The Pilot Trap

Almost everyone now has an AI pilot. Far fewer have AI in production. That gap, between a promising proof of concept and something people actually use every day, is where most organisations are quietly stuck.  

The numbers bear this out. In its 2025 State of AI in Business study, MIT’s Media Lab found that only around 5% of enterprise AI pilots reach production with measurable value; the other 95% show no discernible return. The striking part isn’t the failure rate. It’s the cause. MIT concluded that the difference wasn’t model quality or regulation; it came down to approach.  

That matches what we see at Marra. As a North East technology consultancy working with organisations, providing business solutions using AI and low-code solutions, we’re rarely called in because the technology failed. More often, we’re called in because everything around the technology has stalled progress.

The demo is not the destination 

A pilot is built to answer one question: could this work? A production system has to answer a harder one: will people actually use it, day to day, as part of how they already work? These are different problems. The effort, discipline and mindset needed to cross the gap from the first to the second are routinely underestimated.  

Too many pilots are designed to impress a steering group rather than to survive contact with real work. They run on clean sample data, in a safe sandbox environment, with the enthusiast who championed them sitting next to the screen. Remove any one of those supports and the whole thing wobbles.  

Why pilots stall  

When we look under the bonnet of a stalled initiative, the same handful of causes come up again and again. None of them are really about the model. Here are five of the most common issues we see. 

1. It started with the technology, not the problem

“We should do something with AI” is an ambition, not a business case. If there is no clear problem to solve, there is nothing to scale. 

2. Nobody defined what “good” looked like

If you cannot measure the value a pilot creates, you cannot make the case to invest in scaling it. “It felt quicker” is not enough in a budget conversation. 

3. Adoption was treated as a launch, not a change

AI changes how people work. If you drop a tool on a team without bringing them with you, they will quietly go back to the manual spreadsheets they trust.

4. The unglamorous work was deferred

Integration, data quality, permissions, security, and governance are easy to skip in a pilot but impossible to skip in production. If left too late, they become the reason the project stalls.

5. There was no one to hand it to

Pilots are often run by innovation or leadership teams who cannot own a live service. When the pilot works, there is no team, no budget, and no owner ready to take it on. 

All of these issues point to something we see often at Marra: the biggest challenge with AI is not the technology. It is working out what to do with it, when to do it, and who to do it with.

What “ready to scale” actually looks like  

One of the questions we like to put to leaders is deceptively simple: at what point does an AI pilot become something worth scaling, what signals do you look for? A few we’d flag include:  

• A clearly defined business problem – AI is being used to solve a real issue in a business function or process. It is not technology looking for a use. 

• A measurable baseline – You know how you will measure success, so improvement is clear and not just anecdotal. 

• A repeatable workflow – The team knows how they will use the AI tool to automate a process or solve an issue, not just as a one-off demo. 

• A named owner – Someone is responsible for the AI tool after go-live, so it does not get neglected or underused. 

• Users who miss it – The clearest sign an AI tool is ready to scale is when people ask for it back if you take it away. 

Crossing the gap

Closing the gap between pilot and adoption is less about better models and more about better delivery. At Marra, our approach comes down to a few key principles:

1. Start with the problem, not the tool

Begin with the business outcome you’re trying to achieve and let that decide where, and whether, AI is needed in your organisation.

2. Diagnose before you deliver

A short, structured triage helps separate ideas worth piloting from those that only sound good in the room.

3. Design the pilot for production from day one

Use real data, real integration, and real governance in any pilot. Build the path to scale in from the start, not as an afterthought.

4. Treat adoption as delivery work

Enable the business users who will actually use it. Make sure they know how to get value from it, rather than assuming a good tool will sell itself.

5. Resource the transition

Managed delivery and team support give a pilot the external expertise it needs to become a service. Leaving it in the hands of an already busy team without the experience or time rarely leads to success. 

Conclusion 

The organisations that pull ahead with AI will not be the ones with the most impressive demos. They will be the ones who treat the pilot as the first step of delivery, not the last step of an experiment, and who focus on what matters once the novelty fades: measurable value and enabling their teams to use the tools well. 

If you want help shaping your own AI or digital strategy, our team at Marra is always happy to talk. Get in touch to speak with one of our consultants.

Written by Ben Dawson, Business Development Executive