The Building Blocks: Scaling AI in Regulated Markets with Colin Carmichael, BAIA

Jim Clark: Colin, great for you to be joining us. You’re Client Partner at 4most, working with major financial institutions on data and AI transformation, and you’ve also co-founded the Business AI Alliance to give SMEs and mid-market firms a voice on AI. How far ahead are big organisations versus the firms you represent in the Alliance in terms of AI adoption?

Colin Carmichael: First thing to be aware of: the Business AI Alliance is a community of AI experts. When we’re comparing my day job in financial services to the community within the Alliance, there are service companies, product companies, and actually some banks in there — a whole plethora of different sectoral focuses. They’re all AI experts in their organisations, either building products or services using AI. Then there are organisations that don’t specifically focus on AI but have champions from within them. It’s a peer network where they can ask questions and get answers, because as everyone knows, it’s a very fast-paced environment.

Colin Carmichael: The bigger organisations in financial services are throwing a lot of money at AI — it’s probably the number one priority for all the big banks. Some of the smaller organisations say it’s their strategy, but a lot of times it hasn’t really been their strategy in practice. They’re thinking about it and just need a little helping hand along the way.

Jim Clark: What proportion are really mature versus finding their way?

Colin Carmichael: In financial services, the neo-banks — the digital banks, your Starlings, Monzos, Atom Banks — are definitely more mature. The reason is they’ve got the infrastructure in place. They’re effectively a little bit more nimble about how they operate. They’ve not got bogged down with legacy. They’ve got cloud, they’ve got all the latest technology, and they’re organised in a way that’s more aligned to how you would build products and roll them out quickly and experiment quickly. That’s one distinction.

Colin Carmichael: Then you’ve got some of the challenger banks and building societies. Some of them are going through digital transformations themselves — putting the building blocks in place, cloud provision, digital services — but they’re not quite as mature. Once they’ve got those building blocks in place, they’ll be at a point where they can start evolving their estates. But it really does come down to the energy of the board, the focus of the board, and whether individuals have time in their busy day jobs to start to experiment with AI as well. Everyone’s in completely different places.

Jim Clark: Of the firms in the Alliance, are they finding the gap is widening as they’re still thinking about where to move forward?

Colin Carmichael: The SMEs in the Business AI Alliance are building products using AI — they live and breathe AI every day. I think the gap is probably not as big as everyone thinks. If you look at the last three and a half years since generative AI came along, in the last twelve months most organisations rolled out products like Copilot to their employees. There’s a lot of familiarisation across most organisations with what to do with AI, and people are just experimenting and trying to understand where they have opportunities or challenges to address through AI.

Colin Carmichael: Last year was definitely experimentation. This year, I would imagine across the sector, it’s all about the MVPs — the minimal viable products — that firms can start to focus on. The ones with the highest value that they really want to try and industrialise. So the noise has been filtered out, people have got used to it, and now it’s: here are two or three great ideas we want to take forward because we see the best value there. Next year then, as long as organisations have their technology, data and operating model in place, it becomes how do we start scaling that? But everyone is working at the same pace because generative AI only came out at the same time, and they’re all just lifting their knowledge base and experience at the same time.

Jim Clark: As firms move forward looking at how they can make best use of AI, they’re having conversations with you about data quality and governance. Is that something that comes up first or afterwards?

Colin Carmichael: When people are experimenting, they find that AI is nothing without the data — I think anyone who’s been using it understands that. A lot of organisations have obviously been going through their own digital transformations and data transformations, and even in the financial services sector, regulation has driven a lot of focus on data and the governance and controls that go with it. AI is just another lens really on top of all the other programmes of work going on within an organisation. It is hugely important, but it’s been hugely important for most organisations for many years. Without good data, they’re unable to interact properly with their customers, personalise their services and products, or answer the questions regulators are always asking. AI is accelerating the need to get the right governance around data and transparency and trust in the data you’re using.

Jim Clark: Big financial firms know they can’t just deploy AI and hope it works. But many mid-market firms in other sectors are basically doing exactly that. What should they be testing before they go live?

Colin Carmichael: I think it’s more about considering the outcomes. If you’re going to be pushing out solutions to your customers or even internally, who owns the outcome? Who’s signing off the ongoing use of that outcome? What if something goes wrong? Who’s accountable for that? And what happens if you’ve pushed out a solution to customers and something goes wrong — what remedial activities do you put in place? It’s all about ownership and understanding what to do if something goes wrong. Don’t just push something out without considering it all.

Colin Carmichael: When you work for a small business or a start-up, you tend not to think like that. You’re excited about the technology and you want to push things out to customers and deal with the fallout later. But if you’re working in a regulated environment, you can’t do that. Focus on your customers and what your remediation activities will be — and who’s going to own that remedial activity. What is that process if you need to apologise to customers?

Jim Clark: Do you think concern about governance has given some firms an excuse not to move forward quicker?

Colin Carmichael: I don’t believe so anymore. I think there was a small moment in time when generative AI was first released where that may have been the case, but that was just a lack of understanding of what it was. Now it’s such a focus. The board sets out the mandate, the CRO aligns to the mandate, and they have to make it happen. They own the risks. The board is going to want AI or a strategy around AI, so they need to support it. I generally don’t think there are any inhibitors any more when it comes to the risk function, certainly in financial services.

Jim Clark: Where are the firms making the most progress? What characteristics are contributing to it?

Colin Carmichael: The tier one banks — the large regional banks in the UK — there was one I was working with in particular that had put the building blocks in place, the ways of working in place, the innovation functions in place. They’d been experimenting for quite a number of years but were very focused on scaling. When generative AI came out, they were effectively ready. They could start building things quickly and get them rolled out. Going back to the core point again: operating model, technology, data, governance. If you get all four of those ingredients right and they’re in place, you can just start quicker. It’s as simple as that.

Jim Clark: Where does the talent question fit into this? Budget and talent shortages are often cited as holding mid-market firms back.

Colin Carmichael: I don’t think talent is the problem. I think it’s the time. The big organisations have the investment to build innovation teams, whereas small banks don’t. So you’re really reliant on pressing on challengers across a smaller organisation to experiment, perhaps even in their own time, or trying to find time in their day jobs to get their heads around it. But when you’re busy already, it’s quite challenging. Some organisations that are smaller will go: this is important, we’re going to set up a team and we’re going to invest in it. But everyone’s at completely different parts in their journey with different reasons for where they are.

Jim Clark: Moving from experimentation to operational AI — what’s the jump there and why are organisations struggling?

Colin Carmichael: Proof of concepts are great — they get everyone understanding the technology. But moving to an operational state is different. It’s about the building blocks: your model, your technology, your data, your governance. You need to start thinking about how you interact with the foundation models you’re effectively building solutions on. It’s not just a case of putting a prompt into an agentic workflow and getting an answer out of the box. There’s actually a lot of customisation that needs to go on. That’s when you start to get a lot more technical. If you’re really looking to customise your outcomes to give you the trust and transparency that something in an operational model requires, that is the big jump. A lot of organisations are struggling with it. Because when you’re operationalising something, it needs to be more or less right all the time. So how do you customise your prompt engineering to ensure you’re getting as close to perfection as possible? There’s a bit of a dark art around that one.

Jim Clark: How are the organisations making progress actually justifying the AI investment?

Colin Carmichael: In the service industry, AI is built into how you deliver outcomes for your clients now. If you get accelerators and products right to complement your deliveries, the cost to the client will drop. You’re more likely to win the deal if you’ve got complementary accelerators alongside your services rather than a pure service play, and that drives higher gross margins as well. But for the clients themselves in terms of justification — I think it’s gut feel, to be perfectly honest, a lot of the time. They know they want to invest so they have to just take the plunge. It’s quite hard to really justify return on investment, certainly in the early days or months. Eventually they’ll get there. But it is hard, especially going through that learning curve.

Jim Clark: What’s an indication that they’re on the right track before they get to that end result?

Colin Carmichael: They tend to do a lot of experimentation internally first, rather than pushing it straight out to the customer. Let’s try and streamline our own internal processes and then try and qualify what that has saved in costs. So there’s already a saving. Most costs — certainly in the banking sector — they’ll be focusing on their own internal process improvements. What took ten person-days is now taking two. Multiply that across an organisation and there’ll be a definite cost benefit in the longer term.

Jim Clark: If a mid-market CEO came to you and said their team’s not ready for AI, what’s the first question you’d ask?

Colin Carmichael: I think it’s going to come down to: are they not ready to embrace AI? The key thing for the UK government and for any organisation is education. AI unfortunately makes a lot of people uncomfortable about their own futures. By embracing it, does that mean their future is going to be compromised? There’s an obvious nervousness around it. But the key thing is taking them on the journey, making them feel part of it. Hopefully they’ll gain from it and start to see a little more of a future for how they can apply their experience with AI.

Colin Carmichael: Think about it this way: you’ve got kids coming through today who are using AI all the time and are comfortable with it. They pick up technology quickly. Older generations have been doing sometimes repetitive jobs and haven’t fully embraced new technology because they’ve never really needed to. But what they’ve got are all the business problems. And the kids coming through have the technology without the business problems. So it has to meet in the middle somewhere. The chief executive just needs to educate — and there are lots of education programmes out there from government and from organisations like the Business AI Alliance. Leave from the front. Once you help people realise that freeing themselves from the weeds means they can spend even more time with clients, that realisation really kicks in.

What is the Business AI Alliance and what does it do?

The Business AI Alliance (BAIA) is a community of AI experts founded by Colin Carmichael and others to give SMEs and mid-market firms a voice on AI — particularly in conversations with government that had previously been dominated by large tech companies and big corporates. Members include AI product companies, AI service companies and organisations from a range of sectors working on AI adoption. The Alliance functions as a peer network where members can ask questions, share experiences and get answers in a fast-moving environment, as well as acting as a convening voice to government on the needs of smaller businesses.

Why are neo-banks more advanced in AI adoption than traditional banks?

Neo-banks such as Starling, Monzo and Atom Bank have a structural advantage in AI adoption because they were built on modern cloud infrastructure from the outset. They have no legacy systems to migrate, no technical debt to manage, and their internal ways of working are already aligned to rapid product development and experimentation. Traditional challenger banks and building societies are going through digital transformation programmes to reach a similar starting point, which means their AI readiness depends heavily on how far along that transformation they are. Colin Carmichael argues that the four ingredients — operating model, technology, data, governance — are the deciding factors in how quickly any organisation can start.

What should mid-market businesses test before deploying AI to customers?

Colin Carmichael’s core question is: who owns the outcome? Before any AI solution goes live — whether internally or customer-facing — an organisation needs to have clearly defined who is accountable for the ongoing use of that outcome, what happens if something goes wrong, and what the remediation process looks like. In regulated environments this accountability is mandatory. In unregulated environments it is frequently overlooked, with organisations excited about the technology and keen to ship quickly. Colin’s advice is direct: don’t push something out without considering it all, and make sure you know who owns the apology if you need to make one.

What are the four building blocks for scaling AI successfully?

Colin Carmichael identifies four core ingredients that determine whether AI can be scaled: operating model (the internal ways of working and team structures that support rapid experimentation and deployment), technology (the modern cloud and data infrastructure that enables AI products to be built and maintained), data (clean, governed, interoperable data that AI can actually use), and governance (the controls, accountability structures and validation processes that ensure AI operates reliably and compliantly). Organisations that have all four in place can start quickly. Those missing one or more will struggle to get beyond proof of concept.

Why is moving from proof of concept to operational AI so difficult?

The jump from proof of concept to operational AI is where most programmes stall, according to Colin Carmichael. Proofs of concept are valuable for building understanding and enthusiasm, but operational AI is held to a much higher standard: it needs to be more or less right all the time, not just impressive in a demo. This requires significant customisation of foundation models — fine-tuned prompt engineering, guardrails against hallucinations, and real integration with existing data and workflows. It also requires clear governance: model validation, accountability structures, and processes for when the model gets something wrong. The skills required to do this are more technical and more specific than the skills required to build a proof of concept.

How should mid-market firms justify AI investment to their boards?

Colin Carmichael is candid that in the early stages, justifying AI investment often comes down to gut feel. Hard ROI is difficult to quantify during the learning curve. The most reliable early signal is internal process efficiency: what previously took ten person-days now takes two. Multiplied across an organisation, that creates a demonstrable cost saving. For service companies, the business case builds further as AI-enabled accelerators allow them to deliver at a lower cost, win more deals, and generate higher gross margins. The key discipline is to start with internal processes before customer-facing applications, build evidence of real savings, and use that evidence to justify the next stage of investment.

What is the role of curiosity in AI adoption for mid-market firms?

Colin Carmichael argues that most organisations have a very small number of people with genuinely curious mindsets and the rest are simply getting on with their day jobs. For AI adoption to succeed, that needs to change. He draws a parallel between experienced operators — who hold all the business knowledge and problem context — and younger, technically confident joiners who hold the tools but lack the business depth. The most successful AI programmes find ways to bring those two groups together. Leadership commitment is also essential: if the board and CEO are not actively driving the agenda, curious individuals at lower levels of the organisation will struggle to create systemic change from the bottom up.