Earlier this year, several large tech companies started tracking how many AI tokens—the fragments of text AI vendors meter and bill by—their employees burned through each month. Leadership began ranking its employees, with internal leaderboards like Amazon’s “KirkoRank” and Meta’s “Claudeonomics.” Firms handed out titles to employees who used the most tokens, on the logic that they were achieving the most value with AI. The practice became known as tokenmaxxing, and most companies have since shut the leaderboards down.
The leaderboards were a substitute for real ROI analysis of AI, and most firms still have not found a better method. Spend is easy to monitor but value is not, so organizations measured the one they could. Establishing what that consumption delivered would take a single owner tracking AI’s impact on a business process from start to finish, and in most companies nobody holds that view.
One team might use an AI copilot, while another uses a recommendation engine. AI performs well within these types of isolated tasks, but once teams try to scale its use across the organization, the exceptions and errors add up. While the underlying models are technically sound, the orchestration is often missing across end-to-end processes. Research shows that 44% of employees have manually overridden AI outputs because the underlying process wasn’t set up correctly.
When exceptions become expensive
Most AI spending happens between when teams pilot a project and launch it in production—and exception handling is often the reason why. Pilots are typically built for the happy path, not the edge cases. When an AI-powered process encounters a case it wasn’t designed for, it has no defined path forward, so the work gets routed back to the AI model. Every subsequent attempt gets billed again, and the bill keeps climbing long after the original task should have finished.
A recent study documented nearly 700 cases of AI agents ignoring their instructions, with researchers comparing their behavior to that of a slightly unreliable junior employee. In a separate case, a Replit coding agent wiped out a production database in 11 seconds, while the instructions meant to prevent that sat untouched in a prompt.
Cleaning up after an incident routinely costs many time what the original task would have cost on its own. While prompts provide written guidance for how AI should behave, an agent can choose to disregard that guidance altogether. A better option is to clearly define AI instructions within the business process itself. To do this, most organizations need to rethink business processes for an AI-native world.
Rebuild, don’t bolt on
Most large organizations run hundreds of business processes at once. Nearly all these processes qualify as legacy, because someone built them for a world where AI didn’t exist. While layering AI on top of legacy processes might look like a faster, cheaper option, this leads to AI’s most common token-consumption inefficiencies.
Instead of bolting AI onto an existing process, rebuilding a process around what AI can—and cannot—do produces numbers worth showing a board. These metrics might include improvements to how long a case takes from start to finish, what it costs to resolve, and how often it closes without a human stepping in.
Take corporate lending as an example. A credit application passes from one team to the next before it reaches approval, with every handoff dependent on someone opening the file and re-entering details. Adding AI to the document review may shorten one stage, but the handoffs behind it stay where they were. Re-engineered for an AI-native world, agents assemble and instantly check the documents, flagging anything falling outside policy directly to the credit officer who holds authority for that decision. The process then pauses only where judgment is genuinely required.
Historically, a process redesign like this could take up to a year or more. Fortunately, AI has started taking on much of the redesign work itself, compressing these projects into a matter of weeks. In this model, legacy modernization changes from a multi-year commitment into something leadership can greenlight in the same quarter.
Bottom line
The age of measuring token consumption or relative effort as value is over.
The better metrics come from rethinking processes from scratch and evaluating AI’s real impact on business results. By redesigning processes for an AI-native world, organizations can optimize otherwise wasted token consumption from errors and exceptions—and demonstrate real efficiency to their leadership and their board.
Jakob Freund is CEO and co-founder of Camunda.