Outsourcing My PM Work to AI: What Remains Human After 15 Years

After 15 years in product management, I've been experimenting with WorkBuddy — an AI project assistant — for the past few months. The experience has fundamentally reshaped how I view the product manager role, which I break down into three core activities: producing proposals (PRDs), reviewing designs, and making final decisions.

Producing Proposals: From Writing for Humans to Writing Specs for AI

Traditionally, writing a detailed B-end PRD took me hours: outlining, filling details, checking logic. Now I feed WorkBuddy rough requirements, competitor links, and flow diagrams with a structured prompt: "As a senior PM, output functional specs, exception flows, and data tracking needs per standard PRD structure." Within five minutes I get an 80%-complete draft. My remaining 20% is reviewing logic and adding business-specific nuances only I know. Efficiency has more than tripled.

The deeper shift: my output is moving from "documents for humans" to "specifications for AI." A raw prompt yields generic templates. To get usable drafts I did two things: (1) fed it past high-quality PRDs as style examples, and (2) locked the output structure into hard constraints — background first, then solution, no free-form deviation. This turned a generic chatbot into a context-aware junior product assistant. It exemplifies the industry shift from prompt engineering to context engineering: the skill is no longer writing perfect documents, but specifying problems precisely enough for AI to execute accurately.

Reviewing Designs: Red-Teaming at the Design Stage

Beyond drafting, I use WorkBuddy as a "picky user" and "dominant stakeholder" to attack my own proposals. I instruct it: "You are an extremely demanding user and a powerful business owner; tear this proposal apart." It uncovers gaps I missed — unhandled exception flows, permission conflicts, user paths that deviate from my design.

This works because humans have blind spots for their own work; colleagues pull punches in reviews. An emotionless AI explicitly told to "go for the jugular" surfaces production bugs before they ship. This is essentially red-team testing — standard for AI product validation — moved upstream into the design phase. The first gate of design review is shifting from people to AI.

Making Decisions: Execution Commoditized, Judgment Appreciated

Data analysis illustrates the new boundary. I drop messy Excel files (thousands of rows) into WorkBuddy with: "Analyze field distribution, find outliers, summarize top 5 trends in a Markdown table." Minutes later I have results with preliminary insights — previously an hour of Python scripting.

But data is where hallucination is unacceptable. I enforce three hard guardrails: (1) admit ignorance rather than fabricate numbers; (2) force conclusion-first reporting; (3) every key claim must cite its source row for instant traceability. The more I delegate execution, the more critical these guardrails become — and deciding which guardrails to set, where human sign-off is mandatory, is itself judgment that cannot be outsourced.

This reveals the core shift: as AI flattens execution barriers, the PM's value moves from "how much can I do" to "how well can I judge." Drafting, number-crunching, and adversarial review are now delegatable. The freed cognitive capacity goes to the irreplaceable parts: defining the right problem, allocating bets, weighing trade-offs, and ultimately signing off: "This is the version we ship."

Tool Specialization: IDE Assistants vs. Project Assistants

WorkBuddy differs from IDE coding assistants. The latter are code editors — fast, single-file, immersed in coding. WorkBuddy is a cross-file, long-context project assistant: complex logic synthesis, PRD writing, Excel analysis, extended Q&A. One ensures "this code is correct"; the other ensures "this problem is understood." Professional specialization lets each do its job better.

A telling moment: I pasted an obscure legacy error with zero context. WorkBuddy retrieved an interface spec I'd uploaded two weeks earlier from chat history and explained the error in context. It remembered what I was doing, not just answering in isolation. That continuity is prerequisite for trusting it with judgment support.

Conclusion: A Lever for Thinking, Not a Vending Machine for Answers

"Outsourcing myself" is a misnomer. I offloaded the lowest-value 15-year skills — polishing documents, crunching spreadsheets, finding my own bugs. What remains is the scarcer, higher-value work: problem definition, resource allocation, trade-off decisions, and the final call. AI is a lever for thought; it moves effort from where it's depreciating to where it's appreciating.