Why AI Projects Fail
in Production

I recently attended DataVision EMEA 2026 at UBS in London. The conference brought together senior figures from banking, market infrastructure, technology and data management under the theme "Gold In, Gold Out: Building Trusted Data to Optimise AI."

That framing was deliberate. It moved the conversation beyond the familiar warning of "garbage in, garbage out" and towards a more useful question: what does good data actually look like when an organisation wants to use AI properly, at scale, and inside live business processes?

The clearest lesson I took away was this: the main obstacle to successful AI is often not the model. It is the condition of the data within the organisation beneath it.

The Gap Between an AI Demonstration and a Production System

It has become remarkably easy to produce an impressive AI demonstration. A small team can take a carefully selected dataset, define a narrow problem, and show a convincing result in a short period of time.

Putting that same system into production is a different matter entirely. Production requires working with old and new technology side by side, inconsistent definitions, incomplete metadata and local workarounds. It also requires privacy, confidentiality, security, regulatory and contractual obligations to be addressed properly. An AI demonstration proves that a model can do something interesting. Moving it to production requires the organisation to prove it can do the same thing repeatably, lawfully and at scale.

The additional cost of reaching production is not really the model. It is the work needed to make the organisation's data dependable enough for the model to use.

One panellist used the image of children building a sandcastle on a beach: something impressive goes up quickly, but when the tide comes in it destroys what was built, because it never had solid foundations — just sand. Many AI pilots are digital sandcastles: convincing in a protected environment, but unable to survive contact with the operating business.

AI Amplifies What Is Already There

James Dallas of UBS described AI as an amplifier sitting on top of data. Where the data is strong, AI can accelerate sound analysis and better decisions. Where the data is weak, it accelerates poor decisions just as fast.

AI is also exposing weaknesses that have usually stayed hidden, because experienced staff quietly fix the errors and make imperfect systems work. An experienced team knows which source is usually right, which field is unreliable, and which exception needs checking — and corrects the errors before the information ever reaches a report or a spreadsheet. Those protections are not there when the processing moves to an AI, because the AI has none of the local knowledge that a team builds up over hours of working with the data and the systems around it.

If an organisation does not have a coherent way of identifying its important data, tracing where it came from, assessing its reliability, and controlling how and by whom it may be used, it has no supporting structure. And without that structure, AI has nothing dependable to stand on.

AI Strategy Is Data Strategy

The strongest lesson from DataVision was not that organisations need less ambition around AI. It's that they need stronger foundations beneath that ambition.

The eventual winners are unlikely to be decided simply by who has access to the largest or most advanced model — models will keep improving, and access to them will keep broadening.

The more lasting advantage will come from the ability to provide those models with data strategy discipline: data that is trusted, well defined, reusable, and lawfully accessible.