If 2025 was the year of expensive lessons, 2026 is about who actually learned them. Discover the forces reshaping enterprise AI and what data leaders must get right to move from pilots to production.
"Agents amplify everything" is the sentence that should be on every data team's wall right now. One specific failure mode I'd add to the list: entity fragmentation. Agents don't just retrieve data, they act on entities — customers, accounts, products. When two records referring to the same customer haven't been resolved, an agent might send the same suppression twice, or trigger a high-value offer for a churned account because it's registered under a different ID. Entity resolution is one of the last pieces of data infrastructure that still runs in batch in most stacks, which means the "trusted entity" an agent is acting on could be a stale snapshot. The fix isn't hard — incremental ER on the modern data stack is solvable — but it has to be in the architecture before agents go to production, not after.
"95% of AI pilots never made it to production" is the statistic that should be in every AI strategy deck. The cause is almost never the model.
The framing of metadata as "active AI infrastructure" is important. I'd add a specific mechanism: metadata that isn't versioned, owned, and connected to a change process isn't active infrastructure it's documentation in disguise. The distinction matters because agents will act on stale metadata with the same confidence they act on current metadata.
The forcing function for getting this right is the gap between LLM accuracy on raw warehouse schemas (~20% on real enterprise data) versus accuracy with a governed semantic layer (~83-95%). That gap isn't closed by a better model. It's closed by better governance. I wrote about the specific failure pattern "semantic drift" and how OSI v1.0 addresses the exchange side of the problem (while governance addresses the ownership side): https://www.nilayshah.me/semantic-layer-ai-analytics-osi/
The organizations that learn from 2025 are the ones that treat this as infrastructure work, not a best practice.
The '95% of AI pilots never reached production' statistic validates what I've observed: the bottleneck is almost never the model capability.
Data readiness, governance, metadata quality - these aren't sexy topics but they're where implementations actually fail. I spent more time on data plumbing for my autonomous agent than on prompt engineering.
Your framing of metadata as 'active infrastructure for autonomous systems' is exactly right. Static documentation doesn't cut it when agents need to make real-time decisions about what data to trust.
This reads like an accurate autopsy of why so many 2025 pilots collapsed. The common thread isn’t model weakness, it’s that foundations built for analytics were pushed into systems that retrieve, infer, and act.
One pressure point that shows up repeatedly is the transition from context-aware reasoning to execution. Metadata and semantics help agents decide what to do, but governance thins out at the moment decisions turn into side effects. Architecturally, that handoff is already formalised as an execution control layer.
Which reinforces your core point: scaling AI isn’t about better models, it’s about treating context, governance, and execution as infrastructure rather than assumptions inherited from the analytics era.
"Agents amplify everything" is the sentence that should be on every data team's wall right now. One specific failure mode I'd add to the list: entity fragmentation. Agents don't just retrieve data, they act on entities — customers, accounts, products. When two records referring to the same customer haven't been resolved, an agent might send the same suppression twice, or trigger a high-value offer for a churned account because it's registered under a different ID. Entity resolution is one of the last pieces of data infrastructure that still runs in batch in most stacks, which means the "trusted entity" an agent is acting on could be a stale snapshot. The fix isn't hard — incremental ER on the modern data stack is solvable — but it has to be in the architecture before agents go to production, not after.
"95% of AI pilots never made it to production" is the statistic that should be in every AI strategy deck. The cause is almost never the model.
The framing of metadata as "active AI infrastructure" is important. I'd add a specific mechanism: metadata that isn't versioned, owned, and connected to a change process isn't active infrastructure it's documentation in disguise. The distinction matters because agents will act on stale metadata with the same confidence they act on current metadata.
The forcing function for getting this right is the gap between LLM accuracy on raw warehouse schemas (~20% on real enterprise data) versus accuracy with a governed semantic layer (~83-95%). That gap isn't closed by a better model. It's closed by better governance. I wrote about the specific failure pattern "semantic drift" and how OSI v1.0 addresses the exchange side of the problem (while governance addresses the ownership side): https://www.nilayshah.me/semantic-layer-ai-analytics-osi/
The organizations that learn from 2025 are the ones that treat this as infrastructure work, not a best practice.
The '95% of AI pilots never reached production' statistic validates what I've observed: the bottleneck is almost never the model capability.
Data readiness, governance, metadata quality - these aren't sexy topics but they're where implementations actually fail. I spent more time on data plumbing for my autonomous agent than on prompt engineering.
Your framing of metadata as 'active infrastructure for autonomous systems' is exactly right. Static documentation doesn't cut it when agents need to make real-time decisions about what data to trust.
This reads like an accurate autopsy of why so many 2025 pilots collapsed. The common thread isn’t model weakness, it’s that foundations built for analytics were pushed into systems that retrieve, infer, and act.
One pressure point that shows up repeatedly is the transition from context-aware reasoning to execution. Metadata and semantics help agents decide what to do, but governance thins out at the moment decisions turn into side effects. Architecturally, that handoff is already formalised as an execution control layer.
Which reinforces your core point: scaling AI isn’t about better models, it’s about treating context, governance, and execution as infrastructure rather than assumptions inherited from the analytics era.
thanks for sharing my article!
Insightful read 🙌 🙌
🙌 🙌 🙌