Enterprises are racing to get their workforces AI-ready. The instinct is understandable. The execution is becoming a problem.
A new behavior is emerging inside enterprises. They are tying employee performance reviews to AI usage. Internal leaderboards track token consumption. Engineers compete over who burns the most compute. The practice now has a name: “token-maxing”. Meta made AI-driven impact a core criterion in performance evaluations, with bonuses up to 200% for top performers. Jensen Huang envisions Nvidia engineers needing an annual token budget, potentially worth half their base salary. The signal from leadership is clear. Use AI. Use it now. Use it a lot.
So here is the problem: Volume is not value.
AI Slop: The Hidden Tax
Stanford and others studied over 1,000 U.S. workers and found that 40% had received what researchers call "work-slop" in the past month. AI-generated content that looks polished but lacks substance. Emails that sound authoritative but say nothing. Slide decks that are structurally sound but intellectually empty.
AI slop can take hours to resolve. At scale, the lost productivity compounds fast. For a 10,000-person organization, researchers at Stanford estimate over $9 million per year in rework costs alone. And that does not account for the cultural damage: 42% of recipients said they trust the sender less. Half viewed them as less capable and reliable.
The data reveals a cruel irony. The people generating slop think they are being productive. The people receiving it are doing the actual work of cleaning it up. AI did not eliminate the labor. It relocated it.
The Two-Tier Workforce
Across the three theaters I operate in: large enterprise transformation, early-stage ventures, and a global technology think tank, leaders describe the same pattern. Employees who understand how to use AI well are seeing genuine, sometimes dramatic, improvements in throughput and quality. They use AI as a thinking partner, not a replacement for thinking. They know how to prompt, how to validate, how to edit, how to integrate AI output into real workflows.
And then there is everyone else. Tons of engagement, no real economic results.
We Have Seen This Before
Robert Solow won the Nobel Prize in 1987, the same year he observed: "You can see the computer age everywhere but in the productivity statistics." Businesses had invested billions in PCs and mainframes. Productivity growth actually declined. It took nearly a decade of organizational redesign, process reengineering, and workforce training before IT investment translated into measurable economic gains.
A recent Fortune article makes the point that the parallel is clear. AI in 2026 mirrors IT in 1987. Adoption is high. Impact is low. The missing link is not the technology. It is the organizational work required to make technology productive. Computers in the 1980s produced too much information. Agonizingly detailed reports printed on reams of paper. Sound familiar? AI in 2026 produces too much content. The medium changed. The problem did not.
100,000 Seeds vs. 100 Trees
This is the core tension leaders need to confront. Broad AI rollouts are necessary. You cannot identify your power users without giving everyone access. You cannot build institutional fluency without experimentation at scale. You need the 100,000 seeds. But seeds without cultivation produce weeds, not orchards.
The companies getting this right are shifting from adoption metrics to outcome metrics. Not "how many tokens did your team burn" but "what did those tokens produce." Not "are you using AI" but "can you demonstrate what AI changed about your output, your decisions, your speed to insight." Shopify's Lütke articulated this well. He did not just mandate usage. He mandated that teams demonstrate why AI cannot do the job before requesting headcount. That is an outcome frame, not a volume frame. The distinction matters.
A Framework for Enterprise AI Rollouts
- Define acceptable use, not just permitted use. First-generation AI policies focused on what is allowed. The next generation must focus on what is effective. Where does AI augment judgment versus replace it.
- Assign human ownership for AI-assisted work. Every AI-generated deliverable needs a named human accountable for its quality. The moment you remove that accountability, you get slop.
- Measure the outcome, not the activity. Track what changed because of AI. Cycle time reductions. Error rate improvements. Decision velocity. Revenue impact. Not token counts.
- Invest in the 100 trees. Identify your power users. In your high return use cases Study what they do differently. Build training around their workflows, not around generic prompt engineering courses.
A Real Risk
The real risk is not that enterprises adopt AI too slowly. It is that they adopt it too broadly without an intentional strategy. That they optimize for visible metrics (tokens consumed, tools deployed, leaderboard rankings) while ignoring the invisible ones (rework hours, trust erosion, decision quality). The question is not whether AI works - it does - at the task level, the evidence is overwhelming. The question is whether your organization is structured to capture that value or just structured to spend on it.
100,000 seeds sown is a strategy for discovery. 100 trees blooming is a strategy for value. The leaders who figure out how to do both, in sequence, with discipline, will separate from the pack. The rest will have impressive dashboards - and a lot of slop to clean up.
