Most enterprise AI programmes reach a point where pilots work, but nothing scales. Every new use case starts close to scratch, and every team ends up solving the same problems again: data access, orchestration, security, integration. The way through is to stop building one solution at a time and start building capabilities that many use cases can share. That is what we mean by an AI factory, which is neither infrastructure nor a platform teams have to migrate onto, but a layer of ready-made capabilities they can pull from while keeping their own logic on top.

We build these with clients, and we have watched them run in production. Delivery time decreases substantially, and most of the technical stack gets reused rather than rebuilt.

The session goes through the architecture and then through the capability model, which is where it gets interesting. A capability sits somewhere between a shared service and a finished application, carrying its own logic, prompts and semantic layer, aligned to how a business process actually works, so that a new use case is composed from capabilities rather than written from nothing. We will also be clear about where the boundary sits, because knowing what a team still has to build for itself matters as much as knowing what it gets for free. From there we take one capability apart, end to end, including the parts that were hard to get right.

The workshop concludes on where things are heading. Agents work across systems rather than inside a single use case, and that raises questions nobody in the industry has fully settled, around keeping visibility of an agent once it leaves your own systems, handling identity for something that is neither a person nor a service account, and holding semantics steady across platforms that were never designed to agree. We would rather put those in front of the room than pretend they are closed.

About a third of the session is small group work, and it should be useful whether you are building agents inside a large organisation or selling into one.