Two Agentic AI Case Studies – One Success, One Reversal And What Separates Them
I have spent this year documenting agentic AI wins. This edition is different. I want to put a genuine success story next to a genuine reversal, side by side, because the gap between them is not about the technology. It is about one missing stage in the roadmap.
What Went Right
A large U.S. health system rolled out a HIPAA-ready agentic AI platform for prior authorization, one of the most administratively burdened workflows in medicine. The agent connects to coverage databases and clinical records, checks a request against payer policy, assembles supporting evidence, and proposes a determination. In one pilot, it worked through more than 1,400 pages of clinical notes. The large majority of clinicians using it reported meaningful time savings with no drop in accuracy.
What made this work was not the model. It was the architecture around the model. Every proposed determination routes through a human before anything reaches a payer. The system draws from authoritative data connectors rather than free-form generation. On my Five-Stage Roadmap, I place this deployment at Stage 3, Governed, moving toward Stage 4, Assured. The health system is running this at real volume with outcomes now public, which is exactly the evidence base assurance requires before autonomy expands further. Full case study here: Agentic AI in Healthcare: The Prior Authorization Case Study.
What Went Wrong
A European fintech company pushed the opposite direction. It moved straight to a claimed Stage 5, Autonomous, and skipped Assured entirely. The headline metric looked strong, hundreds of thousands of conversations handled in month one. But the agent could not reliably tell which conversations belonged to it and which needed a human. Disputes, hardship cases, and anything involving judgment about money got the same confident handling as a routine order-status question.
The company’s own postmortems pointed to weak escalation handoffs. When a case did route to a human, the context did not travel with it, so the customer had to start over. That is an architecture gap, not a model capability gap. AI-driven hiring freeze had cut headcount significantly before the company reversed course and began recruiting again. Full breakdown here: Agentic AI’s Autonomy Trap – A Case Study.
Put the two side by side and the same five questions split them cleanly.
Roadmap stage claimed versus stage evidenced:The health system claimed Governed moving to Assured, and its own numbers back that claim. The fintech claimed Autonomous, but did not have enough error containment that Assured, let alone Autonomous, requires.
Human-in-the-loop placement: In healthcare, the human sits before the determination reaches the payer, on every case, with no exception carved out. In the fintech, the human sits after the agent already decided the case did not need one. The first design catches errors before they leave the building. The second depends on the agent correctly self-identifying its own blind spots, which is precisely the judgment call it was weakest at.
Escalation design: The health system’s workflow was built around escalation from day one, since every case escalates by design. The fintech treated escalation as an edge case bolted on after the fact, which is why context did not travel with the handoff and customers had to restart.
Evidence base before scale: The health system published outcomes from a bounded pilot before expanding. The fintech scaled to hundreds of thousands of conversations in month one, before it had evidence its edge-case handling worked at all.
What broke when it broke: In healthcare, the failure mode is bounded, a determination gets flagged for human review. In the fintech, the failure mode compounded, since a bad handoff on a dispute or hardship case is also the case most likely to escalate into reputational and regulatory exposure.
The Key Learning
Both companies automated a real workflow. Both had scale. The difference is that one built the audit trail, the risk-tiered escalation, and the human accountability loop before claiming the next stage, and the other announced the destination before doing the work to get there.
If you are evaluating a customer-facing or judgment-involving agentic AI deployment, ask this before go-live, not after: what happens to the cases my agent cannot handle well and does the human handoff carry full context, or force the person to start over. Get that answer before you claim autonomy.
For the full framework behind Governed, Assured, and Autonomous, see The Agentic AI Roadmap: 5 Levels From Prompted to Autonomous.
© Dr. Harish Kotadia, Ph.D. All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.
Disclaimer: This blog post is based on recent events and news items drawn from reputed media sources available in the public domain and linked above. This post is intended for educational purposes, to help the enterprise agentic AI community learn from public information on the application and use of agentic AI tools and technology in Fortune 500 companies. Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of published news reports quoted above and does not constitute legal, financial, or consulting advice of any kind.