Build a useful internal service
AI Centre of Excellence Setup
Establish an AI centre of excellence around recurring business demand, a clear service catalogue, real staff capacity and evidence of value to delivery teams.
An AI centre of excellence should solve recurring problems that several teams cannot efficiently solve alone. Its value may come from a reusable evaluation method, a reliable route to specialist advice or a maintained body of knowledge. A new name, steering group or central inbox does not create that capability.
Begin with demand. Identify the teams that would use the centre, the work they repeatedly struggle with and the outputs they would accept as useful. If the business has one small initiative, a named owner and occasional specialist support may be enough. Establish a centre when the shared need and capacity justify an enduring service.
Write a charter with an internal customer
The charter should define the purpose, service customers, initial catalogue, resources, boundaries and review conditions. A useful illustrative charter could say: “The centre will help participating business teams frame AI opportunities and produce reusable evaluation evidence. It will maintain agreed methods and provide specialist support within published capacity. Business owners retain responsibility for their outcomes and release decisions under the organisation’s operating model.”
This wording makes the centre’s role concrete without giving it every AI responsibility. The AI operating model should settle enduring decision rights. An AI council or steering committee addresses oversight and escalation. A service team and a decision forum can work together, but they should not be treated as the same thing.
State what the centre will not provide initially. It may not have capacity for production support, legal opinions, every business-unit build or independent audit. Publishing those boundaries is useful service design. It prevents colleagues from assuming that submitting a request transfers their operating responsibility to the centre.
Build the first service catalogue from repeated demand
The US General Services Administration’s AI CoE service catalogue is a public example of describing a centre through services and deliverables. Its government context is different from a Saudi company. The catalogue below is an original starting design to adapt to demonstrated internal needs.
| Service | What the requesting team supplies | Usable output |
|---|---|---|
| Opportunity framing clinic | Process owner, observed problem and redacted cases | A bounded opportunity card and the next evidence question |
| Evaluation design support | Proposed task, users, failure cases and domain reviewer | Agreed evaluation approach and evidence requirements |
| Source-readiness coordination | Named information owner and proposed sources | Documented readiness questions and remediation ownership |
| Reusable method library | A reviewed method and its limitations from completed work | A maintained version with instructions and appropriate reuse conditions |
| Supplier evidence review | Buyer-owned brief and comparable proposal material | Questions, evidence gaps and a recorded review within agreed expertise |
| Role-based coaching | A practical task and staff with time to learn | Demonstrated ability to perform the task, with remaining support needs |
For each service, also record its owner, intake conditions, typical effort assumptions, prioritisation method and completion test. These are internal service commitments to agree, not universal service-level promises. A catalogue that says “support AI innovation” without defining an output cannot guide either demand or staffing.
Use use-case discovery for the opportunity service and the vendor scorecard for a structured supplier review. Keep the centre’s role proportional to its expertise. A review that identifies missing security evidence should route the question to the appropriate specialist rather than imply the centre has certified the supplier.
Test the catalogue with a real capacity model
Consider an illustrative multi-brand services group. Several teams repeatedly ask for help defining whether an AI trial is useful. The sponsor can allocate 20 hours a week of actual shared specialist capacity to evaluation support. This is a planning example, not a recommended staffing level.
If five new requests each need an assumed six hours, demand is 5 × 6 = 30 hours. The gap is 10 hours before accounting for older work. Publishing a one-week response promise would not create the missing capacity. The centre must reduce scope, improve reuse, schedule requests or obtain more capacity. It should also distinguish an initial acknowledgement from completion of the analytical work.
Measure actual effort after the first requests. An apparently simple review may depend on a domain owner who is unavailable. Record elapsed waiting separately from specialist effort so the team does not diagnose every delay as a staffing shortage. Also reserve time to maintain methods and learn from completed work; otherwise each request consumes the capacity intended to make the next one easier.
Staff for services, with explicit time allocation
Start from the catalogue and identify the capabilities each service requires. Some can be supplied by a small core team, others through named contributors from architecture, data, security, procurement or business operations. A virtual centre is possible if those contributors have agreed time and responsibilities. A list of volunteers without allocated capacity is a network, not yet a dependable service.
Name an accountable service lead and a counterpart in each requesting business team. The lead coordinates the service and capacity; the business counterpart provides context and accepts the output. Do not use the centre to remove process ownership from the team that must adopt the result.
External specialists can help establish methods and coach staff, but define what remains inside the organisation. The capability-building guide helps specify practical learning outcomes. A centre that cannot maintain its own catalogue or explain its methods after the founding supplier leaves has an unresolved operating dependency.
Use intake to route work, not to create a new barrier
The intake should ask for the business problem, owner, requested service, relevant evidence and timing reason. It should be short enough for a colleague to complete without hiring another consultant. A request may be accepted, returned for clarification, routed to an existing function or deferred with an explanation.
Publish prioritisation principles. For example, urgent questions affecting an already authorised release may need a different route from speculative ideas. Shared methods that unblock several teams may be valuable, but they still need an identified user and evidence of reuse. Do not let seniority alone determine the queue or disguise an unstaffed service as a pending approval.
Keep service routing separate from investment or risk approval. If the centre identifies a decision beyond its remit, refer it to the authorised body using the steering committee operating model. The request should not become an indefinite handoff between two inboxes.
Establish the centre through evidence milestones
- Confirm recurring demand. Interview likely internal customers and inspect recent cases where shared help was missing. Select the smallest useful catalogue.
- Launch with a capacity boundary. Assign real staff time, agree inputs and outputs, and test the service on a limited set of requests.
- Check transfer and reuse. Ask a second team to use a method without its original author explaining every detail. Record the adjustments needed.
- Review whether to expand. Add services only when demand, competence and capacity support them. Remove offerings that duplicate existing functions or do not produce usable outcomes.
This sequence is not tied to a fixed number of weeks. Progress depends on evidence from the services. A charter signed quickly is not proof that the centre is established; a small service used reliably by several teams is more informative.
Measure usefulness to internal customers
Track the time from a complete request to a usable answer, the effort required from the requesting team and whether the output enabled its next decision. Examine repeated questions, recurring defects and the conditions under which methods are reused. Count completed requests only alongside their value; closing many low-value tickets can conceal an ineffective service.
Review the broader capability profile with the enterprise maturity model. The centre is successful when teams gain reliable capability, not when every activity must pass through it. Evidence may support expanding shared services, delegating a mature practice or stopping a service that does not help.
Open for AI’s agentic AI consulting can help examine use-case demand, readiness and a sequenced establishment plan. Discuss an AI centre of excellence setup with your proposed internal customers, recurring requests and available capacity. Agree the charter, catalogue and capability-transfer outputs required before commissioning the work.
FAQ
Your questions, answered.
Does every company need an AI centre of excellence?
No. A centre is useful when recurring shared needs justify an enduring service and the organisation can allocate real capacity. One small initiative may need only a process owner and targeted specialist help. Test the demand before creating a structure. The centre should make work easier for its internal customers rather than become another organisational requirement without a clear service.
Can an AI centre of excellence be virtual or part-time?
Yes, if contributors have named responsibilities, agreed availability and a service lead who manages demand. Record the time allocated rather than assuming people will help around their existing jobs. A virtual model can provide a practical starting point, but its commitments must fit its capacity. If requests consistently exceed that capacity, change the catalogue, priorities or resourcing.
How is an AI CoE different from an AI steering committee?
The centre provides defined services and practical capability to teams. A steering committee makes or escalates decisions within its mandate, such as priorities and material exceptions. Their responsibilities can interact, but committee attendance does not deliver an evaluation method or maintain a knowledge service. Keep the service catalogue separate from the committee’s decision rights and reporting requirements.
Should the centre charge business units for its services?
Choose a funding approach that supports the behaviour you want and can be administered proportionately. Central funding may support shared foundations, while clearly attributable specialist work may justify a different allocation. Avoid a charging model that discourages early questions or hides the cost of maintaining reusable methods. Start with visibility of demand and effort before adding complex internal billing.
Can an external consultancy establish the centre for us?
It can help design the charter, service catalogue, methods and initial coaching if those deliverables are agreed. The organisation still needs internal owners, allocated capacity and acceptance of the ongoing responsibilities. Test whether staff can maintain and use the methods after the engagement. A launch presentation and a set of templates do not by themselves establish an operating centre.
What indicates that the AI centre of excellence has become a bottleneck?
Look for requests waiting without an identified decision, repeated reviews that add no new evidence and teams unable to use methods without central intervention. Separate a capacity problem from unclear inputs or responsibilities. Improve the service or delegate a mature practice where appropriate. The goal is reliable organisational capability, so a reduction in unnecessary dependence on the centre can be a positive result.