The model is the easy part: why your data decides what AI can do for you

The closing session at AI Scotland landed on an idea I have been turning over ever since: applying what industry knows, and the data it holds, to build small language models tuned to a sector. Not another general-purpose chatbot, but a compact model that understands the language of subsea maintenance, care home rostering or freight compliance, because it learned from the records of people who do that work every day. It is a compelling picture, and for Scotland it plays to a real strength. We have deep sector expertise in places like Aberdeen, and much of it lives in data nobody outside the sector will ever see.

The technology has moved in its favour. Small models of a few billion parameters now run comfortably on modest hardware, so they can be hosted in the UK, or even on a business's own infrastructure, rather than sent to a data centre in another jurisdiction. That matters for anyone following the sovereignty debate. If inference runs where you control it, on data you own, many of the harder questions about who can see what start to answer themselves.

My view is that the model is the easy part. Base models are becoming close to a commodity, and they get cheaper and better every few months. What makes a sector model valuable is the data behind it, and that is exactly where most businesses are least ready. In 30 years in technology, most of them running IT for Scottish businesses, I rarely met a company whose records were clean, consistently labelled and clearly owned. Maintenance logs live in three systems and a spreadsheet. Customer history sits partly in a CRM and partly in someone's inbox. None of that is unusual, but none of it is ready to teach a model anything useful.

There is also the question of what it takes. Fine-tuning a small model no longer needs vast datasets, and a few hundred to a thousand good examples of a specific task can be enough. The catch is the word good. Those examples need to be consistent, correctly labelled and yours to use, and someone needs the skills to prepare them, run the training and check the results. For most businesses, the better route today is to leave the model alone and let a capable one work over their own documents, with sensible controls on what it can reach. That delivers most of the value at a fraction of the cost and risk, and it avoids locking you into something you will need to rebuild when the next generation of models arrives.

Where sector models become interesting is when an industry pools what it knows, and oil and gas is already doing it. EnRG-LLM, built by Aramco, the Society of Petroleum Engineers and i2k Connect, was trained on SPE's technical papers, books and reference material to create a model that understands petroleum engineering better than general-purpose alternatives. That is the pattern worth copying: shared, properly governed knowledge rather than one company's records on their own. Getting there is a convening job as much as a technical one. It means agreeing what gets shared, on what terms, who owns the result and where it runs. Scotland's economic development bodies and industry groups are well placed to lead it, and the North East's energy sector is an obvious place to start.

None of this happens without the unglamorous groundwork first. Before any business can contribute to a sector model, or benefit from one, it needs to know what data it holds, where it lives, who is responsible for it and what it is allowed to do with it. That is data readiness, and it is also good governance. The businesses that do it now will have options later. The ones that do not will be buying whatever someone else built, on someone else's terms.

The question is not whether you should be building a model. For almost everyone, you should not. The question is whether your data is in a state where you could, and whether you would be ready to take part if your sector decided to build one together.

Small models trained on sector knowledge will be part of how Scotland competes. But the advantage will not go to whoever has the best model. It will go to whoever has their data in order.

Source references: closing session, AI Scotland conference, 2026. "Building EnergyLLM: A Domain-Specific Large Language Model Trained on SPE Content", SPE Annual Technical Conference and Exhibition, 2025.