Most organisations we talk to have already tried their hand at AI. They ran a pilot, saw a promising demo, perhaps built a model that performed well on a slice of historical data. And then it stopped. The solution never reached production, or it got there and was never used. The pattern is common enough that it is worth being honest about why it happens.
The short version: AI projects rarely stall because of the model. They stall on everything around the model.
The bottleneck is the data foundation, not the algorithm
A model is only as good as the data it is fed. That sounds obvious, but the consequences are almost always underestimated. In a pilot you work with a clean, curated dataset that someone has hand-picked and scrubbed. In production the same model meets reality: data arriving from five different source systems, with inconsistent formats, missing fields, duplicates, and definitions that mean different things in different departments.
This is where most initiatives actually run aground. Not because someone picked the wrong algorithm, but because there is no data foundation delivering the right data, at the right time, in a format the model can use – continuously, not just the one time in the pilot. Without data flow, data quality and a clear understanding of what the data actually means, even a good model becomes unreliable the moment it meets real traffic.
In our experience, the bulk of the work in a successful AI project lies before the model becomes interesting at all. It is about collecting, structuring and quality-assuring data – building the very foundation the solution will rest on. It is less glamorous than talking about large language models, but it is what decides whether the value is ever realised.
Demo to production is a discipline of its own
A demo has to impress once. A production solution has to give the right answer a thousand times a day, including when something goes wrong. The distance between the two is greater than most people assume, and it calls for an entirely different skill set.
In production you have to answer questions a demo never asks: How does the model fit into the workflow people already have? What happens when an input looks unlike anything the model has seen before? How do we monitor that the model still performs as the world shifts underneath it – what we call model drift? Who is alerted when something fails, and what is the fallback? How do we handle traceability, access control and privacy once the solution is suddenly processing real customer data?
This is not afterthought work you can defer. It is delivery discipline, and it has to be built in from the start. An AI solution destined for production is first and foremost a software system – with everything that implies in terms of architecture, testing, operations and ongoing development. The model is one component among many.
Buy the delivery, not the hype
Our starting point is that customers do not really buy “AI”. They buy a problem solved, and a solution that holds up over time. AI is a means, not an end. When an initiative is framed as “we have to do something with AI”, it often starts at the wrong end – with the technology rather than the business problem and the data needed to solve it.
We believe in the opposite order. Begin with a concrete problem where there is data to build on. Assess honestly whether the data foundation is good enough, or whether it has to be built first. Plan for production from day one, not as a phase that comes “later”. And let one delivery partner own the whole journey – from architecture and data foundation, through development, to operations and further development – so nothing falls through the cracks in the hand-off from model to system.
There is no shortcut here, but there is a viable path. AI reaches production when you stop treating it as an experiment and start treating it as a proper delivery. It is less exciting to say in a board meeting, but it is the difference between a pilot that impressed and a solution that actually gets used.