AI Doesn't Fail in Isolation. Organizations Do.

 
Most AI Failures Are Not AI Failures

A regional bank in the northeastern US came to us with a familiar ambition. Move from a hierarchical structure to a project based operating model, faster decisions, less layered approval, technology and operations working as one team instead of two.

The technology side was never the hard part. The hard part was that nobody had touched decision rights. People kept reporting the way they always had, escalating the way they always had, getting evaluated the way they always had. The bank wanted agility without redesigning who owned what, and that gap is where the actual work began. We ended up redesigning the operating model for the technology and operations group, building a new talent platform and reward framework around it, because the structure had to change before any process inside it could.

This is the pattern we keep seeing, and it rarely gets named correctly.

Pilots succeed on a narrow, well defined task with a small group of engaged users. Deployment introduces something different, multiple teams, competing incentives, and decision rights that were never updated to match the new capability. The technology did not fail. The organization around it stayed exactly the same shape it was before the new capability arrived.

Escalation paths stay unclear. Risk ownership gets fragmented across teams that each assume someone else is holding it. Business units expect outcomes they were never structured to deliver. What looks like a technology problem, on closer inspection, is an operating model problem wearing a technology costume.

The hidden cost here is not accuracy or adoption rates. It is coordination.

Every intelligent actor we introduce into a decision system, human or machine, adds another accountability pathway that has to be defined, assigned, and maintained. Add an AI capability to a workflow, and we have added a new participant in a decision that used to involve two teams; now it involves three, plus a question nobody answered about who has final say. Unless governance, incentives, and operating structures evolve alongside the technology, organizations end up scaling complexity faster than they scale value.

This connects to something we explored more formally in recent research with Prof. Arpan Kumar Kar at IIT Delhi, published at BIGS 2025. The paper looks at how systems built on retrieval and reasoning, the architectures increasingly sitting underneath enterprise AI, introduce new actors into decision chains that traditional governance models were never designed to account for. One of the arguments we made there is that responsibility does not stay fixed when a system can retrieve, reason, and act somewhat independently of the people who built it; it has to be actively redistributed, the same way decision rights have to be redistributed when an organization moves from a hierarchical model to a more distributed one. Organizational design research has long shown the underlying mechanism, as decision authority becomes distributed, coordination costs tend to increase unless accountability structures evolve accordingly. AI does not break this pattern. It accelerates it, because the actors multiplying inside the decision system are no longer only human.

The lesson is simple, even if it is not the one most postmortems reach for. AI rarely fails in isolation. It fails when organizations attempt technological transformation without organizational transformation alongside it.

What is the most common non-technical reason you have seen AI initiatives struggle after a successful pilot?

Source: a regional bank in the northeastern US. The research referenced is "From Agentic AI to RAG: A Framework for Responsible AI," co-authored with Prof. Arpan Kumar Kar, IIT Delhi, published in the proceedings of BIGS 2025 and available on the AIS eLibrary.