AI’s Productivity Bottleneck Is Corporate Bureaucracy, Not the Models

Companies are pouring tens of billions into AI and trimming headcount to chase the savings. Two years into the spending spree, the return is landing in a much smaller place than the press releases promised. McKinsey’s State of AI 2025 survey found 88% of organizations now use AI in at least one business function, up from 78% a year earlier. The same survey found that only 39% of respondents saw any positive impact on earnings before interest and taxes, and a sliver, just 6%, qualified as “AI high performers” with material returns of 5% or more of EBIT.

The pattern is consistent across research houses. A 2025 paper by a team at MIT, summarized in a 2026 analysis from the Belfer Center for Science and International Affairs at Harvard Kennedy School, found that 95% of companies that invested in generative AI got no return at all. Adoption is not the bottleneck. The bottleneck is what companies do with the AI after the contract is signed.

The Adoption Ceiling

McKinsey’s November 2025 survey, published as the November 2025 enterprise AI survey and unpacked in a 2026 analysis of generative AI value levers, put the gap in plain terms. 88% of organizations now use AI in at least one business function. 39% see any EBIT impact at all. 6% report a 5% or greater lift in earnings before interest and taxes, the threshold McKinsey uses to define a high performer. The 6% are not the same 6% quarter to quarter; the report found the cohort turns over, with only a small set of companies holding the high-performer title from one year to the next.

Boston Consulting Group, in a September 2025 study reported in the same Belfer analysis, found that 40% of companies are achieving material value from AI, of which only 5% received substantial value. The shape of the gap is what one industry voice, writing in a recent column, called a “productivity paradox.” Individual workers can ship more output per hour, but the organizational structure around them is not absorbing the speed. The result is a cost squeeze that one industry read framed as the cost war that has replaced the AI adoption race.

The numbers are blunt in a different way from a 2026 set of surveys: 50% of generative AI projects are abandoned after proof of concept, the same Gartner analysis that broke the reasons into four failure points, none of them a model problem. Adoption is not the bottleneck. The bottleneck is what companies do with the AI after the contract is signed.

• 88% of organizations use AI in at least one business function (McKinsey, November 2025)
• 39% see any positive EBIT impact from AI (McKinsey, November 2025)
• 6% qualify as “AI high performers” with 5% or greater EBIT impact (McKinsey, November 2025)
• 95% of companies that invested in generative AI saw no return (MIT, 2025, via Belfer Center)
• 50% of generative AI projects are abandoned after proof of concept (Gartner, January 2026)

The 13-Hour Tax on Finance

Finance teams are now the most visible example of the gap. An IDC survey of more than 2,000 senior finance leaders, commissioned by accounting software firm Sage and reported on CFO.com, found finance leaders spend an average of 13 hours per week verifying outputs from AI tools. The number grows at the top of the distribution: 48% of respondents said they spend more than 15 hours a week on verification, and 19% spend more than 30 hours a week, the level at which, the report’s authors wrote, “AI creates more work than it saves.” The lost hours are now defined in the industry as a “verification tax.” The full breakdown ran in a July 2025 finance-leader survey on AI verification hours.

The 13-hour figure is not lost productivity in the abstract. The IDC-Sage report priced it at about $78,000 a year per senior finance professional, assuming all-in compensation of $250,000. That is a quarter of a fully loaded finance leader’s time redirected from forecasting, scenario work, and stakeholder reporting to a manual review of machine output. The same report found that finance organizations remain overwhelmingly manual: 62% said their operations are primarily manual or rules-based, and only 4% said they are running largely autonomous finance. Sage CTO Aaron Harris, in the interview, called the 26% average productivity the report estimates is lost to reverse-engineering AI outputs a tax that “still leaves you with 74% return,” but added that the priority has to be cutting the 26%. The friction is in the chain around the AI, not in the AI itself.

Why Half of Pilots Die in Place

Half of the generative AI projects that make it past a proof of concept are then abandoned, according to a January 2026 analysis of generative AI project failures. The reasons cluster around four failure points, not technology: poor data quality, inadequate risk controls, escalating total cost of ownership, and unclear business value. The first three are operational problems. The fourth is a strategy problem.

The pilot purgatory shows up in the deployment numbers. The November 2025 survey found that 23% of organizations are scaling agentic AI in at least one function, and another 39% are experimenting. The remaining majority is still piloting. In any single function, the survey found that only up to 10% of organizations are truly scaling agents, and usually in one or two functions. A separate 2026 survey of 1,100 senior business and technology leaders by Kyndryl found that 57% of organizations have embedded AI in core processes or deployed it broadly, but only 32% achieved at least one of their two main AI objectives, and just 11% met both.

The pattern shows up in the same sequencing in a recent Kyndryl and AvePoint survey covered on Oton Technology: agents landing in core processes before jobs are redesigned, training is built, or governance catches up. Companies ship the agent, then try to redesign the workflow underneath it. McKinsey’s data describes the same sequence: most organizations are sprinkling AI on top of existing processes instead of rewiring how work gets done.

The winning companies will not be those with the fastest technology, but those with the shortest bureaucracy.

The line comes from a column by the CEO of Robin Pay, who argued that no language model, however capable, can resolve departmental turf wars, sign legal contracts, or alter a corporate culture rooted in redundant approval chains. The same argument shows up in McKinsey’s data on what separates the 6% from everyone else. Speed, in other words, is bought in the operations layer, not the model layer.

The Playbook of the 6%

The cohort McKinsey calls AI high performers, the 6% reporting 5% or greater EBIT impact, behaves differently from the rest on four dimensions. A breakdown of the November 2025 survey data, captured in a November 2025 breakdown of the high-performer cohort, shows that the high performers are roughly three times more likely to say they have fundamentally rebuilt how work flows, three times more likely to scale agents across the business, and three times more likely to say senior leaders sponsor and own AI initiatives. They also spend more on AI: roughly one in three of the high performers invest more than 20% of their digital budget on AI, against about 7% for the rest.

The high performers are not buying more tools. They are re-architecting the workflow around the tools. The November 2025 survey found that point-level cost benefits are easy to come by in software engineering, manufacturing, and IT, while revenue and experience gains show up in marketing and sales, strategy and corporate finance, and product development. Enterprise-wide change at the bottom line, the survey found, “is still rare.” These companies do not have access to better models. They have access to operating models their peers have not built.

The same survey reported that the high performers are also more likely to set growth and innovation, alongside efficiency, as core AI objectives. They report improvements in customer satisfaction, competitive differentiation, profitability, and market share at higher rates than the rest. Companies that treat AI only as a cost lever, the survey concluded, “lock yourself into incrementalism while others build entirely new value chains.” That is the dividing line the headline adoption number is hiding.

The Transparent Layer Under the AI

Every operations leader has seen the friction. A 2024 employee survey by Mesh Payments, captured in a 2024 employee survey on expense management friction, found that 78% of employees think business travel expense report processes are confusing and take too long, and 45% said the underlying travel policies and approval workflows are confusing. The friction shows up everywhere: a $42 dinner awaiting manager approval, a duplicate data entry into a spreadsheet that does not match the one in the ERP, a second policy reminder that lands before the first is answered. None of it is exotic.

This is the layer the AI cannot see. The CEO of Robin Pay described it as the third, transparent layer of corporate operations, the one that receives zero organizational attention because it does not show up in product roadmaps or infrastructure diagrams. A developer who writes code in half the time with AI assistance still waits weeks for a manual code review and security clearance, while a sales rep who drafts a complex quote in minutes still waits a quarter for procurement, legal, and finance to clear the deal. The AI works. The process around it does not keep up.

From Tasks to End-to-End

The fix is not more AI. It is a different shape of integration. Companies in the high-performer cohort are not deploying more tools, the November 2025 survey found. They are redesigning the workflow around the tools, with AI embedded in legal, financial, and operational systems, not bolted on top of them. Harris of Sage, in the interview, said the focus has to move from human-as-doer to human-as-reviewer, with the cost of verification built into the deployment plan from day one. The 13-hour verification tax is the bill for treating AI as a feature rather than as the backbone of the workflow.

Recent surveys of 1,100 senior business and technology leaders by Kyndryl and 750 enterprise leaders by AvePoint, reported in a June 2026 enterprise AI governance study, found that 46.9% of employees rely on AI agents weekly or daily, 81% expect autonomous agents to make decisions with material business impact within the next 12 months, and 88.4% of organizations had an agent-related security incident in the past 12 months. Only 25% of leaders said they completely trust AI systems to operate without human oversight. The gap between trust and expected autonomy is a workflow problem as much as a model problem. Closing it requires a redesign of the entire process, not another model upgrade.

The conclusion the column by the CEO of Robin Pay reached, and the data now confirms, is that in the AI era the winning companies will not be those with the fastest technology, but those with the shortest bureaucracy. The next round of AI investment will not be settled by who buys the biggest model. It will be settled by who rebuilds the expense report process, the procurement chain, and the legal review queue, and gets the verification tax below the productivity gain. Most are still in the first stage. The high performers are already past it.

Frequently Asked Questions

What is the “AI productivity paradox”?

The AI productivity paradox describes the gap between rising individual output per worker and flat or falling enterprise-level returns. McKinsey’s November 2025 survey found 88% of organizations use AI in at least one business function, but only 39% see any EBIT impact, and just 6% report material returns of 5% or more. The disconnect sits in the operations layer around the AI, not the AI itself.

Why are AI pilots failing to scale?

According to a January 2026 analysis, at least 50% of generative AI projects are abandoned after proof of concept. The four leading causes are poor data quality, inadequate risk controls, escalating total cost of ownership, and unclear business value. None is a model limitation. All four are operational and strategic failures inside the deploying organization.

How much time do finance teams spend verifying AI outputs?

An IDC survey of more than 2,000 senior finance leaders, commissioned by Sage, found finance leaders spend an average of 13 hours per week verifying outputs from AI tools. 48% spend more than 15 hours a week on it, and 19% spend more than 30 hours a week. The report calls the lost hours the AI “verification tax.”

What makes the 6% AI high performers different?

McKinsey’s November 2025 survey found AI high performers are roughly three times more likely than the rest to have fundamentally rebuilt how work flows, three times more likely to scale agents across the business, and three times more likely to say senior leaders sponsor and own AI initiatives. About one in three of the high performers invest more than 20% of their digital budget on AI, against about 7% for the rest.

What is the “transparent layer” in corporate operations?

The transparent layer is the operational substrate under the visible work of the business: expense approvals, manual reconciliations, procurement, legal review, and finance reconciliation. It is “transparent” because it does not appear in product roadmaps or infrastructure diagrams, even though it is where the AI’s productivity gains are absorbed or erased.