Six months after the first team works well with AI, the usual outcome is ten ways of working across the department. The first has written rules and context the agent respects; the others started on their own, and leadership asks what happens to the other nine.
Every team ends up working its own way
Nobody decides that every team should work differently: it happens. Each one picks its tools and writes its own rules, and what it learns stays in that team. People also copy what they see next to them. A study of the rollout of coding agents at Microsoft found that first use spread mostly through social channels, and what travels with the use is the way of using it, bad habits included when there is no shared method.
DORA, Google Cloud’s research programme on software delivery, explains why this matters in its 2025 report. With nearly 5,000 professionals surveyed, AI adoption was associated with higher delivery throughput and better product performance, but still associated with lower delivery stability. Ten teams with ten ways of working give ten different results. The useful question for the leadership is how many teams share one way of working with AI.
Three ways to set up a centre of excellence
A centre of excellence is a small group that looks after how the whole organisation works. Arlotto and Irby, in an HFMA article on data analytics centres of excellence, distinguish three models: in the centralised one, a central group manages all resources, tools and methods; in the decentralised one, teams and tools live inside each unit; and the hybrid, with a core and several spokes (hub-and-spoke), combines the two. In the table I call them centralised, distributed and federated.
| Centralised | Distributed | Federated | |
|---|---|---|---|
| Who decides how to work | A central group | Each team | A small core sets what is common and each team decides the rest |
| What it gives | A standard and oversight across the organisation | A fast response to what each team needs | A common standard and a go-to person in each team (onext’s own judgement) |
| What it costs | It responds less well to each department’s needs | It is harder to scale across the organisation | It takes time from people who already have work (onext’s own judgement) |
onext’s own summary of Arlotto and Irby (HFMA, 2020; updated 2022).
For software development we pick the federated one, and we say so as our own judgement, with two supports. The first is the same article: the spokes are vital because they are more focused and agile than the hub. The second is Team Topologies, the framework by Matthew Skelton and Manuel Pais: an enabling team helps another team overcome obstacles and detects the capabilities it is missing. That is the role of the practice’s core: to help teams, because a group that approves every use of AI puts a queue in front of teams that deliver continuously.
There is one difference with the original model. In Team Topologies, the enabling team strengthens another team’s capabilities and then moves on. The practice’s core stays, because what it looks after (the guide, the shared context) changes with every new tool. The champion in each team follows the same logic as the AI champions of the Chief AI Officer: one person per team who drives change from within.
The limits: the HFMA article is about data analytics in healthcare organisations, Team Topologies is a framework built from practice, and DORA measures associations in surveys. None of the three sources is a controlled study of AI centres of excellence in development teams.
What the practice needs
One person in charge of the practice, a champion in each team and four contents:
- Guide to good practices. Specifications, team rules and automated checks that run on their own.
- Shared context. Common templates, rules and assistants, so the second team does not start from zero.
- Safe use of AI. Criteria on tools, data, security and licences.
- Training and onboarding. A learning path for the teams and a kit for whoever joins.
How it stays alive
A guide nobody updates falls behind when the tools change. The practice follows a four-step cycle: a team proposes a practice; it is tried in that team; if it works, it goes into the guide; and it is shared at the monthly meeting. Each practice reaches the guide after working in a real team.
Each step has its reason. A team proposing it avoids a guide written far from the real work. Trying it in that team gives evidence before asking anything of the others. Moving it into the guide gives it a single place to live, instead of spreading through conversations. And the monthly meeting is where teams discover what another one has already solved.
How it stays alive
With one of our clients, a monthly meeting is held where the teams share the good AI development practices they have adopted. It is the last step of the cycle: what one team has tried reaches the others.
Three simple indicators
Few indicators, and ones that are already familiar. The three, in line with DORA:
- Teams working with the method. How many teams follow the guide and use the shared context.
- Delivery time. DORA defines it as the time it takes for a change to go from committed to version control to deployed in production.
- Errors reaching production. The DORA equivalent is the change fail rate: the ratio of deployments that require immediate intervention after release.
None carries a figure promised in advance. They are measured at the start and again with every team that joins, and the reading is how each team moves against its own starting point. The three fit on one sheet and can be read at the monthly meeting, next to what each team has tried.
A practice that never ends
A change like this only lasts if it also reaches whoever joins later. If every new person learns to work with AI on their own, within a year the department is back to everyone working their own way. That is why the practice includes training and onboarding documentation, not just sessions with today’s teams.
But there is something more important: the transformation of the IT department never ends, because models, tools and ways of working keep changing every few months. The practice is the tool to steer that change continuously, from the inside. Our goal is to help the company build that capacity to adapt and keep improving: to accompany where it is needed and to get our hands dirty where it is needed.
Frequently asked questions
What is an AI centre of excellence in software development?
It is a small group that looks after how all of a company’s development teams work with AI: the guide to good practices, the shared context, safe use and training. Its job is to help the teams work better.
What is the difference between a centralised, distributed and federated model?
In the centralised one, a central group manages resources, tools and methods. In the distributed one, each team decides its own. In the federated one, a small core sets what is common and each team has a champion who brings it into daily work.
What is an enabling team?
It is one of the four team types in Team Topologies, by Matthew Skelton and Manuel Pais: it helps another team overcome obstacles and detects the capabilities it is missing.
Who is part of the practice?
One person in charge and a champion in each team. They are people from the company itself, with a guide to good practices, a shared context, criteria for safe use and a learning path.
How do you know the practice works?
With three indicators: how many teams work with the method, delivery time and errors reaching production. They are measured at the start and again with every team that joins, with no figures promised in advance.
How is the practice kept up to date?
With a four-step cycle: a team proposes a practice, tries it in its own team, if it works it goes into the guide and it is shared at the teams’ monthly meeting.
Sources
- Matthew Skelton and Manuel Pais, “Key concepts”, Team Topologies (section on enabling teams); no visible publication date, accessed on 17 November 2026.
- Pam Arlotto and Susan Irby, “Why healthcare organizations should consider implementing a data analytics center of excellence”, HFMA, 29 October 2020, updated 7 November 2022.
- Nathen Harvey and Derek DeBellis, “Announcing the 2025 DORA report: State of AI-assisted Software Development”, Google Cloud, 23 September 2025.
- DORA, “DORA’s software delivery performance metrics”, last updated 5 January 2026.

Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, automated checks and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.
LinkedIn →