The AI super-user divide is a workforce intelligence blind spot

Two numbers from this year's enterprise-AI research refuse to sit in the same room. On one side, AI super-users, employees fluent with the tools across their core work, save nearly nine hours a week, four and a half times the two hours saved by laggards, and 87% of leaders judge them at least five times more productive.¹ On the other, only 29% of organisations report significant return on their generative-AI investment, and MIT's GenAI Divide study finds that 95% of enterprise pilots produce no measurable impact on the profit-and-loss account.¹² Either AI works or it does not. The data insists, awkwardly, on both.

Resolve it this way: the gains are real but local. The same MIT research notes that the tools in widest use, ChatGPT and Copilot among them, mainly lift individual productivity rather than enterprise outcomes.² Individual wins do not aggregate on their own. A five-person team with two super-users does not become a five-times team; it becomes two people carrying more than their share and three who are not sure what the tools are for.³ Adoption is uneven, hand-offs break, and the saved hours dissipate somewhere between the individual and the income statement. The productivity is genuine. The mechanism to capture it is missing.

The metric is part of the problem

Before reaching that conclusion, the strongest objection deserves room, because it is correct. MIT's 95% figure measured whether pilots produced rapid P&L impact within six months, and it skewed towards sales and marketing use cases; analysts who reconcile the headline point out that firms reaching production average around 1.7 times their investment, with 72% of enterprises now running at least one AI workload live.⁴ Berkeley's executive-education researchers go further, arguing that the standard ROI metric misframes early AI altogether, because value first appears as individual efficiency that quarterly accounting is poorly built to capture.⁵ On that reading the panic is a measurement artefact, not a verdict.

Grant all of it. The concession sharpens the argument rather than dissolving it. If the return is real but invisible because it lives at the level of individuals and small teams, the organisation still cannot see where it is happening, which roles convert AI access into genuine leverage, or how to reproduce the pattern elsewhere. Misreading the metric and lacking the signal are two halves of one failure. A leader who believes the 95% headline gives up too early; a leader who waves it away still cannot point to the teams quietly pulling ahead.

Leverage is not logins

This is the workforce-intelligence blind spot, and it is widening into something with teeth. The career consequences are already visible: super-users are three times more likely to have received both a promotion and a pay rise in the past year, and 60% of firms say they plan to lay off employees who cannot or will not adopt AI.¹⁶ Decisions of that weight are being taken on a remarkably thin evidence base, because the instruments most organisations hold measure adoption rather than leverage. Counting logins and prompts tells you who has opened the tool. It says nothing about who has turned access into capacity, and a workforce that suspects those counts feed promotion and redundancy will optimise for the appearance of use rather than its substance.

The distinction that matters is between usage and leverage, and only one of them appears in a tool-adoption dashboard. Seeing leverage means watching how work actually changes when AI enters it: whether freed hours become deeper output or simply evaporate, whether a team's collaboration load eases or its rework climbs, whether the saved time lands on the balance sheet or on someone's commute home. Those are behavioural signals, and they stay honest only when the people generating them are not being scored against them, which is exactly why a surveillance-first approach produces the gamed data that makes the blind spot worse.

The temptation, faced with two irreconcilable numbers, is to pick a side and declare AI either a breakthrough or a disappointment. Both readings let leaders off the hook. The harder truth is that AI is working unevenly, in places they cannot currently see, and the real question is not whether to believe the nine hours or the 95%. It is whether they can tell, inside their own organisation, which teams are converting the promise into capacity, and which are simply busy opening the tool.

Footnotes

1. Writer & Workplace Intelligence. (2026). AI adoption in the enterprise: Why 79% face challenges despite high investment. Writer. https://writer.com/blog/enterprise-ai-adoption-2026 ↩ ↩² ↩³

2. MIT Project NANDA. (2025). The GenAI divide: State of AI in business 2025. MLQ.ai. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf ↩ ↩²

3. LaunchReady.ai. (2026). The AI super-user divide: What the 2026 data reveals about your team. LaunchReady.ai. https://launchready.ai/insights/ai-workforce/ai-super-user-divide-2026 ↩

4. Sykes, B. (2026). The state of AI adoption in the enterprise [Q1 2026 review]. Substack. https://bsykes.substack.com/p/the-state-of-ai-adoption-in-the-enterprise ↩

5. UC Berkeley Executive Education (SCET AI Commons). (2025). Beyond ROI: Are we using the wrong metric in measuring AI success? UC Berkeley. https://exec-ed.berkeley.edu/2025/09/beyond-roi-are-we-using-the-wrong-metric-in-measuring-ai-success ↩

6. Writer. (2026, April 7). WRITER survey finds 60% of companies plan to lay off employees who won't adopt AI [Press release]. Business Wire. https://www.businesswire.com/news/home/20260407140918/en/WRITER-Survey-Finds-60-of-Companies-Plan-to-Lay-Off-Employees-Who-Wont-Adopt-AI ↩