Fluency Is Not a Subscription

The companies buying the most copilot seats are not the companies whose people are getting smarter. They are often the companies whose people are getting faster at being shallow. That distinction will define the next five years of enterprise productivity, and most of the dashboards measuring it are pointed at the wrong thing.

This is the part of the AI cycle where the honest measurement problem becomes uncomfortable. The vendor decks are full of double-digit productivity gains. The instrumented studies are pulling in the opposite direction. The macro labor numbers say something stranger still — that the cohort losing ground hardest right now is the one with the most AI use, not the least.

The numbers tell two stories

Microsoft's 2026 Work Trend Index, based on what the company calls a privacy-preserving analysis of more than 100,000 chats in Microsoft 365 Copilot and a 20,000-worker survey across ten countries, says 66% of AI users report spending more time on high-value work, and that 80% of "Frontier Professionals" — the top 16% of AI users — claim they are now producing work they could not have produced a year ago. Microsoft also reports that only 13% of workers say their employer rewards reinventing work with AI when initial results fall short, and only 26% say leadership is consistently aligned on AI. The first set of numbers is a self-report from people who use the tool. The second set is a structural admission that the organization around them has not caught up.

Anthropic's Economic Index for March 2026, built from roughly a million conversations across Claude.ai and the developer API, reports that about 49% of jobs already have at least 25% of their tasks performed through Claude. On the consumer side, augmentation has crept back ahead of automation — 52% vs. 45% — while on the enterprise API side, automation dominates at 75%. Both numbers are interesting; neither is evidence of skill growth in the humans nearby. They measure what the model did. They do not measure what the person learned.

These are vendor reports. They should be read like vendor reports. Microsoft is the world's largest seller of AI-assistance subscriptions; Anthropic is one of the two firms with the strongest commercial incentive to demonstrate the indispensability of large models. The methodology is more careful than most vendor work, and the directional signal — that real usage is substantial inside Fortune 500 environments — is solid. The self-reported productivity uplifts are softer than the press release implies.

The junior-developer trap is the canary

The cleanest leading indicator that fluency is not a subscription sits in the labor data for 22-to-25-year-old software developers. The 2026 Stanford AI Index reports their employment has fallen nearly 20% from its 2024 peak, with the curve tracking almost exactly the adoption curve of coding assistants. Research summarised this year shows junior developers accepting above 50% of copilot suggestions while seniors accept far fewer; junior developers complete coding tasks faster with assistance and demonstrate measurably worse skill retention when assistance is removed.

Read that pair together. The cohort with the highest copilot acceptance rate is the same cohort whose roles are evaporating. The mechanism is not a mystery. If the first three years of professional coding consist of accepting suggestions you cannot evaluate critically, you do not become a senior engineer. You become a faster intermediate one — at the moment when faster intermediate is exactly the kind of work that gets eaten next.

Some of the better engineering teams have responded by running "copilot-free Fridays" — a deliberate-practice carve-out designed to surface where the tool is a crutch and where it is a multiplier. The framing matters. It treats fluency as something you build under load, not something you switch on with a license.

What McKinsey actually said

The State of Organizations 2026 finding to take seriously is the gap between optimism and readiness: just over half of leaders expect positive outcomes from current transformation efforts, but 72% describe their organizations as unprepared to execute, and 86% say their organization is not prepared to integrate AI into day-to-day operations. McKinsey's own upskilling guidance is the part most enterprises ignore: 80% of tech-focused organizations say upskilling is the single most effective way to close skills gaps, while only 28% have any plan to fund it over the next two to three years.

That ratio is the entire story. The CFO line in most 2026 budgets reads "AI tooling, up; learning and development, flat or down." The result is an organization where the seat count grows quarterly and the capability to use those seats well does not. That is the upstream of the slop pipeline. Most enterprises confuse a vendor-supplied training video for a learning programme; the two things are not in the same category and never will be.

The honest measurement problem

When someone shows you a 30% productivity improvement from a copilot rollout, ask three things. Was it self-reported, instrumented, or output-graded by a third party? Was the comparison a randomized assignment or a willing-volunteer cohort? Did the measurement run long enough to detect the productivity J-curve Brynjolfsson, Rock, and Syverson have been writing about since the 1990s? In most enterprise rollouts, the answer to all three is uncomfortable.

A recent NBER study referenced in the same MIT Sloan piece found over 80% of surveyed firms reporting zero productivity gains from AI, even as Brynjolfsson points to US productivity growth doubling to 2.7% in 2025. Both can be true at once. The macro number reflects the few firms that have actually done the complementary work — workflow redesign, retraining, software rewiring — while the median firm is still in the trough of the J-curve, paying licensing and seeing flat output.

This is not a reason to ignore the technology. It is a reason to stop pretending the dashboards mean what the vendor says they mean. The right comparison is not pre-rollout to post-rollout self-report. It is grader-blind output assessment on matched tasks, six and twelve months out, against a control group that did not get the tool. Almost no one does this work. The few who do find smaller and more uneven gains than the slide decks promised — and find the gains concentrated in workers who already had strong domain skill before the tool arrived.

What works at scale

Siemens has been running its SiTecSkills Academy since 2022. Across production, service, and sales, the programme has passed roughly 24,000 employees, with roughly 40% female participation, and six-month post-training reviews showing close to a 100% reskilling success rate. The single most interesting thing about Siemens' approach is not the headcount. It is that the curriculum was designed by people who understood that digital, green, and AI skills are not separable categories in industrial work — they are a single composite capability, and you teach them as one. The implicit theory of skill underneath the programme is closer to a graduate apprenticeship than a corporate e-learning portal, and that is why it actually moves capability.

The EU is now attempting the same thing at population scale. The Union of Skills initiative announced in March 2025 has spawned an AI Skills Academy intended to aggregate sectoral training from European Digital Innovation Hubs, AI Factories, EIT communities, and the Interoperable Europe Academy. The earlier 70%-basic-digital-skills-by-2025 target was missed; the AI-specific targets should be read with that in mind. The mechanism the Commission has chosen — federated, sector-specific, employer-anchored — is the right shape, but the execution is still front-loaded with administrative ceremony and back-loaded with measurement. Work inside the HCAIM and PANORAIMA programmes has been targeting exactly this gap inside European universities; the open question is whether the supply of trained instructors can keep up with the demand the policy is about to generate.

A stake

I would bet against any organization that spent 2025 buying enterprise copilot subscriptions without simultaneously committing a hard floor — at least 1.5% of fully-loaded labor cost — to grader-blind capability assessment and deliberate-practice carve-outs for the staff using those tools. The seats are cheap. The capability is not. Most CFOs will discover by Q2 of 2027 that they bought the cheap thing and are now paying for the expensive thing on a delay, in the form of senior hires they could not promote internally because the internal pipeline thinned out.

The companies that quietly bent the curve in the other direction this year are the ones that will be hard to compete with for the next five. The asymmetry will be visible in two places — the speed at which their senior staff can ship novel work without ceremony, and the willingness of their mid-career hires to stay. Both are downstream of having taken capability seriously while everyone else was buying seats.

Fluency is not a subscription. It is a deliberate-practice habit, instrumented and graded, that an organization either invests in directly or imports later at a steep premium.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan, an Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.