Most AI Centers of Excellence die the same way. They launch with a manifesto, hire a few data scientists, run three pilots, and eighteen months later the only thing shipped is a slide deck. I have watched this happen. I have also helped fix it. The difference is never the technology.
An AI CoE that ships is not a research lab. It is a factory. Its job is to move a business problem from idea to production to adoption, then measure whether anyone actually got value. Everything else is decoration.
Here is how I build one that survives contact with reality.
Start with a portfolio, not a platform
The first mistake is buying infrastructure before you know what you are building. Teams spend nine months choosing an MLOps stack for use cases that will never exist.
Flip it. Begin with a ruthless use-case pipeline. I score every candidate on three axes: value, feasibility, and readiness of the underlying data and process. Then I fund the top few and kill the rest publicly.
- Value: quantified in euros, hours saved, or risk reduced — not "strategic importance."
- Feasibility: can we ship a usable version in one quarter?
- Readiness: does the data exist, and does anyone own the process it touches?
A pipeline forces choices. Killing a beloved pet project in front of the steering committee is the single most credibility-building act a CoE leader can perform.
Governance that speeds you up
Governance has a bad name because most of it is a brake. Good governance is a guardrail — it lets you drive faster because you know where the edge is.
I keep it to three questions, answered before any build:
- Who owns the decision the AI influences, and are they accountable for the outcome?
- What is the human fallback when the model is wrong or unavailable?
- How do we monitor drift, bias, and cost once it is live?
Notice what is missing: a 40-page policy. Policies are read once. A one-page intake form that every use case must pass gets used daily. Make the right thing the easy thing.
A CoE is judged not by the models it trains, but by the decisions it improves.
Change management is the product
This is the part engineers underestimate and where I spend most of my time. A model that is 90% accurate and 20% adopted delivers less than a model that is 75% accurate and used by everyone. Adoption is not a rollout phase at the end. It is the product.
Concretely, that means I embed the CoE inside the business unit, not beside it. The people who will use the tool help shape it from week one. I ship narrow and early, get it into real hands, and let the first users become internal champions who sell it better than any executive email.
I also budget for the unglamorous work: retraining people, redesigning the workflow around the model, and retiring the old process. If the manual spreadsheet still exists, people will quietly go back to it the first time the AI surprises them.
Metrics that a CFO respects
If your CoE reports "number of models deployed," you have already lost the room. That is an activity metric, not a value metric.
I track three tiers:
- Adoption: active users, frequency, and share of eligible decisions the tool actually touches.
- Business outcome: the euros, hours, or error rate the use case was funded to move — measured against a real baseline.
- Health: model performance, cost per prediction, and incident count, so ROI does not quietly erode.
Every funded use case gets a baseline before it launches. No baseline, no funding. It sounds bureaucratic; it is the only thing that lets you say "this saved 12,000 hours" and have the number survive an audit.
The operating rhythm
A CoE that ships runs on a heartbeat. Quarterly, I refresh the portfolio and kill what is not working. Monthly, I review live use cases against their outcome metrics. Weekly, the delivery squads ship something a real user can touch.
That cadence does two things. It creates a steady stream of small wins that build political capital. And it makes failure cheap and fast, so the expensive failures — the ones that show up in the annual report — never happen.
What actually changes
The organizations that get this right stop treating AI as a project with an end date. The CoE becomes the muscle that turns any business problem into a shipped, adopted, measured capability. That muscle is the real asset. The models are just its output.
Build the factory, not the demo. Fund value, not novelty. And remember that the hardest problem in AI transformation has never been the AI. It is getting a busy human to change how they work on a Tuesday morning — and to be glad they did.
Cédric Bignet is an AI & ERP Change Management expert at Novartis and founder of AInspire. He writes about change management, AI adoption and enterprise transformation.