50x more than an occasional user, making budgeting unpredictable.

Guardrails, not lockdowns

The instinct to restrict AI access is understandable but counterproductive. Snowflake CEO Sridhar Ramaswamy acknowledged that token consumption is rising across his internal teams. He's worried about spending. But he's not cutting access.

"Are we worried about how much we are spending on AI inference across our different internal teams? Absolutely," Ramaswamy said at Snowflake's Summit 2026 in San Francisco. "But do I see that spend as a reason not to use AI? Absolutely not."

His logic: agents unlock new business opportunities that justify exploration costs. Snowflake views agentic AI as a way to turn internal processes into products for customers. The upside potential exceeds the downside risk of overspending during the learning phase.

Matt Luizzi, VP of Analytics at Whoop, takes a similar approach. The wearable technology company is investing in agentic exploration specifically to find competitive advantages. "If we want people to push themselves out of their comfort zones, we're going to need to be OK with them taking risks and understanding that you can't break anything," he said.

How leading enterprises manage agent costs

The pattern emerging from these companies combines freedom with visibility. Whoop maintains guardrails and observability systems that let people know when their usage spikes. The goal isn't to prevent experimentation but to make costs transparent.

This approach requires infrastructure. Tools like ClickUp or Notion can track AI project allocations, while workflow automation platforms like Zapier or Make help standardize agent deployments across teams. The combination of project management and automation reduces ad-hoc token consumption by channeling AI usage through monitored pipelines.

Business leaders interviewed emphasized context over constraints. Rather than hard limits on token usage, they give employees guidelines and the information needed to make cost-effective model decisions. A task that doesn't require GPT-4 class reasoning shouldn't use it. An agent workflow that runs once daily doesn't need to run hourly.

The shift from token celebration to token discipline

The cultural shift required is significant. During the generative AI boom, organizations celebrated heavy AI usage as a sign of innovation adoption. Leaderboards tracked who used the most tokens. That metric now looks naive.

In the agentic era, token maxing signals waste, not progress. The metric that matters is value per token: what business outcome did that consumption produce? A marketing agent that drafts campaign copy in minutes might justify heavy token usage. An agent that reformats the same document seventeen different ways before a human chooses option three represents pure cost.

The analogy to cloud computing is instructive. Organizations went through similar cycles with AWS and Azure bills. Early cloud adoption often meant runaway costs as teams spun up resources without accountability. Mature cloud operations involve cost allocation, usage monitoring, and reserved capacity planning. AI inference is following the same trajectory, just faster.

What this means for your AI strategy

If your organization is deploying agentic AI without token monitoring, you're likely overspending. The executives interviewed aren't guessing about their costs. They're measuring them and making explicit decisions about acceptable expense levels.

The practical steps: implement usage dashboards, establish per-project or per-team token budgets, and create guidelines for model selection based on task complexity. Not every task needs the most powerful model. Not every workflow needs to run constantly.

Most importantly, frame AI spending as an investment requiring returns, not an entitlement requiring justification. Lucas's 10x increase in Claude spending isn't inherently bad. It's only bad if it didn't produce commensurate value. The discipline of tokenomics forces that question into the open.

Frequently Asked Questions

What is token-maxing in enterprise AI?

Token-maxing refers to professionals consuming large volumes of AI tokens to maximize personal productivity, often without considering aggregate organizational costs. It becomes problematic when individual optimization creates unsustainable enterprise spending.

Why do AI agents cost more than chatbots?

Agents perform multi-step reasoning and tool calls autonomously. A single agent task might involve dozens of API interactions, each consuming tokens. Traditional chatbot queries use far fewer tokens because they involve a single request-response cycle.

How should enterprises budget for agentic AI?

Implement usage monitoring at the team or project level, establish token budgets, create guidelines for model selection based on task complexity, and measure value produced per token consumed rather than total usage volume.

Source: Latest news