AI success is not about adoption. It is about business outcomes.

The boardroom conversation about AI has changed. A year ago most discussions centred on use cases, copilots and proofs of concept. The question executives ask now is a harder one. What did AI change that we can point to on a set of numbers we already report?

The published research is unusually clear about where most organisations sit. McKinsey's 2026 State of AI survey found that nearly nine in ten respondents use AI regularly in at least one business function, and that 44 percent now describe AI as scaling across the enterprise, up from 38 percent a year earlier. Eighty percent report individual productivity improvements. Yet the share attributing any EBIT impact to AI sat at 37 percent, essentially unchanged on the previous year, and only about 6 percent both attributed 5 percent or more of EBIT to it and described the impact as significant. Adoption moved. Reported financial impact did not.

Gartner points at the same gap from the other direction, predicting that more than 40 percent of agentic AI projects will be cancelled before the end of 2027, and naming escalating costs, unclear business value and inadequate risk controls as the reasons. Its separate work on data readiness makes a related point: organisations that have not prepared their data are the ones whose AI projects stall.

What we have found doing this ourselves

We have applied AI across our own operations: consulting, software engineering, knowledge management and delivery. One lesson keeps surfacing. Successful AI programmes are not built around technology. They are built around knowledge somebody trusts enough to act on.

Many organisations begin with models and platforms. The ones creating durable value begin somewhere else.

Name the decision you want to be better

The useful starting point is not a place where AI could be deployed. It is a business problem or a decision that somebody already complains about. Connect it to something the organisation already measures: revenue, cost, risk, customer experience, or the time it takes to decide.

When the outcome is named that precisely, prioritising the investment becomes an ordinary commercial exercise. When it is not, every AI proposal looks equally attractive, which is how budgets get spread thin across things nobody can later evaluate.

It is also the first thing we write down with a client. Our engagements are now set up so that before anything is built, the client signs the objectives the work is meant to move, stated against numbers they already report. After go-live, a value review is designed to measure against those same objectives, so that the conversation at the end of an engagement can be about which decision changed. The model is described on AI-assisted engagement and delivery, and the advisory work that comes before a build under enterprise AI and Copilot.

What the model reads decides what it is worth

Knowledge fragmentation is the most underestimated obstacle in enterprise AI. The expertise that matters sits in documents, mailboxes, file shares, presentations and, above all, in what people have learned and never wrote down.

AI produces its best work when it can draw on that rather than on general information from the open internet. Capturing it, structuring it and keeping it current is what makes it reusable, and it is slow, unglamorous work that has to happen before the interesting part.

The unit of change is the decision cycle

Most AI initiatives automate individual activities. The larger opportunity is in the whole cycle: gathering information, forming a view, supporting the decision, coordinating the action, then measuring what happened.

Producing a report faster is a small win. Changing how the decision behind the report gets made is a different order of return, and it is also much harder, which is why most programmes stop at the report.

Governance is what makes a recommendation usable

As autonomy increases, governance stops being a technology consideration and becomes a business requirement. People act on a recommendation when they can see where the information came from, how the recommendation was produced, what controls applied, and who remains accountable.

Designing that from the start costs far less than retrofitting it after deployment. Without it, a good recommendation gets read and quietly set aside, which looks in the numbers exactly like a bad one.

What comes next

As AI moves from copilots towards agents that coordinate more complex work, the conversation will shift further from technology towards organisational readiness. McKinsey's data shows large enterprises scaling agents faster than smaller ones, while Gartner keeps returning to governance, data readiness and demonstrable value.

The question on the table is no longer how much AI we are using. It is whether a decision got taken differently because of it, and whether we can name which one.

The next piece is about what sits underneath that, which in our own work turned out to be the questions we were asking.