Amy Mitchell: Why AI initiatives break normal product manager instincts

Amy Mitchell is a product management practitioner who writes about AI transformation at the team level. This July 2026 post grew out of her direct involvement in a project to prioritize which PM workflows were suitable for AI agents, starting with PRD writing.

The unexpected expansion of scope

Mitchell describes being given a narrow, well-scoped task: evaluate how AI agents could take over PRD creation. In practice, the scope expanded immediately. The work required making decisions about what constituted business value in the context of agent output, who owned governance when the agent’s PRD diverged from stakeholder expectations, and how to measure whether the agent was actually doing something useful. What started as a workflow design question turned into something closer to product ownership of the transformation itself.

This expansion, she argues, is not a scoping failure. It is structural. AI initiatives are systems changes rather than feature additions, and they surface questions that conventional PM frameworks were not designed to handle. Trying to scope tightly around a single workflow without addressing these upstream questions means they reappear at every subsequent decision point.

A different operating mode

The core argument is that AI initiatives require a fundamentally different operating mode from standard product management. In traditional PM work, the default orientation is operational stability: ship reliably, reduce risk, and iterate on user feedback within defined parameters. In an AI transformation context, the right default is learning. Teams that approach AI initiatives with an operational-stability mindset tend to over-specify early, invest in solutions before the problem is understood, and measure success in ways that obscure whether the AI system is producing genuine value.

Mitchell’s proposed alternative centers on controlled experiments: run small-scale trials, treat each one as a learning event, and make incremental decisions about where to expand based on what those trials reveal. This is not the same as treating the initiative as a research project — the aim is still production deployment — but the sequence matters. Learning precedes scaling.

Finding the seam in existing processes

The practical guidance stays close to the PRD example. Rather than redesigning the entire document-writing process around an AI agent, identify one specific workflow within that process — for instance, the first draft of a problem statement — and test whether an agent can generate something useful in that narrow scope. Mitchell describes this as finding the seam in an existing process rather than replacing the process wholesale.

The post is most useful for product managers being pulled into an AI transformation project for the first time and finding that their usual instincts are not producing results. The PRD framing makes the argument concrete, but the underlying logic applies to most PM workflows an organization might want to automate: start with one seam, learn from it, and expand from there.