Insights / Perspective / AI Rapid Prototyping and MVP Development
PerspectiveWhy most AI projects stall before they pay back
Many organisations use AI, but few have scaled it. Why projects stall between pilot and payback, and how to start with one measurable, governed use case.
AI has moved from novelty to normal in a remarkably short time. Almost every leadership team can point to a pilot, a proof of concept or a team experimenting with a new tool. Far fewer can point to an AI system that runs every day in a core process and delivers a benefit that shows up in the accounts.
The figures confirm the gap. McKinsey’s 2025 global survey found that 88% of organisations use AI in at least one business function, but about two-thirds have not yet begun to scale it across the enterprise, and only 7% report full scale. Gartner has predicted that through 2026 organisations will abandon 60% of AI projects that are not supported by AI-ready data.
Neither figure suggests that AI does not work. They suggest that the path from a promising pilot to a dependable system is harder than it looks, and that most organisations are still learning how to walk it.
Why projects stall
When we review stalled AI initiatives, four causes appear more often than any others. They are rarely technical in the narrow sense. They are about value, data, ownership and design.
Unclear value
Many pilots begin with a capability rather than a problem: we should try generative AI, or we should see what this model can do. Without a specific problem and a measure, there is no way to tell whether the pilot succeeded, and no business case for the investment needed to scale it.
Poor data
A demonstration on clean sample data hides the problems that appear with live information. Real customer records contain duplicates. Product codes differ between systems. Fields that should always be filled are empty a third of the time. Definitions vary: one region’s “active customer” is another’s “lapsed”. A model that performed well in the pilot struggles as soon as it meets the real thing.
No owner
Pilots are often run by an innovation team or a technology function. When the pilot ends, nobody in the business owns the result, has budget for the next stage or is accountable for the benefit. The pilot report is filed and the team moves on to the next experiment.
Pilots never designed for production
A pilot built quickly on a laptop, with manual data extracts and no security review, cannot simply be switched on for the whole organisation. It has to be rebuilt, and the rebuild often costs more than the pilot itself. By then, enthusiasm has faded.
The data question to ask first
Before any AI pilot, ask: does the data we need exist, can we get to it, and would we trust a decision made on it today? If the answer to the last question is no, the first project is a data project.
A better path
The organisations that move from pilot to payback tend to follow a similar sequence. It is not glamorous, and it is slower at the start, but it produces systems that last.
- Pick one problem people feel every week, with a measure you can track, such as hours spent, errors made or response times.
- Check whether the data you need exists and can be trusted. Look at real records, not summaries, and involve the people who create and use the data.
- Build a small version on real data and let real users test it in their normal work, not in a demonstration room.
- Decide in advance what result would justify going further, and what would mean stopping. Write both down before the results arrive.
- Add basic governance from day one: a named business owner, rules for the data used and a human check for anything consequential.
Consider a typical insurance broker that wants to speed up the handling of renewal documents. A pilot that extracts key terms from fifty hand-picked policies looks impressive. A better pilot takes a random week of real renewals, including the scanned, handwritten and incomplete ones, and measures how much time the tool saves the people who process them. The second pilot is less flattering, but it tells the truth about value and readiness.
From prototype to production
A prototype proves value. Production needs much more: security and access controls, integration with live systems, monitoring of accuracy and cost, a support process, training for users and documentation for the people who will maintain it.
None of these needs to be built during the prototype, but all of them should be planned. Ask early which systems the solution will need to connect to, who will support it, what it will cost to run at full volume and which security reviews it must pass. Answering these questions during the prototype means a promising pilot does not have to be rebuilt from scratch.
After launch, review cost and quality regularly. AI systems can drift as the data they see changes, and running costs can grow faster than expected as usage rises. A monthly review with the business owner keeps both under control.
The pilot took six weeks. Getting it ready for every branch took four months, and we would have saved two of them if we had asked the production questions at the start.
Head of operations, financial services client
What leaders can do
Leaders set the conditions that decide whether AI projects pay back. A few choices make the biggest difference.
- Fund fewer pilots, with clearer measures, rather than many loosely defined experiments.
- Require a named business owner before any pilot starts.
- Treat data quality as an investment in its own right, not a side task for the AI team.
- Reward teams for stopping weak ideas early, not only for launching new ones.
The takeaway
Progress with AI comes from a sequence of small, governed wins on real data, each with an owner and a measure. Stopping a weak idea early is a good outcome, not a failure: it frees time and money for the ideas that work.
If your organisation has several stalled pilots, review them against the four causes above. Choose the one with the clearest value and the most usable data, appoint an owner, and plan its path to production. One AI system running reliably in a core process will teach you more than ten pilots on clean samples.
Source: McKinsey, The state of AI in 2025; Gartner, February 2025.
