The AI question top-of-mind for business leaders has recently changed. While previously they were focused on some version of ‘where should we be using AI?’, now business leaders want to know ‘why are our AI pilots not paying off?’ The people asking are rarely AI sceptics. They have a budget, board backing and teams who genuinely want to use the technology. What they lack is a reason to believe the next twelve months will look different to the last.
Look beneath the surface of those organisations and the same story keeps surfacing: somebody bought a licence for an AI tool and sent an email encouraging everyone to have a go. Then, the company waited for the AI to deliver and for the value to appear.
That approach did have a logic to it in 2023, when nobody knew what these tools could really do and the fastest route to understanding was to let people experiment. However, this tentative mindset from three years ago still prevails, and it is one of the main reasons AI programmes stall.
The real cost of AI experimentation
The analogy of a blank sheet of paper is a useful one, because that is what many businesses have unwittingly given their employees when it comes to AI. The worry is that the blank sheet isn’t a strategy. While experimentation without guidelines will generate plenty of activity, it creates very little value in the long run.
What is most concerning is that the money companies are now writing off on AI projects is becoming harder to ignore and, in some instances, is impacting their bottom line. In the US, MIT’s NANDA initiative reviewed more than 300 enterprise AI initiatives and reported that 95% of generative AI pilots delivered no measurable return, set against 30 to 40 billion dollars of enterprise investment.
What stands out about it, is that technology is rarely the only problem. The recurring pattern is a lack of measurable goals. At this point, the business hasn’t worked out a route from pilot to production and, to make matters worse, with no clear business owner accountable for whether it works, the whole exercise becomes a gamble.
As businesses further invest in AI projects that don’t deliver tangible benefits, they will be more wary of investing in the technology in the future. It is a theme that has been picked up by Gartner, which, in its latest research on the topic, predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027. The research firm cites escalating costs, unclear business value and inadequate risk controls as being the key reasons for projects stalling. Gartner concludes that the gap between pilot and production is structural. In other words, this keeps happening because of how these projects are set up, and running more of them the same way produces more of the same result.
The cost to humans
The blank sheet approach also has a human cost, receiving far less attention than the financial one. Handing someone a powerful, unfamiliar tool with no brief attached feels like giving them a test they haven’t had a chance to prepare for. A good example came from Henley Business School which surveyed 2,900 UK workers and found 61% feel overwhelmed by the pace of AI change at work.
Giving employees tools they don’t know how to use, with no clear brief, will affect both their morale and how they see the technology. It can turn enthusiastic participants into cynical ones, undermine their confidence in the technology, and, in some instances, their own abilities.
Creative and marketing functions are often first in line for AI experiments, on the reasonable assumption that imaginative people will find imaginative uses. Creativity and technical confidence are different attributes though, and a brilliant copywriter handed an agent-building platform and a fortnight to explore will usually produce a polite summary of why they ran out of time.
Focusing on problems
So what does the alternative look like in practice? Put simply, it starts with a problem rather than a product.
Businesses need to choose a problem, something that genuinely costs them money or time. This could be, for example, a claims process that takes nine days, or a finance close that eats the first week of every month, or even a support queue where 40% of tickets are the same four questions.
Once the issue is identified, businesses need their teams to own that specific problem, providing the tools they need to solve it. But before anyone starts building, the company needs to define what success looks like, and map out a trajectory to reach it, including milestones and deadlines.
That single reframe makes AI dramatically more approachable for the people who were previously anxious about it, because the ask is now concrete. It gives finance teams a way to judge what happened and to assess the project’s value for money. It also solves the governance question, since a project with a defined problem, a named tool and a measurable target is a project somebody can actually oversee.
Rethinking AI strategies
So how should businesses proceed as they rethink their AI strategies. There are three core adjustments that they need to make.
The first is to introduce role-based access to AI in place of all-you-can-eat AI. Not every role will benefit from using AI in the same way or with the same frequency. Workato runs AI across the whole team because we build software for a living. A logistics firm or a manufacturer should reach a different and equally valid conclusion. Matching access to genuine need controls cost and reduces the noise people have to wade through.
The second is bringing domain engineers and architects in early to shape the initial discussion around planning AI. When the people who understand the surrounding systems shape a build from day one, what emerges is something that connects to the rest of the business and has a chance of reaching production.
The third is unified visibility. Many organisations are unable to produce a single reliable list of every AI agent, application and model running inside their business. This makes measuring value close to impossible and governance largely aspirational. One observability layer across all of it turns AI from a collection of local experiments into something leadership can steer.
Most technology shifts feel chaotic before they feel obvious. Deciding what not to use AI for is a key part of leading a strategy. The businesses finding AI success are not the ones with the longest list of pilots. They are the ones who can answer two questions about every piece of AI work they have running: does this connect to something that matters, and can we prove it?
Hand a team a blank sheet and you will get activity. Give them a problem and you will get a result.
