What we have learned turning AI into value

It is tempting to treat AI as a feature, a model dropped into an existing process to save a few minutes. The organisations pulling ahead treat it as a change agent: a reason to rethink how work is done, who does it, and where judgement sits.

Production AI forces clarity. It exposes where data is poor, where decisions are undefined, and where accountability is vague. Confronting that is transformation, not a plug-in.

This is why AI and the operating model cannot be designed apart. We design them together, and we stand up an AI Centre of Excellence so the capability scales with governance rather than sprawling into shadow projects.

Used well, AI does not just make the old way faster. It makes a better way possible.

As enterprises move from one model to many, the question changes from “which model” to “how do we run many models safely.”

A model garden is an architecture pattern: a governed, central place where models, foundation, fine-tuned and bespoke, are catalogued, served, monitored and swapped behind a stable interface.

Its power is separation. The application talks to a capability, not a specific model, so you can adopt the best model for each task, route by cost or quality, and replace one without rewriting the product.

Around the garden sits the discipline that makes AI safe at scale: evaluation, guardrails, cost controls, observability and clear ownership. It is the difference between a clever demo and AI you can run a business on.

Too much enterprise architecture ends as shelfware: elegant diagrams no one uses and that delivery quietly ignores. Architecture earns its keep only when it speeds decisions and delivery.

We treat architecture as a living asset. A clear target state, a small set of standards that actually matter, and a transition path tied to the business case, not a model maintained for its own sake.

Done well, it is the difference between change that compounds and change that collides with itself. It is also how you stop paying twice for the same capability.

Most business cases are written to win funding, then never reopened. Benefits erode quietly, and no one notices until a review concludes they were never realised.

We instrument value from day one: baselines, owners, and a small set of metrics tracked as hard as cost, through a Value Realisation Office that answers to the board, not the programme.

It changes behaviour. Teams stop reporting activity and start moving the number, and investment flows to what actually works.

Innovation theatre, the labs, the hackathons, the demos that never reach a customer, is comfortable, photogenic and largely a waste. It produces motion, not change.

We bias hard to production. Small, real bets, built and run, with a fast path from idea to live and an honest decision to stop what is not working.

Innovation that ships changes the business. Innovation that demos only changes the slide.

The most expensive myth in enterprise AI is that you must perfect your entire data estate before you can start. It sends organisations into multi-year foundation programmes that bank no value and quietly exhaust the appetite for AI.

The evidence points the other way. Data readiness is the most cited obstacle to AI, yet the organisations that win do not wait for it. They sequence by value: they take the use case worth the most, build only the slice of data that case needs, and let that case fund the foundation underneath it.

Done this way, the foundation grows in the shape of real demand rather than a theoretical model, and the first value arrives in a quarter rather than years. Each case leaves behind governed, reusable data that the next one stands on.

Perfect data is not the entry ticket to AI. It is a by-product of doing AI well, one valuable use case at a time.

The quiet failure mode of transformation is not a missed deadline. It is the slow erosion that begins the day the advisors leave, when the new way of working depends on people who are no longer in the room.

Resistance to change is the single biggest barrier to AI adoption in the region, and it is rarely solved by a training deck. Change holds when the people who will run it have built it, owned it, and been left genuinely able to carry it.

So we treat adoption as the deliverable, not a workstream bolted to the end. We embed inside the teams who will operate the change, transfer the capability deliberately, and stay long enough to see it run without us.

The test of a transformation is not how it looks at go-live. It is whether it is still working, and still improving, a year after the consultants have gone.

Each wave of technology asked more of the enterprise underneath it. Agentic AI asks the most of all, because it does not just inform a decision, it takes the action. Point it at a broken process and it will execute the broken process faster.

That is the opportunity and the risk in one. The value is real: agents can compress work that took weeks into hours. But an agent inherits every weakness in the data, the process and the controls beneath it, and then acts on them at machine speed.

So we treat agentic AI as an operating-model decision, not a tool rollout. We give agents clean inputs, clear boundaries and an audit trail, and on anything safety-critical or in operational technology, a person stays in command of the action, by design.

We have watched this pattern repeat for two decades. A new technology arrives, expectations spike, budgets follow, and roughly seven in ten programmes fall short of what they promised. The names change, from SOA to cloud to robotic automation to generative AI, but the shape of the failure is identical.

The reason is consistent too. Across the failures, about four-fifths of the problem is change, people, process and operating model, and only about one-fifth is the technology itself. Organisations that skip the human and operational work do not escape it; they simply pay for it later, at a higher price, on the next wave.

The lesson is not to slow down. It is to put the weight where the failure actually lives. Fix the foundations just enough to carry this wave, build the change in with the technology, and you stop re-learning the same expensive lesson every few years.

In this region, AI does not run in a vacuum. Data residency and sovereignty rules, sector regulators and model-risk expectations all set the terms, and they are tightening. Bolt governance on at the end and it becomes the thing that stops a working pilot from ever reaching production.

Built in from the start, it does the opposite. Explainability, model risk, data lineage and human authority designed into the architecture are what let a bank put a model in front of a regulator, or an operator put automation onto a safety-critical line, with confidence rather than hope.

We design for the rules of the place we work in, from national data and AI authorities to financial and free-zone regulators, so that governance accelerates the safe use of AI instead of blocking it.