Every founder pitch deck now has a slide about AI-driven efficiency: leaner teams, faster output, lower overhead. It's tempting to treat that slide as settled fact. But the 2026 data on AI productivity tells a more complicated story than either the AI optimists or the AI skeptics are selling, and the part of that story getting the least attention is what efficiency may be quietly costing the humans doing the work.

The Promise

Start with what's true. When researchers from Harvard, MIT, and Boston Consulting Group ran a preregistered field experiment with 758 BCG consultants, giving half of them access to GPT-4 on realistic client-style tasks, the results inside AI's "capability zone" were striking: consultants completed 12.2% more tasks, worked 25.1% faster, and produced work rated 40% higher in quality than colleagues working without it, according to the study published in Organization Science.

A separate study of 5,172 customer-support agents found that access to a generative AI assistant increased productivity by about 15% on average, with substantially larger gains among less experienced workers, per the "Generative AI at Work" paper by Brynjolfsson, Li, and Raymond.

These aren't fringe numbers, and they're exactly the case founders make when they budget for AI tools instead of headcount. Under the right conditions, the gains can be real and substantial.

The Gap Between the Lab and the P&L

The trouble starts when that logic gets stretched from a controlled study to an entire company. A National Bureau of Economic Research survey of nearly 6,000 CEOs, CFOs, and senior executives across the US, UK, Germany, and Australia found that while 69% of firms actively use AI, 89% report no measurable impact on labor productivity over the past three years, and more than 90% report no impact on employment.

That survey measures reported effects over a three-year window rather than a controlled measurement of AI's underlying productivity effect, so it shows that executives generally aren't yet observing measurable enterprise-level gains, not that those gains don't exist.

PwC's 2026 AI Performance Study of 1,217 executives across 25 sectors found a related divide from a different angle: nearly three-quarters of AI's measurable economic value is captured by just one-fifth of organizations — a gap PwC describes as widening, not closing.

Employee sentiment tells a matching story. Microsoft's 2026 Work Trend Index, built on Gallup survey data, found 65% of employees at AI-adopting organizations say the tools improved their individual output, yet only about one in ten strongly agree AI has transformed how work actually gets done at their organization. Accenture's research shows a similar mismatch: 86% of C-suite leaders plan to increase AI investment in 2026, but only 32% say they've achieved sustained, enterprise-wide impact, and just 27% of employees say they're comfortable delegating tasks to AI agents at all.

Task-level speed and company-level results are two different measurements, and most organizations are optimizing the wrong one. The founders capturing real value aren't the ones with the most AI tools installed — they're the ones who redesigned the workflow the tool sits inside.

What Efficiency May Be Quietly Costing Employees

This is the part of the AI conversation founders hear least about, because it doesn't show up on a productivity dashboard. A few patterns are showing up consistently across 2026 research — though this evidence is more preliminary than the productivity numbers above, and is worth reading as a set of open concerns rather than settled conclusions.

The confidence trap

Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers about 936 real instances of AI use at work and found that workers who were more confident in the AI's output reported applying less critical-thinking effort to AI-assisted tasks. The opposite was true of self-confidence — workers who trusted their own expertise kept scrutinizing AI output even when it took more effort.

The study measured self-reported effort during AI-assisted work, not a change in workers' underlying critical-thinking ability, and the authors don't claim to have shown the latter. A separate study of 666 participants, published in the peer-reviewed journal Societies, found a statistically significant negative relationship between reported AI-tool use and measured critical-thinking performance, mediated by cognitive offloading, handing a mental task to a tool rather than doing it yourself.

The study is observational, so it can't rule out that people who already relied more on cognitive shortcuts were simply more likely to use AI tools heavily; the finding is suggestive rather than proof that AI use causes critical-thinking skills to deteriorate.

The verification tax

The job doesn't disappear when AI takes over a task, it shifts. Boston Consulting Group researchers, writing in Harvard Business Review, surveyed 1,488 full-time US workers and coined a term for what they found: "AI brain fry," reported mental fatigue from managing and double-checking multiple AI tools beyond a person's cognitive capacity.

About 14% of AI-using workers reported the symptoms directly, foggy thinking, headaches, slower decisions. The researchers distinguish this reported form of mental fatigue from conventional burnout, linking it to the cognitive demands of managing, evaluating, and correcting AI-generated work, it's a newly coined research term, not a diagnosed medical condition.

The self-perception gap

Perhaps the clearest warning sign is that employees can't always tell when AI is actually helping. A randomized controlled trial by METR gave experienced open-source developers real coding tasks, some with AI tools available and some without. Developers using AI took 19% longer to finish, yet afterward, they estimated AI had made them roughly 20% faster.

METR has since flagged this as a snapshot of early-2025 tools rather than a permanent verdict, and later internal data suggests the picture is shifting as tools and workflows mature. The experiment suggests that self-reported productivity can be a poor substitute for objective measurement of AI's effects in a given setting, a gap founders need to measure around, not assume away.

Zoom out and a pattern emerges across these findings: AI is very good at reducing the visible effort of a task, and the evidence so far suggests it's less reliable at reducing the underlying cognitive load, because someone still has to catch what it gets wrong.

What the Top 20% Actually Do Differently

PwC's leading organizations weren't distinguished simply by how much AI they'd deployed. They were more likely to redesign workflows around AI, pursue growth opportunities, and build stronger governance around deployment, which is what let employees rely on outputs without either blind acceptance or constant re-checking. ActivTrak's analysis of workplace activity data found something related at the individual level: employees who spent 7–10% of their working hours in AI tools had the highest measured productivity of any usage group, outperforming both light and heavy users.

Yet only 3% of employees fall into that range, while more than half spend under 1%. The same data shows a workplace-wide pattern sitting alongside the adoption numbers: as productive hours ticked up, focus efficiency — the share of the workday spent in uninterrupted, focused work, fell to a three-year low. ActivTrak describes this as an association between AI use, collaboration and multitasking patterns, and falling focus, not a proven causal effect of AI itself.

The throughline is a founder's job, not an engineering one: know where your team's version of the capability zone actually ends, staff the verification work deliberately rather than assuming it's free, and resist measuring success by AI adoption instead of outcomes.

Where This Leaves Founders

None of this argues against using AI, the gains for founders who get the workflow right are large enough to matter. What the research supports, taken together, is narrower than a simple productivity story: AI can produce real task-level gains, enterprise-level gains are harder to observe and concentrated among a minority of organizations, and there are legitimate open questions about verification burden, cognitive offloading, and focus that conventional productivity metrics don't capture.

Whether those add up to employees quietly losing judgment and skill isn't settled by the evidence yet, but it's worth watching rather than assuming away. If the machine now handles the easier parts of a task, is the muscle for the harder parts still being built anywhere in your organization, or is everyone just getting faster at forgetting how to do it themselves?

‍

read Being Right Isn't a Business Model: What Leopold Aschenbrenner's Hedge Fund Actually Teaches Founders About Thesis, Leverage, and Time Horizon

‍