Why Widespread AI Adoption Has Not Produced Widespread Returns

Eighty-eight percent of organizations report using AI in at least one business function, up from 78 percent a year earlier. The evidence on returns is different. Around 6 percent qualify as McKinsey's high performers and attribute at least 5 percent of EBIT to AI.¹

Both figures are based on executive self-reporting to a consultancy. The exact percentages therefore require caution, but the gap is clear: adoption is widespread while attributable returns are not.

What distinguishes organizations that create value

The most-quoted answer is workflow redesign. Among the high performers, 55 percent report having fundamentally redesigned workflows to incorporate AI, against 20 percent of the rest.¹ That is the largest single gap in the survey, and it is not about model quality or compute budget.

BCG offers similar guidance through its 10-20-70 principle: allocate 10 percent of the effort to algorithms, 20 percent to technology and data and 70 percent to people and processes.² This is a recommended allocation of work, not evidence that results arise in those proportions.

Gartner predicted in July 2024 that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, naming poor data quality, weak risk controls, rising costs and unclear business value.³ None of those is a modelling problem. They are all problems about what happens between the demo and the decision.

Team composition is only part of the explanation

Three capabilities are needed, although they need not sit with three different people: practical AI engineering, product judgement and knowledge of the domain in which the system will operate. Domain knowledge is often the missing capability.

The best-known study here complicates the story rather than confirming it. Dell'Acqua and colleagues ran a pre-registered field experiment with 776 professionals at Procter and Gamble, randomly assigning people to work alone or in pairs, with or without AI.⁴ Ideas in the top 10 percent were about three times more likely to come from teams using AI than from individuals working without it. But the finding that matters here is a different one: individuals with AI matched the performance of two-person teams without it, and AI narrowed the gap between R&D people and commercial people, who otherwise proposed predictably different solutions.

The experiment suggests that AI can supplement some missing cross-functional expertise. It does not show that every specialist must be present in the team from the outset.

Wu, Wang and Evans found across 65 million papers, patents and software projects that smaller teams are disproportionately responsible for disruptive work while larger teams extend existing directions, with coordination cost as the mechanism.⁵ That is a real constraint on how fast you should staff up. It is still a claim about team shape, not about whether the work gets checked.

The Dutch picture: experience, not ambition

In 2024 about 23 percent of Dutch firms with ten or more employees used at least one AI technology, up almost 8 percentage points in a year.⁶ Among firms that considered using AI and then decided against it, lack of experience was the most-cited reason, at 74.6 percent. That figure is about the group that considered and declined, not about all non-users, and privacy was also cited by more than half of firms with 100 or more staff.

The figures primarily indicate a lack of experience. They do not show that strategy and cost are irrelevant, but they do suggest that many organizations still lack a reliable way to develop and introduce AI.

A reviewable process from idea to release

A useful process has explicit decision points. Each stage produces a result that an authorized person can review before the next stage begins.

Forenta for Business applies that structure across eight stages, with an explicit decision at each transition. A stage produces a reviewable result, findings and required follow-up. The transition is a decision to proceed, revise or stop rather than a routine status update.

None of that is a substitute for the three capabilities. It is what makes their absence visible early instead of at the end.

What this analysis does not cover

Forenta for Business is available as a controlled pilot and guided engagement, with pricing on request. It is not a fully self-service offer. The individual entry point is Forge, which is in public beta and analyses text supplied by the user.

The survey numbers also carry the usual weaknesses. McKinsey and BCG figures are self-reported by executives with an interest in the answer, the CBS data covers adoption without distinguishing internal productivity tools from customer-facing products, and none of it establishes causation between redesigned workflows and returns. The direction is consistent across three independent sources. The mechanism is inferred.

And nothing here addresses the design of AI systems themselves, the regulatory context for Dutch healthcare, or what happens when an organisation scales past its first team.

Model choice does not explain the return

The survey evidence does not support model choice as a sufficient explanation for returns. What consistently differs is whether the work between an idea and a release includes clear ownership, workflow redesign, validation and explicit decisions.

That is a choice about process, which makes it one of the few variables here that you actually control.