AI didn't replace PMs. It repriced them.

Four or five years ago, working inside a corporate structure, I could take a week to "think properly." A couple stakeholder syncs. A PRD. A doc review. A tidy plan. Engineering builds it. We learn something in a month. That was the rhythm, and nobody questioned it.

Then ChatGPT dropped in late 2022, and the perspective started shifting. Slowly at first. I was experimenting with drafts, some teams playing with summaries. But the ground was already moving under us.

Now, in 2026, I'm no longer interested in debating what to build in a meeting about a meeting. It's genuinely faster to build a prototype, test the riskiest assumption, and bring back real results. The cycle that used to take a quarter now fits inside a week. If you let it.

Here's the uncomfortable part:

AI didn't kill product management. It killed the luxury of slow cycles.

The anxiety PMs feel right now is rational. The market is compressing. Expectations are rising. And the weakest position in this new game is the one too many of us were trained for:

The spec-only PM.

Not because specs are useless. Because if your main value is translating ideas into documents, AI just made that commodity-cheap.

“AI didn’t kill product management. It killed the luxury of slow cycles.”

Nikita Simakov, Experienced CPO and AI Expert

The job isn't "write better docs." The job is to turn ambiguity into execution, cut decision latency, and ship learning loops fast enough that strategy doesn't detach from reality.

That's the recalibration.

The shift: from "spec writer" to "hybrid operator"

When cycles are long, you can get by as a coordinator. When cycles collapse, coordinators become bottlenecks.

Andrew Ng put it bluntly at AI Startup School:

"I don't see product management work becoming faster at the same speed as engineering. Just yesterday, one of my teams proposed not having 1 PM to 4 engineers, but 1 PM to 0.5 engineers. For the first time in my life, managers are proposing having twice as many PMs as engineers."

This isn't abstract. It's a structural shift happening in real time. Lenny Rachitsky's large-scale productivity survey from December 2025 showed that PMs are already clawing back at least half a day per week through AI tools — and that number keeps climbing. Kevin Weil from OpenAI nailed it: "The AI model that you're using today is the worst AI model you will ever use for the rest of your life."

The competitive PM in 2026 is a hybrid operator — product sense plus technical literacy plus AI leverage across discovery, delivery, and GTM. We become unstoppable.

Not "I'm a PM who prompts." That's not a thing.

The edge isn't prompting. It's engineering fundamentals, so you can reason about constraints. Systems thinking, so you optimize the whole loop rather than a single step. And workflow integration, so AI output becomes execution.

AI is the multiplier. You still need the machine.

What I got wrong

"Prompting is a skill." No. Prompting is typing. The actual skill is building verification loops — assumptions, constraints, failure modes, checks against source-of-truth, and observable outcomes. Without those, AI just helps you be confidently wrong faster.

"Engineering literacy is optional for product leaders." In compressed cycles, that's the baseline. If you can't reason about APIs, auth, data models, deployments, and observability, you can't make fast calls responsibly. You'll either slow everything down or ship landmines into production. As LogRocket's 2026 breakdown noted:

"Technical fluency isn't about coding every feature yourself. It's about knowing enough to ask the right questions, recognize quality outputs, and guard against catastrophic logic failures."

"Speed is mostly about moving faster." Speed is mostly about removing rework. AI makes it trivial to generate output. The bottleneck shifts to whether it's correct, testable, measurable, and safe for prod.

Four workflows where I actually use AI (without kidding myself)

There are the actual workflows I run because they create throughput and cut rework.

1. GTM strategy into execution. I state a strategic bet in one sentence: "If we position X for segment Y, we get outcome Z." AI unpacks that into 3 micro-ICPs with different triggers, objections, and proof needs. For each, it generates message angles, disqualifiers to filter out garbage pipeline, outreach sequences, and a one-pager sales can run without me. Then I force reality — pick one channel, ship the sequence, measure response and conversion. I know it's working when pipeline velocity moves, reply rates climb, sales stops asking "what do I say," and we learn which ICP slice actually bites.

2. Metric diagnosis: anomaly to truth, fast. I define the anomaly precisely — "Conversion dropped 12% WoW in this segment starting on this date" — and AI proposes a short root-cause set: tracking break, traffic shift, UX regression, latency issues. It then builds diagnostic queries and a decision tree. I validate instrumentation before blaming the product. Time-to-diagnosis shrinks from weeks to hours. The next action becomes obvious — fix tracking, rollback, isolate a segment, or run a controlled test.

3. Interviews and pre-sales calls: extracting signal while exposing my own gaps. AI standardizes messy call notes into pains, desired outcomes, constraints, decision process, objections, and exact phrases. Across multiple conversations, it clusters themes — recurring objections, buying triggers, reasons people walk away. Then I do the uncomfortable part: I ask it to critique me. What questions I skipped. Where qualification was soft. Where framing fell flat. Calls get tighter. Objection handling sharpens. Fewer late-stage "not a fit" surprises.

4. Fast prototyping: testing the uncertainty, not the idea. This is where things changed for me. I use Claude Code to go from hypothesis to working prototype in hours rather than sprints. I define the smallest falsifiable test — often a fake door beats a full feature — and generate a working prototype straight from the terminal, describing what I need in plain English. Claude Code produces the spec outline with edge cases, test cases from acceptance criteria, and an instrumentation plan covering events, properties, dashboards, and guardrails. Then we ship a controlled rollout with feature flags, staged launch, and a rollback plan. This isn't vibe coding. It's a functioning prototype with telemetry that I can put in front of my team an hour after forming the hypothesis. Dennis Yang, a PM at Chime, described a similar approach — he writes a PRD in markdown, opens a terminal, types claude, and twenty minutes later the team is already iterating on something real. The critical checks: if it can't falsify the hypothesis, it's theater; if it's not instrumented, it's not a test; if you're overbuilding when a concierge test would do, you're burning cycles. Test loops compress. Decisions get cleaner — kill, iterate, or scale. Rework drops because instrumentation and QA thinking happen upfront.

The PM skill stack for 2026

These are the capabilities you'll reach for weekly if you want to stay competitive.

Software architecture basics: how systems connect, where complexity hides, trade-offs, non-functional requirements.

Core engineering principles: APIs, auth, data models, event-driven versus request-response patterns, integration thinking.

Dev and prod reality: deployments, feature flags, rollbacks, incident response, observability. 
A testing mindset: unit, integration, end-to-end coverage: what needs to be tested and why, because PMs enforce this to protect velocity.

And analytics literacy in an AI-first world: instrumentation discipline, evaluation loops, experiment integrity, contamination risks.

The bottom line

Product management isn't dying. It's forking. There are PMs who see where things are headed and are already benefiting. And there are those who don't, and every month it gets harder to catch up.

Your "Product Excellence" framework isn't wrong. It's just calibrated for a different environment. The goal right now isn't to be the most correct person in the room. It's to be the one who learns fastest.

Drop the corporate layers. Trust your instincts. And get comfortable with the beautiful mess of building from scratch, powered by tools that didn't exist a year ago.