What Is an AI Factory? How to Set One Up Inside a Large Organisation

What is an AI Factory?

The word factory is deliberate. A lab explores, a centre of excellence advises, a factory produces, and it produces the same kind of thing more reliably each time. I have been Head of AI Factory at Devoteam since 2026, and this article gives the definition I can defend, then what I did.

Why do large organisations need one?

Without a factory, every team rebuilds the same things. One team builds a retrieval pipeline, another builds a slightly different one, a third buys a tool that overlaps with both. Evaluation is improvised, security reviews start from zero and nobody can say which use cases created value. The cost is not only money. It is the time during which good ideas wait in a queue because nobody owns the shared parts.

A large group makes this worse. At Devoteam we are 10,000+ consultants in 25 countries, and demand for AI comes from sales, delivery, internal functions and clients, all at once and in different languages. A factory is the answer to a coordination problem: one place that owns the reusable parts, so that the people closest to the business can focus on the problem in front of them.

It is also an answer to a skills problem. Tools without trained people produce demos, and trained people without tools produce frustration. The factory has to treat both as one system.

What does an AI Factory contain?

My definition rests on three engines that feed each other. Remove one and the factory becomes something else: a platform team, a consultancy or a training department. The table shows what each one means in my case.

How did I set up the AI Factory at Devoteam?

I joined Devoteam in 2023 as Principal Product Manager, became Lead of AI in 2024 and took the Head of AI Factory role in 2026.

The platform pillar is NOVA Suite. It was built by a cross-functional team of 15 people, and I insisted on product discipline: a roadmap, users, a feedback loop and release criteria, not a collection of prototypes. It reached 800 employees.

Client delivery is the second pillar, and it keeps the platform honest. On the DILA platform, shared components and evaluation made delivery faster. At Decathlon, the public figures from my CV show a €4M programme with 13 people and 2.6x productivity gains from augmented teams. At Carrefour, the work was applied machine learning in a retail supply chain. Each engagement fed requirements back to the platform.

Enablement is the third pillar, and it should not wait for the platform to be finished. LEAP, our internal learning programme, was built with a 5-person team and has trained 1,500+ professionals, and Prompt Arena gives people hands-on practice. A factory with no trained users ships tools into silence.

How does an AI Factory run day to day?

A factory is a flow, so the rhythm matters more than the structure. These are habits that make a factory work, in general terms:

• One intake for AI demand, so that a request from sales, a client or an internal function is triaged in the same way.
• A shared backlog where platform needs from engagements compete openly with roadmap ideas.
• A rule that anything built twice becomes a platform candidate.
• Evaluation before scale: no use case widens its audience until it has been tested against agreed criteria.
• Named owners for governance questions such as data classification, human approval and acceptable use.
• A learning loop where delivery teams report what broke, and training content is updated from it.

What goes wrong with AI Factories?

The failure modes below are predictable, which is the good news.

• A factory without demand. Teams build platforms for use cases that nobody asked for. Start from a business problem and from clients.
• A factory that becomes a bottleneck. If every AI request must pass through one central team, people route around it. The factory should own the shared parts, not every decision.
• A platform that ignores delivery. If the platform team never works on client problems, it optimises for elegance, not for use.
• Training as an afterthought. Rolling out a tool to thousands of people without practice produces low usage and a few enthusiasts.
• Governance as a final gate. Security and compliance involved at the end create delays. Involve them from the first design.
• Counting activity instead of value. Number of prototypes, workshops and attendees say little about the business.
• Public repositories mistaken for production. A published design is not a validated deployment, and I try to be clear about that distinction in my own work.

How do you measure an AI Factory?

Measurement is where most factories are weakest, so I treat it as a design constraint. I use four lenses, and I avoid a single headline number, because each lens answers a different question.

What did I see?

I want to separate facts from interpretation. The facts: NOVA Suite is used by 800 employees, built by a 15-person team, inside a group of 10,000+ consultants in 25 countries. LEAP was built with a 5-person team and has trained 1,500+ professionals. The AI Factory carries a €12M+ commercial pipeline.

On DILA, the public-sector platform, the figures I can cite are 70% faster content creation and a 60% reduction in RAG application delivery time, from 10 weeks to 4, with 75% of common technical capabilities reused. This is the pattern I trust most: a shared platform and disciplined evaluation make delivery faster, and the speed is visible in client work, not only on a dashboard.

My interpretation, which is a judgement rather than a measurement: the three engines reinforce each other. The platform gets sharper because delivery exposes real constraints, delivery gets faster because the platform exists and enablement creates both users and demand. I cannot isolate the contribution of each engine, and I would be suspicious of anyone who claims to.