Everyone is using AI. Almost nobody is banking it.

McKinsey's 2025 Global Survey on the State of AI puts regular AI use, in at least one business function, at 88% of organisations, up from 78% a year earlier. Generative AI specifically sits at 72%. On those numbers the adoption argument is over. Every competitor you have is using this technology.

Then the same survey asks what it is worth, and the picture inverts. Only 39% of respondents can point to any measurable effect on the bottom line, and most of those put the figure below 5%. Just 7% describe AI as fully scaled across the organisation. Roughly 6% qualify as high performers, attributing more than 5% of EBIT to it.

Read those three numbers as a funnel and the problem states itself. Getting to 88% required buying a licence. Getting to 39% requires being able to measure. Getting to 7% requires changing how the work is actually done. Most organisations stopped after the licence.

Why pilots stall

In marketing the failure is rarely the model. It is that the pilot was scoped around a tool rather than around a workflow, so the output has nowhere to go. A team generates two hundred ad variants in an afternoon and then discovers the bottleneck was never variant production; it was legal review, trafficking, and the fact that the measurement setup cannot resolve two hundred creatives anyway.

If the constraint downstream is unchanged, speeding up the step before it produces a bigger queue, not a better outcome.

The second failure is the counterfactual. Teams report time saved because time saved is easy to self-report. Almost nobody runs the holdout that would say whether the AI-assisted work performed better than what it replaced. Without that, you have an anecdote with a number attached.

What the 6% appear to do differently

• They redesign the workflow, not the step. The question is what the process looks like when a capable model is assumed, not where a model can be inserted into today's process.
• They instrument before they deploy. The success metric and its counterfactual are agreed before the first output ships, which is the only way a result survives contact with a CFO.
• They concentrate. A small number of use cases taken all the way to production beats a portfolio of pilots, because value only appears after the last mile.
• They keep a human accountable for the output. Not reviewing every asset; owning the standard the outputs are held to.

What we do with this

We treat AI as a production and decision capability, not a content faucet. Concretely: creative systems that generate and version assets against a brief and a brand rulebook; analysis layers that compress reporting from days to minutes and surface the three things that changed rather than the four hundred that did not; and a measurement frame that can actually attribute a lift to the intervention.

The uncomfortable part is that the third item is the one that makes the first two worth paying for, and it is the one nobody wants to fund. The 39% figure is not a technology problem. It is a measurement problem wearing a technology costume.