Closing the AI impact gap

AI adoption is no longer difficult to find. Employees are using copilots and generative AI tools in their daily work, business units are running pilots, and new AI capabilities are appearing in existing enterprise platforms.

What is much harder to find is evidence that all this activity is changing business performance at the same rate.

McKinsey’s 2026 State of AI survey found that nearly nine in 10 respondents regularly use AI somewhere in their organisation. Some 80% reported improvements in individual productivity, while only 37% said AI had contributed to earnings before interest and tax. Just 6% qualified as AI high performers, attributing at least 5% of Ebit to AI and reporting significant value from its use.

This is the AI impact gap. The technology has moved quickly, but many organisations are still working out how to turn AI’s rapidly evolving capabilities into results at scale.

The limitations of early generative AI were largely technical. Models lacked enterprise context, struggled with complex work and could not reliably interact with the systems businesses depend on.

That picture has changed. AI can now handle more complex reasoning, work across different types of information and interact with enterprise systems through agents. Organisations can also provide models with far more of the context they need to perform useful work.

But greater capability does not remove the need for people, judgement or oversight. As AI becomes faster and more autonomous, the consequences of a poor decision can increase, particularly when it is embedded in important business processes. Those safeguards can no longer depend on individual users working around limitations themselves; they have to be designed into how the work gets done.

Giving employees access to an AI assistant is relatively easy. Redesigning a claims process, customer journey, financial operation or supply chain around AI is much harder. It requires decisions about how work will be performed, where responsibility sits, what data is needed, and how the resulting system will be secured and governed.

Those gaps often become visible when a successful pilot tries to move into production. A team may suddenly need production data, engineering support, security approval, governance controls or an owner willing to take responsibility for the outcome.

At that point, the question is no longer simply whether the AI works. It is whether the organisation is ready to operate it at scale.

This also explains why measuring AI adoption tells us only part of the story.

A business can have thousands of active AI users and still struggle to point to changes in revenue, cost, customer experience or operational performance. It can run successful proofs of concept without having a reliable way to move them into production.

Previous articles in this series have looked at parts of that problem from different directions. Talent and operating models need to change as AI takes on more work. Security has to support adoption without becoming the reason projects stall.