Three Departments, Three Different AI Vendors, One Very Confused CEO
A Hong Kong logistics group with around 400 employees recently held its first cross-departmental AI review. The Marketing team had bought a content tool. The Operations team had a separate routing tool. Finance was running an early invoice-extraction pilot. Three teams, three vendors, three sets of contracts, three completely different data-handling postures. None of them had spoken to each other before the meeting.
This is the moment in every enterprise's AI journey when a CEO realises the absence of structure has become a structural problem. The question is no longer whether AI works. The question is who owns it.
The answer most enterprises arrive at is the AI Center of Excellence. By 2026, according to Deloitte's applied AI research, this has become the dominant operating model for any organisation running more than two production AI use cases. This article explains what it is, what it does, and how a Hong Kong enterprise should actually build one.
What Is an AI Center of Excellence?
An AI Center of Excellence is a cross-functional team that sets the standards, owns the governance, manages the shared infrastructure, and helps business units scale AI safely. It is a hub: a small, senior group that defines how AI is done across the organisation, while business units remain responsible for what AI is done in their domain.
The CoE is not a delivery team. It does not build every model. It does not own every project. Its role is to ensure that every project meets the same quality, security, and governance bar, and that the lessons learned in one team become available to the others. This is the difference between an organisation that gets compounding returns from AI and one that pays the same learning costs three times.
Why Does Every Major Enterprise Need an AI CoE Now?
Every enterprise needs an AI CoE now because, without one, the cost of fragmentation begins to outweigh the benefits of speed. The trigger point is typically the third concurrent AI use case. According to Microsoft's 2026 Cloud Adoption Framework guidance, organisations that pass this threshold without a CoE start to see vendor duplication, contradictory data policies, and audit gaps within two quarters.
The numbers are stark. McKinsey's 2026 State of AI research reports that enterprises with a formal AI CoE achieve 5 to 10 times greater productivity gains by the second year of operation compared with those running ad-hoc programmes. The difference is not the technology. It is the reuse of evaluation methods, prompt libraries, security frameworks, and vendor relationships across the business.
For Hong Kong enterprises specifically, there is a regulatory dimension. The HKMA's GenA.I. Sandbox++, expanded in March 2026 across banking, securities, insurance, and MPF sectors, expects participating institutions to demonstrate centralised governance of AI deployments. A CoE is the operating-model answer to that requirement.
What Does the Hub-and-Spoke Model Actually Look Like?
The hub-and-spoke model is the structure most enterprises converge on by their second year of AI operations. The hub is a lean central team that owns standards, platforms, and governance. The spokes are business units that build and run AI use cases in their domain using the hub's shared assets. It is the operating model that works because it balances control with speed.
In a 2026 Deloitte review, organisations using this model reported 73% better scaling outcomes than those using purely centralised or purely decentralised approaches. The reason is structural. A purely centralised team becomes a bottleneck. A purely decentralised approach reinvents the wheel in every department. Hub-and-spoke gives the hub the leverage of standards and the spokes the speed of domain ownership.
In practice, the hub publishes approved patterns. The spokes consume them. When a spoke discovers something new, it flows back to the hub, which incorporates it into the shared standard. This is how the model creates compounding returns.
What Does the Hub Actually Own?
The hub owns five things. Get these right, and the spokes can run at speed. Get them wrong, and the entire AI programme becomes a series of expensive, disconnected experiments.
Standards and evaluation. The hub defines what good looks like. Accuracy thresholds, tone guidelines, allowed model providers, prompt patterns, escalation procedures. These are the rules every spoke must follow.
Shared platforms. The hub provides the common infrastructure: the observability layer, the vector store, the prompt library, the evaluation harness, the cost-monitoring dashboard. Spokes use this infrastructure rather than build their own.
Governance and risk. The hub owns the AI risk register, the data-handling rules, the vendor security reviews, and the connection to legal and compliance. It is the single point of contact for any AI-related audit question.
Vendor strategy. The hub negotiates enterprise contracts with model providers and observability vendors. This is where central scale beats decentralised purchasing.
Capability building. The hub runs the internal training programme, certifies prompt engineers, and maintains the playbook every spoke uses. According to Articsledge's 2026 CoE study, organisations that invest in capability building see use-case throughput double within one year.
What Do the Spokes Actually Do?
Spokes own use cases. Each spoke is a business unit, function, or product team that takes the hub's shared assets and applies them to its specific business problem. The Marketing spoke runs the content workflow. The Finance spoke runs the accounts-payable extractor. The Operations spoke runs the routing assistant.
Each spoke is accountable for the business outcome, the user adoption, the change management within its team, and the use-case-level evaluation. The hub does not do these things. If a spoke fails to drive adoption, that is a spoke problem, not a hub problem. This division of accountability is what makes the model scale.
The most important rule for spokes is that they cannot ignore the hub's standards. A spoke that decides to use a non-approved model, write its own evaluation rubric from scratch, or skip the prompt-library review is not innovating, it is creating future audit liability. CEOs who tolerate this dilute the entire CoE.
How Should a Mid-Sized Hong Kong Enterprise Staff an AI CoE?
A mid-sized Hong Kong enterprise should not start with a ten-person team. According to Tredence's 2026 CoE staffing benchmarks, the right starting size is four to six people for organisations of 200 to 800 employees, growing modestly as scale demands. The discipline is to staff the function before the headcount, not the other way around.
The four foundational roles are an AI CoE lead, an AI architect or technical lead, a governance and risk lead, and a delivery enablement lead who works directly with the spokes. A fifth role, the evaluation and observability owner, should be added by the time a third production use case launches. A sixth, a vendor and procurement lead, is added once enterprise contracts exceed five providers.
The most important hire is the CoE lead. This person needs to be senior enough to chair a vendor negotiation, technical enough to challenge an engineer, and politically credible enough to push back on a department head who wants to ignore the standards. Hong Kong enterprises often underestimate how senior this role needs to be.
What Are the Most Common Mistakes Enterprises Make When Building an AI CoE?
Three mistakes appear in nearly every failed CoE rollout. Each is preventable, but they all share the same root cause: under-investing in the governance and operating-model design before hiring begins.
The first mistake is treating the CoE as an IT function. When the CoE reports into the CIO with no business-side counterpart, business units treat it as a gatekeeper rather than an enabler. Microsoft's 2026 enterprise CoE guidance is explicit. The CoE must have a business sponsor at the C-suite level or it will be ignored.
The second mistake is mandating standards without providing tools. A CoE that publishes a governance policy without delivering an evaluation harness, a prompt library, and an observability dashboard has shifted compliance overhead onto the spokes. The result is a paper compliance regime that nobody follows in practice.
The third mistake is staffing the CoE entirely from outside the business. Pure consultants do not know which workflows matter, which managers will resist change, and which vendors have failed inside the organisation before. The 2026 EPCGroup CoE survey found that hybrid teams, with at least half the headcount drawn from existing business units, outperform pure-hire teams on time-to-first-production by 40%.
How Long Does a CoE Take to Pay Back?
A well-built AI CoE typically pays back within eighteen months in mid-sized Hong Kong enterprises and within twelve months in larger groups. The payback is not one big win. It is the cumulative effect of three smaller compounding benefits.
First is vendor consolidation. A single negotiated enterprise contract for a major model provider typically lands at 30 to 45% below the sum of independent departmental contracts, based on 2026 procurement benchmarks from KPMG's enterprise AI sourcing research.
Second is reuse. When the Operations spoke builds a document extraction pattern, the Finance spoke uses it for invoice processing six weeks later. Without the CoE, Finance would have rebuilt it from scratch and the costs would have appeared twice in the budget.
Third is risk avoidance. The CoE catches problems before they become customer incidents or audit findings. According to Gartner's 2026 governance research, organisations without a CoE are 3.5 times more likely to experience a publicly disclosed AI-related compliance event in the first two years of production deployment.
What Is the First Move This Quarter?
The first move is not to hire a CoE lead. It is to write a one-page operating model document that defines the hub's five responsibilities and the spokes' three obligations. This document forces the executive committee to commit to the structure before the headcount. If the document survives a board review without rewording, the CoE can begin hiring.
The document is also the artefact that determines whether the AI programme stays strategic or becomes a procurement exercise. Hong Kong enterprises that build the document first consistently report cleaner CoE launches than those that hire first and write the structure later.
From Fragmented Pilots to Compounding Returns
The shift to an AI CoE is the shift from running AI projects to running an AI capability. Without it, every department's AI investment is a separate experiment with its own learning curve, vendor relationship, and audit risk. With it, every department's investment accelerates every other department's investment.
The organisations that build their CoE in 2026 will have a structurally cheaper, safer, and faster AI programme by 2028 than those who wait. This is not a technology lead. It is an operating-model lead. And operating-model leads are very hard to close.
We understand the cold edges of AI and the hard parts of your work, and UD has walked with Hong Kong enterprises for twenty-eight years, making technology a partnership with warmth.
Take the Next Step With UD
Now that you have the framework, the next step is mapping your current AI footprint and designing the right hub-and-spoke structure for your organisation. We'll walk you through every step — from operating-model design to staffing, vendor strategy, and the first 90 days of CoE operations.
