A vendor of ours renamed itself a few years ago. A direction to look out for the old name and replace it with the new one has been a staple in all of my CLAUDE.md files of every project that touches this vendor for months. In my agentic workflows, it's been handled for me. But today, legal sent me a draft for a brand new legal document. And there it was: The old vendor name, staring me right in the face. That's when it hit me: We might have an Agentic Silos Problem (ASP, my word choice).
Every company-wide AI rollout eventually collides with the same wall. Two departments have two versions of the same data. The AI plugged into finance says one thing. The AI plugged into ops says another. Both are confident. Both are cited. The exec looking at the two answers has less signal than they did with the two humans.
The failure is not in the model. The failure is that a silo is not a technical boundary. It is a boundary between two definitions of the same word - customer, revenue, signups, new bookings, churn, active users - that a human org papered over with meetings and adult behavior. The model does not do meetings. It does not soften. It takes the definition its data was labeled with and defends it, at machine speed, with plausible reasoning.
The agentic output generated within multiple silos looks like disagreement. It is actually the org chart being read aloud. Every unresolved political question about ownership of a metric now becomes a runtime conflict.
You cannot fix ASP with a better prompt. You fix it upstream, by naming the definitions that were never named. I think it's a great start to introduce a company-wide GLOSSARY.md and make it mandatory to use. And you should make sure that people read the GLOSSARY.md and suggest improvements. The free agentic feedback tool ritualdissent.com is a very helpful tool for exactly that at our company.
Thoughts? Find me on Bluesky.