The AI-Native PM Operating Model

A PRD is an opinion with formatting.

Documents are proxies

For twenty years that was fine. Building the real thing was expensive, so organizations argued about descriptions of the thing instead. Review cycles exist because a document can’t be tested. It can only be believed, and belief takes meetings.

One variable changed

AI collapsed the cost of a working artifact. A PM can now answer a feasibility debate with a deployed prototype in less time than it takes to get the debate onto a calendar. When the real thing gets cheaper than the proxy for the thing, the proxy starts dying.

The procurement trap

Buy every PM a license, run a training, declare victory. What you get is the same documents, produced faster. The organization needed fewer proxies, and it bought a proxy accelerator.

What actually works

Change the system the PMs operate in: a shared bar for what good looks like, guardrails that make speed safe, a commons where one PM’s tooling becomes everyone’s, and honest measurement. That’s the rest of this page.

The ladder

Five levels, defined by what you can observe. The test at every level: what does this person’s work change for other people?

Assisted

A faster keyboard.

Drafting and summarizing on an unchanged workflow. Where most training programs stop.

Fluent

Patterns, not prompts.

Engineered context, deliberate model choice. One person gets faster. Peers start asking how.

Builder

Answers are software.

Prototypes, pipelines, dashboards. Engineering evaluates instead of interprets, and feasibility arguments get shorter.

Orchestrator

Runs agent teams.

Publishes tools colleagues actually install. The first level where the gains leave the individual.

Multiplier

Changes the discipline.

Ladders, gates, and measurement wired into how the org works. The gains survive the author.

Most orgs today: many 1s, a few 3s, zero 5s. Not because the skills are rare. Levels 4 and 5 aren’t skills at all, they’re permissions. Someone with authority has to let you change what the PM job is defined as. That makes this an org-design decision wearing a training program’s clothes.

Seven gates

The questions a product decision answers before AI ships in it. Written as gates rather than principles for a simple reason: you can’t fail a principle.

Grounding

Every answer cites its source, or it doesn’t ship to people whose decisions cost money.

Data boundary

Which data reaches which model, under what agreement. “It’s just a prompt” is how data leaks.

Accountability

A named human owns every output that reaches a customer. No name, no ship.

Failure design

Price the wrong answer before the incident. That’s the cheap time to do it.

Accessibility

WCAG 2.2 AA in the first release, not the retrofit. Excluded users don’t file bugs. They leave.

Measurement

A baseline, and a metric that could show failure. Enthusiasm is not an instrument.

Disclosure

People know when a machine wrote it. They find out eventually anyway; the only choice is whether you told them first.

Principles describe intent. Gates block launches. The test of a governance model is whether a team can fail it. This one can.

Ninety days

Measure, then build the commons, then wire it into promotions. Most programs run this backwards, which is why most programs are slide decks.

Baseline & seed

Instrument cycle times first: discovery to PRD, PRD to decision, question to answer. You can’t claim gains you never baselined, and someone will ask.
Map the org against the ladder, honestly. Expect a pyramid.
Find the level-3s already hiding in the org and make them the seed cohort. Give them a sanctioned sandbox so experimenting stops being shadow IT.

Build the commons

An internal marketplace where PM-built agents, skills, and templates are installable by anyone. One level-4 becomes a hundred faster PMs.
The seven gates go into intake and PRD templates. Governance as a form field, not a memo.
Publish the first before-and-after numbers. Evidence recruits volunteers; mandates recruit compliance.

Wire it into promotions

The ladder goes into the career framework. Promotion criteria are the only mechanism that survives leadership turnover.
A lightweight review board for gate exceptions, so governance has an owner and a queue instead of a vibe.
Rotate the seed cohort into every product family as named practice leads.

Receipts

An operating model is worth exactly as much as its author’s practice.

Production

AI, shipped

I wrote the first version of a RAG engine that now answers field technicians’ questions every day, with cited sources. Engineering took the prototype into the official build. That’s the only compliment from engineering that counts.

Adoption

Tools other PMs install

Agents and skills published to an internal AI marketplace: a notes-to-prototype pipeline, an automated design-standards auditor. Adoption you can count in installs, not slideware.

Scale

Platforms, not features

Digital twins, predictive analytics, IoT monitoring, and work-order systems for one of the largest physical operations in the world. Before that: cloud-native 5G, and three spacecraft from vendor selection to on-orbit commissioning.

Velocity

Audited, not asserted

Discovery to PRD went from weeks to days. That measurement came from independent audits of my own session logs, because gate six applies to me too.

Depth

Personal practice

Fine-tuned models, local open-weight hosting, multi-agent orchestration with cost-aware routing. Built at home, on my own hardware. The tradeoffs I argue at work are ones I’ve already paid for.

This page

The site is the demo

Built with my own agent tooling. One HTML file, no dependencies, no scripts, no trackers, accessible by default. Shipping beats describing.

Accessible by default

Built to WCAG 2.2 AA. Accessibility is a PM competence, and failing gate five on my own homepage would be a bad look.