Building Cross-Functional AI Teams That Actually Work
The AI team that sits in isolation builds models that the business never uses. The AI team that is embedded in the business but lacks technical depth builds dashboards that never become decisions. The structure that works is neither of these.
Prince Kumar
Author
Most enterprise AI teams are structured in one of two ways: a centralised AI centre of excellence that owns all AI development and deploys solutions to business units, or distributed data scientists embedded within individual business units who own their own models. Both structures fail, for predictable and well-documented reasons. The centralised model builds technically excellent solutions to problems the business did not actually have. The distributed model builds solutions that cannot be maintained, cannot scale, and cannot share learning across units. The structure that works is a deliberate hybrid that most organisations have not yet built.
Why the Standard Structures Fail
The centralised AI CoE fails because it is structurally disconnected from the operational context that determines whether a solution is actually useful. A model that achieves 94% accuracy in a test environment but requires data that is only available with a two-day lag in the production environment is not a useful model. The CoE discovers this problem late because it has no visibility into the operational reality of the business unit it is serving.The distributed model fails because data scientists embedded in individual business units lack the infrastructure, tooling, and peer review that produces reliable, maintainable work. They also cannot leverage learning from adjacent units the demand forecasting insight from the retail team is not accessible to the supply chain team because the two are organisationally separate and technically siloed.
The Hybrid Structure
The structure that works has three components: a central platform team that owns infrastructure, tooling, data pipelines, model deployment, and monitoring; domain AI leads who are technically strong but embedded within specific business units and report into both the platform team and the business unit leadership; and business translators in each unit who own the problem definition, success criteria, and change management for every AI initiative.The platform team ensures consistency, reusability, and maintainability. The domain leads ensure technical solutions are grounded in operational reality. The business translators ensure AI initiatives are solving problems that matter and that the organisation is ready to act on the outputs. Remove any one of the three, and the structure degrades into one of the standard failure modes.
Related articles
View all →
Enterprise AIEnterprise AI Is Broken: Why It's Not Delivering ROI
Billions have been spent on AI across enterprises worldwide, yet manual follow-ups still dominate daily operations. Dashboards multiply, pilots launch, and most quietly die. The missing layer is not intelligence it is execution.
Enterprise AIThe Core Problem: AI Gives Insights, But Enterprises Need Execution
LLMs can summarise, forecast, and flag anomalies with impressive accuracy. What they cannot do is close the loop. The gap between an AI-generated insight and a resolved enterprise workflow is where most AI value is lost and it is a structural gap, not a technology gap.
AI ProductivityThe AI Productivity Paradox (2026): More AI, Slower Decisions
Every new AI dashboard added to an enterprise increases cognitive load, not productivity. The paradox of 2026: AI tools reduce tactical work while multiplying the strategic coordination overhead required to act on what they surface. The fix is not more assistants it is autonomous executors.