Workers Feel More Productive With AI. The P&L Hasn't Noticed.

Ask employees whether AI has made them more effective at their jobs, and the answer is overwhelmingly yes. Ask their employers whether that effectiveness has shown up in earnings, and the answer is overwhelmingly no. That is the uncomfortable split running through this year's crop of enterprise AI research: McKinsey's 2026 global survey found that roughly 80% of respondents report AI has improved their individual productivity, yet only about 37% say it has made a positive contribution to their organization's EBIT — a figure that is essentially unchanged from a year earlier. The share of companies McKinsey classifies as true 'AI high performers,' meaning they attribute at least 5% of EBIT to AI, has been stuck at around 6% for two years running.

That is not a one-off finding from one consulting firm with a product to sell. Deloitte's parallel State of AI in the Enterprise research, fielded separately across more than 3,200 leaders in 24 countries, describes the same split from a different angle: only about a third of organizations say they are using AI to meaningfully redesign core processes or transform products, while a comparable share admit they are still using it at a 'surface level' — bolting it onto existing workflows rather than rebuilding around it. Put the two studies together and a pattern emerges that should worry any executive who signed off on an AI budget expecting the investment to show up in guidance: individual-level productivity is real and widespread, but it is evaporating somewhere between the employee's desktop and the finance department's spreadsheet.

Where the money actually shows up — and where it doesn't

McKinsey's data suggests the gap isn't random; it tracks function and company size. Large enterprises — those over $1 billion in revenue — are far more likely to be scaling AI agents in at least one function (40%, up from 27% a year ago) than smaller firms, which have been stuck near 22% for two years. Among companies that do capture value, the pattern is also uneven by function: cost reductions concentrate in supply chain, service operations and manufacturing, while revenue gains cluster in marketing, product development and software engineering. In other words, AI is proving itself as a cost lever before it proves itself as a growth lever — which is exactly the kind of benefit that's easy to feel at the team level and hard to see at the consolidated-earnings level, since it shows up as marginally lower costs scattered across dozens of departments rather than one clean, attributable number a CFO can point to.

There's also a quieter story in the build-versus-buy data: roughly a third of respondents say their organization has skipped buying at least one software product or feature because an internal team could now build it with agentic coding tools — most commonly in technology, followed by healthcare and professional services. That's a real efficiency gain, but it's also invisible to traditional ROI accounting, since avoided purchases don't appear anywhere in a revenue or margin line. Some meaningful share of the 'missing' AI value may simply be showing up as costs that never got spent, rather than costs that got cut.

The data problem nobody budgeted for

If productivity gains are real but financial impact is stuck, the question is what's actually standing in the way — and a separate line of research points less at the AI models and more at the plumbing underneath them. Dun & Bradstreet's quarterly AI Momentum survey of 10,000 businesses across 32 countries found that while 97% of organizations report active AI initiatives, only about 5-6% describe their underlying business data as fully ready to support AI at scale. Half cite limited data access as a leading obstacle, and roughly four in ten point to data quality and integration problems severe enough to undercut whatever a model is asked to do with that data.

The challenge now is that AI adoption has outpaced data readiness.

That line, from Dun & Bradstreet chief data and analytics officer Gary Kotovets, captures something the productivity-versus-EBIT numbers only hint at: companies have largely solved the 'can employees use this tool' problem and are now running into the much older 'is our data trustworthy enough to automate decisions on top of' problem — a problem that predates generative AI by decades and that no amount of model capability fixes on its own.

Deployment is outrunning readiness on the people side too

A third dataset suggests the gap isn't only technical. Kyndryl's second annual People Readiness Report, surveying 1,100 senior business and technology leaders across eight countries, found that 57% of enterprises now have AI embedded in core processes or deployed broadly — up sharply from 35% the year before. But only 32% have hit at least one of their top two AI objectives, and just 11% have hit both. Confidence is actually sliding even as deployment accelerates: only 23% of leaders now say their workforce is fully prepared for AI, a six-point drop from the prior year, even as nearly four in five agree that AI's development is outpacing their organization's operating model. A small cohort Kyndryl calls 'pacesetters' — about 9% of respondents — are 1.5 times more likely to report AI-driven revenue growth, and the trait that separates them isn't a better model or a bigger budget; it's that they redesigned roles and decision rights before they scaled deployment, rather than after.

The takeaway for decision-makers

Lay the three datasets side by side and the implication for business leaders sharpens considerably. The problem enterprise AI has in 2026 is not primarily a capability problem — employees across functions and industries are telling researchers, consistently, that the tools make them individually faster. It's an organizational-design and infrastructure problem: unreliable data pipelines that undercut automated decisions, workflows that were never rebuilt around the new tool, and roles and incentives that still assume a pre-AI division of labor. McKinsey's own data on 'high performers' backs this up directly — nearly three-quarters of that small group have fundamentally redesigned workflows around AI, versus roughly a quarter of everyone else, and they're more than twice as likely to be putting over 15% of their IT budget behind it. None of that requires a better model. It requires treating the rollout as an operating-model project with a technology component, rather than a technology project with an operating-model footnote — which is a far less comfortable thing for a board to hear than 'we need to buy more compute.'