Escaping The Tokenmaxxing Trap With Outcome-First AI Governance

As agentic systems mature, leadership teams are taking a hard look at their AI bills. With projections suggesting AI costs could be four times higher this year, executives are tightening budget discipline heading into the second half of 2026. But in the rush to justify the spend, many organizations are measuring the wrong things. Tracking token usage and prompt counts has sparked concerns over "tokenmaxxing" and artificial adoption, a familiar dynamic where usage becomes a target and stops reflecting actual value.

Anthony Moisant is the dual-hatted Chief Information Officer and Chief Security Officer at Indeed. Alongside his role as SBU Executive Officer at Recruit Holdings, he oversees technology and digital trust for a platform with more than 655 million job seeker profiles and 3.5 million employers. A U.S. Navy veteran who served on nuclear fast-attack submarines, Moisant brings a highly practical lens to technology investment. At Indeed, that lens has produced an operating model where every AI dollar is tied to a workflow the business can measure.

"Once a metric becomes tied to incentives, behavior organizes around it, often in ways that aren't intentional but are still counterproductive," said Moisant. "This is essentially Goodhart's Law: when a measure becomes a target, it stops being a good measure." He drew a parallel to Campbell's Law, which makes a related point: the more weight an organization places on a quantitative indicator, the more likely that indicator is to distort the very behavior it was designed to measure. "We are careful not to confuse AI activity with AI impact. Token usage and prompt counts are useful signals for us to inspect, but they are not the goals."

Getting enterprise AI working across legacy systems is complex, and usage metrics have value as early diagnostic tools. They provide what Moisant called "air cover," the psychological safety for teams to experiment without immediate ROI pressure. The friction starts when those temporary launch proxies become permanent targets. Indeed started where many companies do—briefly tracking the percentage of code written with AI—before realizing the metric wasn't reshaping outcomes. The team dropped it. Fail fast, learn faster.

• The Goodhart glitch: "Any time a metric like usage or token counts becomes part of an incentive system, you create a perverse incentive," Moisant noted. "People start chasing what's easy to measure rather than what's meaningful to improve. They stop optimizing for the mission and start optimizing for the metric." As AI automates execution, activity metrics tend to measure the least strategic parts of a workflow. Human contribution is moving up the stack toward judgment, prioritization, and integration that matters to customers and the business. Usage alone doesn't tell you whether work is getting better, faster, or more impactful.

• Outcome, not activity: To bypass those misaligned incentives, Moisant anchors investment on time-to-value: how quickly Indeed moves from idea to product delivered to customers. Instrumenting it starts with mapping the full lifecycle from product ideation through validation to find where friction actually lives. "The biggest blockers to time-to-value aren't solvable by AI alone. So the first thing instrumentation does is make those necessary process changes visible." From there, each AI investment is attached to a specific friction point in that lifecycle and measured against a single question: did it meaningfully change the outcome? "The goal isn't to track AI activity. It's to measure whether AI is meaningfully accelerating outcomes," he said. "That shift is what turns time-to-value from a concept to a real operating metric."

That methodology produces different metrics than token counting. Outcome measurement revealed that AI coding assistants save Indeed engineers an average of four hours per week, contributing to a 50% boost in overall engineering productivity. The strategy extends beyond developer velocity. By automating administrative tasks like meeting prep, account research, and planning, sales teams reclaimed hours to focus on strategic customer work, tightening feedback loops and shortening hiring cycles. On the customer-facing side, Indeed's AI chat agents now resolve 28% of cases autonomously, allowing employers to spend less time waiting for support and more time getting jobs posted. In each case, the governing metric is not how much AI the team consumed but whether employers and job seekers got to outcomes faster.

As the enterprise transitions to more autonomous workflows, static annual budgets can feel misaligned with the pace of agentic systems. Scaling without accountability allows costs to grow faster than value. Moisant addresses this with an adaptable model built on outcomes, ownership, and visibility, embedding governance directly into the architecture to evaluate usage continuously rather than in quarterly reviews. But rewiring the architecture is only half the work. Leaders also have to change how they manage.

• Context over counting: At Indeed, governance means identifying the highest-consumption users of coding agents and reaching out with the explicit goal of guiding their workflows. "We still look at usage patterns to understand what's happening beneath the surface," Moisant said. "When a team's token usage spikes, the question isn't why they are using more. It's what they are building and whether it is working." He argued that the leadership role has shifted accordingly. "Their job isn't to maximize usage. It's to decide what should be automated, what should be augmented, and what should remain human-driven." Indeed does not cap AI investment. "If we over-index on budget discipline, we risk slowing the productivity gains, product velocity, and customer impact we're trying to unlock," Moisant said. The balance comes from attaching every dollar to a real workflow and maintaining real-time visibility into whether spend is translating into results.

• Innovation and trust: Moisant's dual mandate as CIO and CSO tasks him with balancing speed and security across a platform where AI deployments affect how job seekers and employers connect. Maintaining marketplace integrity goes beyond standard IT policy. "Safe speed means we never trade trust for velocity." That requires clear ownership, strong instrumentation, and guardrails built into the architecture so that quality, risk, and customer impact are inspected as part of the workflow rather than after the fact. "If teams know the guardrails, understand the risks, and have visibility into how systems are performing, they can make decisions with more confidence," Moisant said. "Innovation and trust aren't opposing forces. Done well, trust is what allows organizations to move faster safely and sustainably."

Durable advantage comes from treating AI spending as an input to be governed alongside output, not a line item to minimize or a leaderboard to climb. For Indeed, every investment connects to the same question: is it helping people get jobs faster? If the answer is yes, the green light stays on.

"If AI isn't moving a real metric, it's not productivity," Moisant concluded. "It's just expensive electricity."