Two numbers explain most of corporate AI's trouble. In McKinsey's latest research, 89 percent of organizations use AI in at least one business function, up from 88 percent a year earlier. The share of high performers, those attributing at least 5 percent of EBIT to AI and reporting value from its use, sat at 6 percent both years. Writing in a commentary first published in Fortune, McKinsey senior partner David Pralong argues the gap persists because executives "keep mistaking some of AI's clearest successes for a playbook they can apply anywhere."
He is right, and the point runs further than he takes it. The customer-support and coding results that fill every board deck are evidence that AI pays off where someone already did the slow, unglamorous work of structuring the process. Companies that lift the tool without the structure are buying a faster first step for work with no defined finish line, then wondering why earnings refuse to move.
The headline wins were paid for decades ago
The studies executives cite are real. Pralong points to roughly 5,200 support agents who resolved 15 percent more issues per hour with AI assistance, and randomized trials of about 4,900 developers who completed around 26 percent more tasks with a coding assistant. His sharper observation is that both settings had a long head start. Contact centers had spent decades organizing work into queues, tracking outcomes and accumulating past interactions. Software teams had testing, continuous integration and code review, so new work could be inspected and corrected.
So these gains are a story about readiness as much as technology. AI arrived at systems with clear inputs, measured outputs and a reliable way to catch mistakes, and it made them faster. Most of what a company does looks nothing like that. A small-business loan that passes through sales, credit, compliance and operations has no single owner, no agreed definition of done, and four places to sit and wait. In Pralong's own example, AI document review might save the bank two days, and the customer may barely notice.
Before borrowing the contact-center story, an executive should check whether the target work has what made AI succeed there:
- Work arrives in volume and is organized so it can be counted.
- Outcomes are tracked, so a change in quality shows up somewhere someone looks.
- A body of past cases exists to learn from and to judge new output against.
- Output can be inspected and corrected before it does damage, the way code review catches a bad change.
- One named person or team owns the result from start to finish.
Where the answers are mostly no, the AI initiative is a process project with a software line item attached, and it should be budgeted and led as one. Bain's executive survey fits this reading: 40 percent of software development use cases and 32 percent of customer service use cases are scaling, against "more than 20%" in many other domains. The most structured work, software development above all, scales most often.
Pilots that work mostly build momentum toward more pilots
The best argument against this view is that quick wins are how organizations learn. A visible success in support earns budget, builds data infrastructure, trains people and gives a leadership team the confidence to attempt harder things. Demanding wholesale redesign up front asks executives to commit to a new operating model before they have proof that the technology works for them. Bain itself frames its findings as a counternarrative to the idea that AI never leaves the pilot stage, and reports that 80 percent of generative AI use cases met or exceeded expectations.
The same survey supplies the answer: only 23 percent of executives can tie those initiatives to new revenue or lower costs. Use cases that meet expectations without touching the P&L do build momentum, mostly toward more use cases. The capability a support deployment develops, plugging an assistant into an existing queue, is not the capability the loan process needs. That process needs a leadership team to decide who holds approval authority, which exceptions require specialists and which handoffs can disappear. Those are choices about turf and risk. No number of successful pilots makes them easier, and a portfolio of pilots that "worked" gives everyone a respectable reason to keep postponing them.
Chief executives appear to be drawing the wrong lesson. In the EY-Parthenon CEO Outlook Survey of 1,200 CEOs, 50 percent named AI as the single largest contributor to productivity gains, ahead of business process redesign at 46 percent. Leaders are crediting the visible input and discounting the structural one, which is the template error in a single statistic.
Use cases that meet expectations without touching the P&L do build momentum, mostly toward more use cases.
Redesign means deciding who owns the work
Pralong's best evidence is McKinsey research from July: among organizations in the earliest stage of AI adoption, those that had redesigned workflows were 5.3 times as likely to report enterprise-level value, 32 percent versus 6 percent. That figure deserves two caveats. It shows correlation, and nothing in the data establishes whether redesign produced the value or whether companies already good at capturing value were simply the ones that redesigned. It also comes from a firm that sells redesign, which readers should weigh; Pralong himself concedes his profession bears some responsibility for the problem, since consultants often earn their keep optimizing existing processes. Neither caveat rescues the alternative, though. Even on the most generous reading, a 6 percent hit rate for unchanged workflows is a thin basis for layering assistants onto existing handoffs.
Redesign, done properly, is narrower and harder than the word suggests. A leadership team decides what outcome the work exists to produce, gives one leader authority over the whole process, sets how decisions will be checked, and settles in advance what the freed capacity is for. Skip that last step and the gains leak away. BCG finds that in most cases freed capacity "is not strategically redeployed, as the underlying structure of work remains intact," and warns that many companies see their cost bases unchanged or even rising.
Redesign also forces an answer that tool-first programs avoid. Employees ask, in Pralong's phrasing, "Am I training my replacement?" Managers cannot offer much reassurance while leadership has described only the tasks AI might perform, not the work people will do next. Only once leaders decide what the redesigned process looks like can that conversation happen honestly.
The contact center was never a template. It is a receipt for decades of process discipline that made AI's job easy, and executives who hold it up as proof that AI will remake their own business are citing someone else's investment as their forecast. The companies reporting real value are disproportionately the ones that changed the work; the rest of the 89 percent are still hoping the cheap part, the tool, will do the job on its own.
