AI productivity gains can become delivery commitments before a team has mapped what the workflow depends on. Switching to another model may still leave missing permissions, connectors, evidence and workflow state.

This practitioner talk works through a sample agentic research workflow to distinguish adopting an AI tool from changing the operating model around it. We map the essential service, identify failure paths, define a usable degraded mode and design a controlled recovery exercise. The focus is on ownership, observability, human handoff and the evidence needed to decide whether a fallback actually works.

Attendees leave with a dependency checklist, a recovery-drill structure and concrete measures such as time to restore an essential task and how errors or backlogs are contained. The talk distinguishes proposed exercises from completed tests and discusses limitations openly.

Audience: engineering, DevOps and technology leaders. Presentation option: 25-minute talk plus 5-minute Q&A. Level: intermediate. The recovery exercise is a proposed method, with completed implementation work and hypothetical scenarios distinguished clearly.