Most AI Centers of Excellence operate as IT sandboxes or academic research labs. Funding a CoE provides the appearance of innovation without demanding operational change. An effective CoE is neither a department nor a project. It is a standing enterprise capability that identifies proven AI applications, builds the infrastructure to deploy them, and redesigns the workflows through which cognitive work actually happens.
Three Defaults
Without a cross-functional mandate tied to production outcomes, CoEs default to three dysfunctions:
- The IT Silo: Reporting into the CIO, the CoE treats AI deployment as software installation and cannot alter how work is performed.
- The Ivory Tower: Reporting into research, the CoE builds technically impressive models against problems the business does not have.
- The Hollow Bench: Reporting into short-term efficiency, the CoE automates the entry-level cognitive work — research, drafting, pattern recognition — through which junior practitioners historically built judgment. As Jeff Raikes recently observed, this accrues as talent debt. The bill comes due when the organization needs the senior experts the automated-away work used to produce.
The Operating Model
Three conditions are required:
- Literacy Over Competency: Competency is prompting, summarizing, running analyses. Literacy is the capacity to extend one's own reasoning with AI, interrogate its outputs, and recognize flawed analysis. A CoE designs workflows that require literacy — and preserve the cognitive reps through which judgment forms. This is the operational expression of Human-in-Command.
- Standardized Enablement: The CoE provides reusable architectures, integration patterns, and evaluation harnesses. Business units apply them; the CoE maintains the foundations.
- Embedded Expertise: CoE personnel sit alongside business unit operators. Systems built in a vacuum fail on handoff.
The Line Between Build and Govern
A CoE builds, scales, and redesigns. A governance board evaluates risk and sets boundaries. Collapsing these functions produces either a rubber stamp or a bottleneck. Executives stand up CoEs and governance boards together because the CoE needs defined boundaries to navigate, and the board needs a counterparty disciplined enough to navigate them.
Execution
Resolve one high-impact business problem end-to-end and document the process as a repeatable template. To make that possible:
- C-Suite Sponsorship: Changing how a company operates is political. The CoE requires a sponsor with the authority to resolve the disputes workflow redesign surfaces.
- Dedicated Capacity: Data scientists alone cannot run a CoE. The team requires ML engineers, human factors experts, and change management personnel.
- Clear Authority at the Seam: The CoE redesigns the workflow. Business unit leaders retain authority over their people, performance management, and pace of adoption. Both roles must be named at stand-up, or the seam becomes the fight.
The Standard Worth Holding
Evaluate a CoE with two questions. Are business units altering daily operations to use CoE-developed capabilities? Are practitioners using those capabilities developing the judgment to become senior experts — or being hollowed out by the systems meant to assist them?
"Yes" to both confirms a sustainable transformation engine. Ambiguity on the first means an IT sandbox. Ambiguity on the second means talent debt that comes due exactly when the organization most needs senior judgment to oversee what it has deployed.