In most businesses, the first stage of AI adoption has happened much faster than expected. Tools that felt experimental just a couple of years ago are now part of daily working life. We’ve leapt headfirst into the AI landslide, and while the dizzying pace has led to a sense of progress and advancement, it is still far too easy to mistake adoption for value.

Because those first AI tests can happen so quickly, they often feel more conclusive than they really are. An early version might summarise a document or automate part of a process well enough to suggest that there is something worth exploring. Everyone in the demo is impressed, and companies talk about committing to an AI strategy without a realistic idea of what’s to come. Then the problems start: the AI struggles with imperfect data, existing systems, staff habits, customer expectations, compliance protocols; the list goes on.

There is clearly a gap between the expectations businesses have for their early AI initiatives and the value they create. By exploring what’s going on and why it’s happening, we can create a road map for founders and decision-makers to put AI strategies in place that will live up to the hype.

Delivering real value

Take a customer service assistant, which is one of the easier AI use cases to imagine. At first, the business case seems obvious: customers get faster answers, staff spend less time on repetitive queries, and the business becomes more efficient. In practice, we learn that the assistant needs reliable information to draw from, a clear way to hand over to a human being when necessary, and a process for checking whether its answers are useful. If it gives confident answers from poor data, or if staff have to spend more time correcting it than they previously spent answering the query, the business has not gained much.

Inside the business, the same pattern appears in less visible ways. An AI tool might help with document review, reporting, knowledge search, or pulling information together before a meeting. The early version may work well enough to impress people, then slow down when it has to fit into the actual workflow, as people struggle to connect its output to the systems they already use and end up flummoxed by the way the tool changes the way a task is done. Suddenly, the AI tool is struggling to make it past the pilot stage as managers experience the gap between the dream they were sold and the reality they’ve bought into.

The facts and figures 

Our recent study among 500 senior decision-makers in UK businesses shone a light on the sizeable gap between AI adoption and measurable value.

As of the start of 2026, 78% of UK businesses said they were already using AI in some capacity, yet only 31% of those using AI had seen a positive return on investment, while less than half had a clear idea of what success looked like when implementing AI.

Managers are usually best positioned to see where time is being lost or where customers are getting frustrated. Before AI, their job was to find ways to optimise these processes and to anticipate and solve issues before they became distractions. However, most managers are not AI experts, so while they understand the impact pilot AI programmes ultimately have on workflow, they cannot be expected to predict how these issues come about, nor how the AI implementation needs to be altered to minimise disruption.

It is no surprise, then, that our study also found that 40% of AI-using businesses said a lack of internal expertise was the biggest barrier to achieving ROI.

How to make AI projects stick

When a project begins with a broad instruction to “use AI”, the work can spread very quickly without becoming any clearer. One team tests a tool, another runs a pilot, a third starts looking at vendors, and eventually there is a lot of momentum without a shared view of what improvement the business is trying to make.

A better starting point is usually more ordinary: which process is taking too long, which decision is being made with poor information, which element is creating avoidable pressure on staff, and what would genuinely improve if AI were applied well?

At this point, within this framework, experimentation is still valuable. Constant, aimless experimentation will likely lead to the figures from our research not improving any time soon, but endless planning and theorising won’t uncover the issues that test programmes raise. What’s most important is that the pilot programme has to allow the business to learn something new, whether it’s that they need better data collection, increased interdepartmental feedback, or just a different solution than AI.

What we’ve learned 

As AI becomes a more regular part of business investment, boards and leadership teams will understandably want clearer evidence of return. That does not mean every AI project has to justify itself through immediate cost savings; some will improve decision-making, reduce risk, support growth, or make a product more useful to customers. The point is that the business needs to decide what kind of value it is looking for before the project is too far advanced; in fact, in an ideal world, this has to be defined before any AI investment or deployment begins.

Otherwise, frustration builds because nobody can quite say whether the technology is working.

For all the attention around AI tools themselves, the harder work is often around the business that has to use them. Data needs to be ready enough, people need to know how the tool fits into their work, managers need enough support to make good decisions. So long as somebody is measuring whether the project has improved the process it was meant to improve, then the project will always be a success. It may be less exciting than the first demonstration, but it is what turns AI from experimentation into something organisations can depend on.

Ritam Gandhi is the founder and director of Studio Graphene.