The CoE design conversation usually starts with a centralized model: one team owns the platform, all changes route through them, governance is uniform. It’s appealing because it’s simple to draw on a slide. It also stops scaling around 200 platform users.
Why centralized CoEs hit a ceiling
A centralized CoE is a queue. Every demand from every business unit flows through one decision-making bottleneck. The queue grows linearly with the platform user base; the team grows linearly with hiring budget. Eventually demand exceeds throughput, the queue gets longer, and the business stops trusting the CoE to deliver. Innovation moves out of the platform team to whoever can deliver fastest — usually a shadow IT alternative.
The federated alternative
A small central team (5-12 people) owns: platform architecture, governance standards, CMDB hygiene, security model, and CoE-wide releases. Embedded delivery squads (3-5 people each, one per major business unit) own: their unit’s use cases, day-to-day delivery, business-unit relationships.
The split: 30% central, 70% embedded.
How governance works without becoming a bottleneck
The central team publishes standards as code (linting rules, automated policy checks, reference patterns). Squads operate within those standards autonomously. Approval-tier escalates to central only for: cross-domain integration, security exceptions, architectural deviations.
For a tier-1 bank we work with, this model brought average decision time down from 4 weeks (centralized) to 5 days (federated) without measurably increasing risk-event frequency.
Capacity sizing — a starting heuristic
- Central CoE: 1 platform engineer per 500 active users
- Embedded squad: 1 squad per major business domain (HR, IT-Service, ITOM, Customer-Service)
- Squad sizing: 4-6 people per 200 active users in that domain
These ratios produce sustainable work-in-progress. Adjust based on demand profile and the maturity of automation in each domain.