As AI capabilities expand, organizations are scrambling to figure out how to organize around them. Should you have a central AI team? Embedded AI specialists in each business unit? A dedicated innovation lab? All of the above?

I’ve advised a dozen organizations on this question over the past two years. Here’s what I’ve learned about what works.

The Common Approaches

Organizations typically try one of these structures:

Centralized AI team. A single team serves the entire organization, taking on AI projects from various business units. The team owns AI expertise and technology choices.

Pros: Concentration of expertise, consistent standards, economies of scale. Cons: Can become bottleneck, may lack domain expertise, can feel disconnected from business reality.

Distributed/embedded model. AI specialists sit within business units, reporting to business leadership. Each unit builds its own capabilities.

Pros: Close to the business, deep domain understanding, clear accountability. Cons: Duplication of effort, inconsistent approaches, talent competition between units.

Hub and spoke. A central team provides platforms, standards, and expertise, while embedded specialists in business units focus on domain-specific applications.

Pros: Best of both worlds in theory. Cons: Coordination overhead, unclear ownership, can create conflict between center and edges.

Innovation lab. A separate unit focused on experimental AI projects, somewhat insulated from day-to-day operations.

Pros: Freedom to explore, attracts certain talent, high visibility. Cons: Often disconnected from production reality, projects don’t transfer to operations, can breed resentment.

What Actually Works

Based on successful examples I’ve seen:

Hub and spoke wins for most organizations. The pure centralized model creates bottlenecks. The pure distributed model creates chaos. Hub and spoke, done right, gets the benefits of both.

The key is clear division of responsibility:

• Hub: Platform and infrastructure, technical standards, specialized ML engineering capability, cross-cutting research
• Spokes: Business case identification, domain expertise, change management, end-user engagement

Innovation labs work only with clear paths to production. Labs that operate as isolated research units produce impressive demos that never ship. Labs that have explicit mechanisms for transferring successful projects to operations can be valuable.

Start with one high-value use case, not an organizational structure. Many organizations design the AI organization before they have a clear AI strategy. Better to start with a specific, valuable use case, build the team to deliver it, and let the structure evolve from there.

The Skills You Need

A capable AI innovation team needs these skills:

Machine learning engineering. People who can build, train, and deploy models. This is the technical core.

Data engineering. ML is only as good as the data. You need people who can build data pipelines and ensure data quality.

Software engineering. Production AI is software. You need people who understand production systems, DevOps, and building reliable software.

Domain expertise. AI specialists who don’t understand the business problem will build the wrong solutions. Either hire for domain expertise or ensure close collaboration with business experts.

Product thinking. Turning technical capabilities into products people actually use requires product sense. User research, interface design, change management.

Project/program management. AI projects have specific challenges - experimental nature, data dependencies, iteration requirements. You need people who can manage this complexity.

AI ethics and governance. As AI regulation increases and risks become clearer, you need people thinking about responsible deployment.

The common mistake: Over-indexing on ML research skills while under-investing in engineering, domain expertise, and product thinking.

Talent Realities

Honest talk about AI talent:

PhD researchers are expensive and often unnecessary. Most enterprise AI doesn’t require novel research. It requires capable engineering application of existing techniques. A strong ML engineer often delivers more value than a research scientist.

The best people are hard to hire. Top AI talent has abundant options. You’re competing with tech giants and well-funded startups. Be realistic about the talent tier you can attract.

Develop internal talent. Given hiring challenges, growing AI capability within existing teams is often more practical than external hiring. Data analysts and software engineers can upskill into AI roles.

Consider partners for specialized work. For specific capabilities you can’t build or hire, working with AI development companies in Sydney or other specialist partners may be more practical than building everything in-house.

Common Mistakes

Pitfalls I see repeatedly:

Building infrastructure before use cases. Organizations that start by building AI platforms without specific applications often build the wrong thing. Start with use cases; let infrastructure needs emerge from real projects.

Hiring too senior too early. Hiring a Chief AI Officer before you have AI projects to run is putting the cart before the horse. Start with practitioners; add leadership as the practice scales.

Isolating the AI team. AI teams that operate separately from the rest of the organization struggle to deploy anything useful. Embed AI capability close to business operations.

Underestimating change management. The technical work is often easier than getting the organization to adopt new AI-powered ways of working. Budget for change management accordingly.

Expecting immediate ROI. AI capability takes time to build. Expecting immediate returns pressures teams toward quick wins that don’t build lasting capability.

Ignoring MLOps. Building a model is maybe 20% of the work. Production deployment, monitoring, maintenance, and iteration are 80%. Staff and invest accordingly.

Practical Recommendations

If you’re building an AI innovation team:

1. Start with one or two high-value use cases that have executive sponsorship and clear success criteria.

2. Build a small team to deliver those use cases - maybe 3-5 people with a mix of ML engineering, software engineering, and domain expertise.

3. Focus on shipping. Get something into production, learn from it, iterate. Avoid extended research phases without deployment.

4. Add structure as you scale. Once you have successful deployments and demand for more, formalize the organizational model. Hub and spoke is usually the right answer.

5. Invest in platforms and standards only after you understand what you’re building on them. Let requirements emerge from real projects.

6. Plan for ongoing capability building - training existing staff, bringing in external expertise, building relationships with research communities.

Building AI capability is a multi-year journey. The organizations that approach it patiently and practically end up further ahead than those that try to leapfrog with aggressive hiring and infrastructure investment.

The goal isn’t to have an AI team. It’s to have AI capability that delivers business value. Keep that north star in focus as you make organizational decisions.