What makes AI hard to deploy is not the model

At the end of October tDigital was at the British Ambassador's Residence in Stockholm for the UK & Sweden Entrepreneurship and Innovation Reception, held during Black History Month. The occasion was a visiting UK delegation, and the room was full of founders, investors and embassy staff from two countries that work together more than they talk about it.

The evening opened with the Ambassador's welcome address, and then the room did what rooms like this are for: introductions, the short version of what everyone is building, and the longer conversations that start once two people find they are working on the same problem.

We went expecting the usual conversations about markets and funding, and we had a few. The one that kept coming back, with different people and in different forms, was something much more practical: why AI is still so hard to get into production. Nobody blamed the model.

The hard part is not the model

One of the best conversations of the night was with Yasaar Biladama, founder of Eggsplain, who runs AI entirely on the customer's own premises. No cloud, no data leaving the building. We ended up comparing notes on what actually slows AI down once the pilot is over.

It is rarely the model. The models are good and getting cheaper. What came up again and again that evening was everything around the model: trust, security and workflow. Who gets to see the data. How security signs off. How the thing fits into the way people already work. An AI feature that lives in a demo tab next to the real system is a feature nobody uses.

What we see with clients: it starts with data

Trust, security and workflow are real problems, and we meet all three. But trace any of them back and you land on the data. Nobody can trust data nobody owns. Security cannot sign off on a flow nobody can map. A workflow breaks the moment the data it needs is missing. The three problems the room kept naming turn out to be one problem.

And the data is usually in poor shape. It sits in systems that were never meant to talk to each other. Parts of it are missing. Ownership is unclear. The people who understand it are the busiest people in the building. A model is only as useful as what you can feed it, and most organisations discover how little that is in the first week.

So the first question we ask is not which model. It is what data this needs, where it lives, who owns it, and what has to be true for security to say yes. That is architecture work before it is AI work, and it is why our AI Transformation accelerator opens with a Foundation phase: data pipelines, platform and guardrails first, use cases after.

Cloud or the customer's own server room makes no difference. If the AI cannot get at the data, it does not count.

Why rooms like this matter

Innovation moves faster when people from different backgrounds and industries sit at the same table. An on premise AI founder and a global consultancy reached the same conclusion from opposite directions. Sweden and the UK have more to build together than most of us use.

Thank you to Ambassador Samantha Job, Ashton Thorp and the British Embassy Stockholm team for the warm welcome. We are looking forward to continuing the conversations.

Where to start

If your organisation has AI ambitions, and data, security or workflow is what stands in the way, that is what our AI Transformation accelerator is built for. It takes you from scattered pilots to durable capability: use cases, platform, governance and people, delivered on live data and handed over as a process your own teams can run.