AI Center of Excellence vs federated model vs enablement team: which to build
A central AI team that builds, a federated model with a small hub, or a small team that serves others. What each does well, where it breaks, and which to pick.
Short answer
An AI center of excellence keeps AI work in one central team. A federated model puts practitioners in each business unit under a small central hub. An enablement team is a small central group that ranks use cases, sets rules, trains and measures while others build. The duties in 62 postings we read look like the third. Start there.
On this page
- Models compared
- Center of excellence, federated (hub and spoke), enablement team
- Evidence
- 62 job postings we read, plus Microsoft, Rework and McKinsey
- Not covered
- Outcome data. We found no study comparing the three.
The verdict
- Pick a CoEif you are early, regulated or short of AI talentOne team gives one set of rules and one list of tools. Skip it once business units wait in its queue, which Microsoft treats as the signal to loosen control.
- Pick federatedif you have several business lines that can each staff an AI leadLocal teams move fast and know the work. Skip it if you cannot staff both a hub and the spokes, or if the hub has no authority.
- Pick an enablement teamif people already use AI and nobody counts or ranks the resultsA few people who choose, teach, set rules and measure. Skip it if nobody anywhere can build, because it does not.
Nolli makes software in this area. We compare from each vendor’s public pages and name our sources below. We did not run hands-on tests.
Key takeaways
- In the 55 postings we read for an employer’s own staff, 42 ask for a use case pipeline and 39 for measurement, while only 8 lead with building something.
- Microsoft and Rework both describe the same end state from opposite sides: a small central group that owns standards while teams own delivery.
- Pick by where the building should sit and who can say no, and expect to move: Rework warns that switching models later is an expensive reorganization.
Three names, one question: where does AI work sit in your company? A center of excellence (CoE) keeps it in one central team. A federated model moves most of it into business units and keeps a small hub for the rules. An enablement team keeps a small central group that does not build for others.
We expected to recommend one. The job postings pointed somewhere narrower. In the 55 postings for an employer’s own staff that we counted in our read of 62 AI enablement postings, the duties are mostly picking use cases, setting rules, teaching and counting. Building is the minority. That describes a small team that serves other teams, and it sits between the two famous models.
Three models, one sentence each
Center of excellence
A dedicated central team of AI specialists that sets strategy and standards and, in its classic form, delivers projects for the business units. Microsoft describes it as an internal team of experts that prevents fragmented or ungoverned AI adoption.
Federated model
Business units hold their own AI practitioners and run their own projects, while a small central hub owns the platform, the standards and the direction. Rework calls the same shape hub and spoke.
Enablement team
Our term. A small central group that chooses and ranks use cases, sets the rules, trains people and measures results, while teams or engineering do the building. No source defines it this way; it is what the postings below ask for.
Do not treat the names as fences. Microsoft says a CoE should grow into an advisory group that sets guardrails while frontline teams own delivery. Rework’s hub is a small team that owns standards without owning every project. Both end states have the shape of the enablement team. We read the three as points on one dial: how much of the building sits in the center.
What 62 job postings actually ask for
We read 62 public AI enablement and AI transformation postings on October 16, 2026. Fifty-five are for an employer’s own staff. Here is how many of those name each duty.
Twenty-nine of the 55 start their duty list with strategy, a roadmap or a portfolio. Eight start with building or delivering a solution. Two of the 62 titles say center of excellence.
Compare that with Microsoft’s CoE. Its list of duties overlaps the postings closely: strategy, skills, pilots, standards, an intake process and a ranked backlog, and measurement. The difference is staffing. Microsoft’s CoE has senior data scientists, machine learning engineers, AI security specialists and AI operations people. The postings mostly ask for someone who sets direction and counts, and several draw the line against delivery in writing. Johnson Controls calls its seat a capability-building and standards role rather than a delivery role. Spectris sends anything that needs production-grade architecture to its Enterprise Architect. At J.D. Power, as we read it earlier, a hub coordinates while champions deliver inside their functions.
The three side by side
| Center of excellence | Federated | Enablement team | |
|---|---|---|---|
| Where decisions sit | In the central team, which approves and prioritizes requests | Standards and platform at the hub; priorities with each business unit | Standards and the ranked list with the team; the call on each use case with the business owner |
| Who builds | The CoE, per Rework (5 to 15 people at mid-market scale) | Practitioners in the business units, on the hub’s platform | Business teams and engineering. The team may build small tools; 8 of 55 postings lead with it |
| Who measures | The CoE reports adoption and business impact to leadership (Microsoft) | Each spoke, unless the hub sets one method | The team, against a baseline. 39 of 55 postings ask for it |
| Speed to a team’s first result | Slow once the queue forms | Fast | Medium: it cannot ship for you, and it can remove blockers |
| Main risk | Bottleneck and distance from the work | Weak hub governance and dual reporting | Responsibility without authority |
| Who it suits | Early stage, regulated industries, little AI talent (Rework). McKinsey finds firms under $500M revenue likelier to centralize fully | Several business lines with staff to place in each. Rework says mid-market firms should not start here | Our reading: a company of 50 to 5,000 people where AI use exists but no one owns the list, the rules or the proof |
What each model does well
| Center of excellence | Federated | Enablement team | |
|---|---|---|---|
| Standards and governance | fully | partly | partly |
| Speed to a first result | none | fully | partly |
| Fit with local work | partly | fully | partly |
| Measurement across teams | fully | partly | fully |
| Control of tools and licenses | fully | partly | partly |
none partly fully · Our reading of Microsoft’s and Rework’s descriptions, McKinsey’s survey and the 62 postings. Full means the model does this well by design, half means partly or only with extra effort, empty means it is a known weak point. Judgments, not measurements.
Read the rows. The CoE wins on control and loses on speed, which is Rework’s whole critique: business units wait in a queue behind a team that lacks their context. The federated model wins on speed and fit and gives up easy control. Rework says weak hub governance produces fragmentation faster than none, because the hub gives false confidence that someone is coordinating. The enablement team is the only one that scores full on measurement without owning the build, and that is why we think it is the cheapest first step. It also leaves tool decisions mostly alone, since just 24 of 55 postings give the lead that job.
Central control against local fit
Our reading, from Microsoft’s and Rework’s descriptions and from the 62 postings. The positions are judgments, not measurements. “Embedded, no hub” is Rework’s name for specialists in each unit with no central function.
McKinsey’s survey shows companies already split the difference. It asked 1,491 participants in 101 nations in July 2024. For risk and compliance, and for data governance, organizations often use a fully centralized model such as a CoE. For tech talent and adoption of AI solutions, respondents most often report a hybrid or partially centralized model. Respondents at organizations under $500 million in revenue were more likely than others to centralize those fully. That supports a central core of rules with local delivery, and a smaller company leaning central.
One correction. Rework credits McKinsey with saying deployment works best through a federated hub-and-spoke structure. The McKinsey page we read reports what respondents say they do, in the words “hybrid or partially centralized”. It does not say that structure works best. The same page finds that CEO oversight of AI governance correlates most with self-reported bottom-line impact, among 25 attributes, with an R-squared of 0.20. Correlation, self-reported, and a weak fit. Microsoft says the same thing in another way: without executive backing, a CoE cannot enforce standards.
When each one fails
A CoE fails with a queue. Microsoft lists the signals: approval delays, knowledge bottlenecks because the experts cannot support every team, and product teams that argue priorities with the CoE instead of delivering. Rework adds the ivory tower, a CoE that builds technically sophisticated tools nobody uses because its engineers do not do the daily work. If you see both, the CoE is working against you.
A federated model fails at the seam. Each spoke lead answers to a business unit head for priorities and to the hub for standards, and Rework says organizations that design that line badly see high spoke turnover. It is also hard to staff, since you need platform and governance people for the hub and domain people with AI skill for every spoke. Many companies build the hub and then cannot hire the spokes.
An enablement team fails when it has the duties and no power. Several of the postings we read earlier expect the lead to influence people without authority over them, and Betts Recruiting says most placements are individual contributors with executive sponsorship. The second failure is drift. Johnson Controls writes into its posting that the seat will be pulled toward execution constantly and has to resist. A lead who becomes the company’s busiest automation engineer enables no one. We read ten postings for this and found five that wanted a builder and five that wanted a change or standards lead. Decide which you are hiring before you write the ad.
How a company moves from one to the next
- 01Policy and literacyRework: too early for any model. Approve tools, classify data, teach the basics.
- 02Small central teamA lightweight CoE or an enablement team. Rework gives 1 to 3 people at the pilot stage.
- 03Embed where the queue formsAdd practitioners in the busiest units. Keep the center for rules and the platform.
- 04Hub and spokesFormalize the federated model once AI sits in several core workflows.
The stages are Rework’s and its size thresholds are its own rules of thumb, shown without data. Microsoft describes the same move as a CoE going from central control to an advisory role.
The trigger matters more than the stage name. Microsoft says the move only works if you embed AI governance in platform operations first, so that standards travel to every team without a gatekeeper. Rework warns that the switching cost between models is high enough that the wrong choice now means an expensive and disruptive reorganization. We read that as a reason to start small and central, not as a reason to start with the final form.
Which one to build
Build a CoE if
you are regulated, have few AI-capable people and need one inventory of tools. Set a date to review it, and hand delivery to the teams once approvals slow down.
Go federated if
you have several business lines, an executive who will back the hub, and people who can be placed in each. Write the dual reporting down before you hire the first spoke.
Start with an enablement team if
AI is already in use, you have no ranked list, no single set of rules and no baseline. Our pick for most companies in the 50 to 5,000 range. Give it a sponsor with real authority, and keep it away from production work.
Build none of them yet if
you have no real deployments. Rework’s advice for that stage is policy and basic training, and we agree.
We found no study that compares results across the three models. The sources are vendor guidance, one survey of what companies do, and job ads. Treat our pick as a reasoned bet on that evidence.
Where Nolli fits
Where Nolli fits, and where it does not
Nolli sells to the team in the third column, so weigh this post with that in mind. Nolli maps how each team really works from the work itself, ranks use cases by hours and cost, and measures the same workflow done by people, by people with AI and by agents, on time, quality and cost per task. That covers two of the five duty areas above, the use case list and the proof, and it works the same under any of the three models. It does not choose your model, write your AI policy, run training or decide who the team reports to, and your company decides data and policy questions. No public price. Request a demo. See Nolli for AI enablement teams and how to measure AI impact.
Request a demoQuestions people ask
What is the difference between an AI center of excellence and a federated model?
A CoE keeps AI specialists and delivery in one central team. A federated model places AI practitioners in each business unit and keeps a small central hub for the platform and standards. Rework describes the hub as typically 3 to 8 people.
Is an AI enablement team the same as a center of excellence?
No. A CoE in its classic form builds and delivers for the business. The enablement team in the job postings we read is asked mostly to rank use cases, set rules, train and measure. Microsoft says a CoE should grow into that advisory role as AI adoption matures.
Which AI operating model is best for a mid-sized company?
No study we found compares them. Rework says mid-market companies should not start with a federated model, and McKinsey found companies under $500 million in revenue more likely to centralize talent and adoption fully. Our reading is to start with a small central team and add embedded people where a queue forms.
When should a CoE become a federated or advisory model?
Microsoft names the signals: approval delays, knowledge bottlenecks where experts cannot support every team, and growing friction between product teams and the CoE over priorities.
How big should an AI center of excellence be?
Rework gives 5 to 15 people at mid-market scale, and 1 to 3 for a lightweight CoE at the pilot stage. It gives no data behind the figures, so treat them as a starting guess.
Sources
- “Establish an AI Center of Excellence (Cloud Adoption Framework; vendor-written).” Microsoft Learn, accessed October 19, 2026.
- Victor Hoang. “AI CoE vs. Embedded vs. Federated: Choosing the Right Organizational Model (vendor-written).” Rework, updated May 19, 2026.
- Alex Singla, Alexander Sukharevsky, Lareina Yee, Michael Chui and Bryce Hall. “The state of AI: How organizations are rewiring to capture value.” McKinsey & Company (QuantumBlack), March 12, 2025.
- “AI Enablement Leader (job posting).” Johnson Controls, accessed October 15, 2026.
- “AI Enablement Lead (job posting, Particle Measuring Systems).” Spectris, via Workday, published August 24, 2026.
- “Head of AI Enablement (job posting R-100190; expired).” J.D. Power / Autodata, via London Tech Jobs, posted June 30, 2026.
- The Betts Team. “What is a Head of AI Enablement?” Betts Recruiting, July 20, 2026.
Plain-text copy: /blog/ai-center-of-excellence-vs-federated-vs-enablement-team.md