Four in five respondents to McKinsey's 2026 state of AI survey say AI has made them personally more productive. Only 37 percent report a positive effect on EBIT. Somewhere between the employee's laptop and the income statement, most of the gain disappears.
We think the leak has a specific location. AI's productivity is created by individuals and absorbed by organizations still built around the constraints AI is removing. Companies that want to keep it have to change their operating model, their talent system and their lines of accountability at the same time, because any one of them changed alone will be undone by the other two.
The occasion is a McKinsey Quarterly article, "AI is changing work. Now it has to change the organization", by Alexis Krivkovich, Brooke Weddle, Holly Price and Vik Sohoni. They argue that agentic AI loosens three constraints that shaped the modern firm (human capacity, expertise and coordination) and set out four shifts: organizing work around outcomes, reshaping talent, building mechanisms that let change compound, and putting clear authority behind the effort. The diagnosis is sound. The risk is that readers treat the four shifts as four separate workstreams, which is precisely how the gap gets preserved.
Freed capacity goes wherever the old structure sends it
An hour saved by an agent is only an hour, and an organization designed for handoffs will spend it on handoffs. The McKinsey authors report that, across their client work, coordinator roles such as project management have grown at roughly 1.5 to 2 times the rate of line roles over the past five years. That growth, which McKinsey reads as a sign of how complex organizations have become, suggests where freed capacity will go by default: more status meetings and review cycles, and more people translating between functions that an agent could connect directly.
EY reaches the same conclusion from outside the McKinsey building. Its CEO Outlook report warns that "without operating model change, gains risk being absorbed by complexity and friction," and the accompanying release finds 23 percent of CEOs struggling to convert AI productivity into measurable financial outcomes, with legacy systems and organizational complexity named among the culprits. The McKinsey piece offers a counterexample: one product development function moved from pods of eight to ten people to hybrid teams of four to six supported by agents. The change came from redrawing the team, something faster individuals could never have done on their own.
The best case against all this is the one most CFOs make instinctively. Pick bounded tasks, automate them, measure the result and bank it, with no reorganization and no consultants redrawing the org chart. At the scale of a single process that works, and it is often the right first move. It stops working at the point where the saving has to show up somewhere other than a pilot dashboard. The same EY release reports that only 16 percent of CEOs have strong confidence in measuring AI ROI with real-time visibility into costs and returns. Targeted automation produces a productivity number; whether that number reaches the P&L depends on decisions about team design, workflow and headcount that targeted automation, by definition, does not make.
A fair reader should also discount McKinsey's headline correlation. Nearly three-quarters of its AI high performers report substantially redesigning workflows, against one-quarter of everyone else. Firms able to redesign workflows were probably better run before any agent arrived, and McKinsey sells organizational redesign for a living. Neither caveat rescues the alternative. If the capacity to redesign is what separates the 6 percent of high performers from the rest, that capacity is the thing worth building.
The talent ladder breaks where nobody is measuring
The authors are sharpest on a cost that will not appear in this year's numbers. Agents absorb research, drafting, analysis and coordination, which is the work through which juniors have always learned. "Today's productivity gain can become tomorrow's judgment gap," they write. The evidence is starting to arrive: CNBC reports on a Stanford study of ADP payroll records finding a 13 percent relative decline since 2022 in employment of 22- to 25-year-olds in the most AI-exposed occupations, such as customer service, accounting and software development, while more experienced workers in those fields held steady or grew.
This is where piecemeal change does the most damage. An operating-model team shrinks a pod from ten people to five and books the saving. Meanwhile the talent function, working from last year's career framework, still expects analysts to learn by doing analysis that no longer exists. Each is right on its own terms, and together they produce a firm with no route to senior judgment five years out. McKinsey's prescription, managers spending less time coordinating and more time creating practice, feedback and progressively consequential decisions, requires changing the operating model and the talent system at once.
Promotion has the same problem. If coordination is what agents absorb, advancement tied to managing larger teams starts rewarding the wrong thing. The authors suggest a senior individual contributor leading agents around a complex outcome may create as much value as a manager with a large team. Companies that agree will have to change pay bands and titles along with job descriptions, and that decision sits nowhere near the AI program office.
An hour saved by an agent is only an hour, and an organization designed for handoffs will spend it on handoffs.
Someone has to own the trade-offs
That leaves accountability, the shift most companies will be tempted to skip. McKinsey argues that reinvention needs a senior leader with real authority to move funding and scarce talent and to settle competing priorities, and it proposes an "agentic mission factory" in which a business owner brings an outcome and a cross-functional team redesigns the work around humans and agents. The factory is a good idea with a familiar failure mode. Its predecessor, the digital factory, used dedicated cross-functional teams to change the business while the rest of the organization kept running it, and arguably worked partly because of that separation. The agentic version only works if the redesign reaches roles, career paths and performance management, which live with HR and line managers. Shipping agents is the easy part; the accountable leader has to be able to force those decisions.
Trust is the part that can be measured. McKinsey's readiness research finds 70 percent of respondents personally ready for AI, but only 27 percent of leaders saying their organizations are ready, and it associates employees' trust that the organization will support them with enterprise value at every stage of AI maturity. That trust is built or lost in exactly the decisions that piecemeal programs leave to whoever happens to own them: who loses a role, who gets a realistic path into new work, whether promises about human oversight survive a difficult quarter. A company that automates in one function and redesigns careers in another sends mixed signals, and employees read them accurately.
What lies between 80 percent and 37 percent is the organization. Targeted automation will keep producing the first number, because it measures what individuals do. Only a company willing to redraw its structure, its career ladders and its lines of authority in one motion gets to keep the second. The rest will go on reporting productivity gains that their own org charts have already spent.
