In February 2025, Andrej Karpathy described a new way to write code. Give direction. Let AI execute. Iterate on results. He called it vibe coding. Collins Dictionary named it Word of the Year.
Within months, the same pattern reached product management. PMs started using AI to write PRDs, run competitive research, synthesize user interviews. But most of them are doing it wrong.
They open ChatGPT. They type a prompt. They get generic output. They spend an hour editing it to sound like someone who actually knows their product. Tomorrow they do it again. The AI remembers nothing. Every session starts from zero.
This is how 80% of PM teams use AI today. It's a search bar with better grammar.
There's a better way. And it has six levels.
Near the top.
Karpathy scored 342 US jobs on AI exposure.
Product management scored near the top.
The adoption gap.
88%
Of organizations say they use AI. Only 6% see meaningful bottom-line impact.
McKinsey, 2025
49%
Of product teams lack time for strategic planning or data analysis. AI hasn’t fixed it.
Atlassian, 2026
+40%
Higher quality output when AI is embedded in structured workflows vs. ad-hoc prompting.
Harvard/BCG, 2024
3x
More likely to see positive AI returns with structured frameworks vs. ad-hoc adoption.
McKinsey, 2025
Everyone uses AI. Almost nobody has a system for it.
The gap between Level 1 and Level 6 isn't the tool. It's the operating system.
Their Monday.
...5:00pm
Nothing strategic happened today.
Tomorrow it starts over.
Your Monday.
Same AI. Same deliverables. Same fire drills.
The difference is the operating system.
The 6 levels.
Most PM teams are at Level 1 or 2.
The teams at Level 4+ are operating at a completely different speed.
The gap compounds every month.
Level 1
AI as a Search Bar
This is where 80% of PM teams are today.
A PM needs to write a PRD. They open ChatGPT, type "help me write a PRD for a checkout redesign," get something generic, spend an hour giving Chat more information about the company and product. Finally, they paste it into a doc and move on.
Next week, different PM, same task. Starts from scratch.
What it looks like
The AI doesn't know your product. It doesn't know your users. It doesn't know your competitors. It's a search bar with better grammar.
Level 2
AI as an Assistant
Your team has shared prompt templates. Maybe a Notion doc with "how to use AI for competitive research." Maybe a few structured skills with detailed instructions.
This is real progress. PMs aren't starting from zero anymore.
What it looks like
This is where most AI-forward PM teams plateau. They think the problem is better prompts. It's not. The problem is the AI has no idea who you are.
Level 3
AI as a Teammate
This is the level that changes everything.
Someone on the team — usually your best PM — has loaded company context into their AI setup. Product details. Customer personas. Competitive landscape. Strategic priorities.
Now the AI doesn't just produce structure. It produces output that sounds like someone who works at your company.
What it looks like
Here's the problem: it's personal, not shared. Your best PM's setup doesn't transfer to the next PM. If they leave, that knowledge walks out the door.
This is where the Head of Product has a decision to make. Do you let every PM build their own system? Or do you build one system the entire team plugs into?
Level 4
AI as the PM Operating System
The whole team runs on the same system. Same company context. Same skills. Same frameworks. Any PM runs the same task and gets consistent, business-aware output.
What it looks like
This is where compounding starts. Every PM workflow makes the system smarter. Every research round updates the shared context. Every competitive shift gets reflected in every PM's output — automatically.
At Levels 1–3, AI helps individual PMs. At Level 4, AI runs the PM function. That's a fundamentally different thing.
Level 5
AI as Autopilot
Workflows run without anyone triggering them.
Weekly competitive scans that show up in Slack Monday morning. Customer feedback synthesis that updates after every support cycle. Metric reports that flag anomalies before the PM even logs in.
What it looks like
This is the agent level. Your PM system delegates tasks to specialized agents. A research agent pulls competitive data. An analyst agent structures the findings. A writer agent drafts the brief. A critic agent scores it before you ever see it.
One PM doing the work of a small team. The coordination cost drops to zero because there are no handoffs.
Fewer than 5% of PM teams are here today.
Level 6
AI as the Engine
The difference between Level 5 and Level 6 is memory.
At Level 5, agents execute. At Level 6, agents learn.
The system tracks what works and what doesn't. It feeds performance data back into the next cycle. Context doesn't just stay current — it gets richer every week.
What it looks like
This is where AI stops being a tool you use and becomes infrastructure you run.
Almost nobody is here yet. The teams building toward it now will have compounding advantages that late adopters can't catch up to.
The tool is not the advantage.
How you use it is.
Foundation
Company context. Product details. Customer personas. Competitive landscape. Strategic priorities. This is the context layer — every skill, every agent, every output starts here. Without it, the AI guesses. With it, the AI knows your product.
Skills
PRD generation. Competitive analysis. User interview synthesis. Sprint planning. Stakeholder updates. Each task runs through a dedicated skill with PM frameworks baked in — Teresa Torres, Marty Cagan, April Dunford. Not a generic prompt. A methodology file.
Agents
An agent reads your goal, plans the steps, picks the right skills, and executes. One command triggers a full workflow. The research agent pulls data. The analyst structures findings. The writer drafts. The critic scores. You review the output, not the process.
Memory
The system learns. What resonated with stakeholders. Which competitive angles landed. What positioning converted. Every cycle feeds the next. The PM OS doesn't just run — it compounds.
A skill is not a prompt.
It's a methodology file.
A Prompt
“Write me a PRD for a checkout redesign.”
A Skill
1,200+ lines of structured methodology:
You write the methodology once. It runs forever. And it runs on your context. The AI doesn't guess — it follows proven frameworks, references your actual product, and scores its own output.
Just a markdown file.
The gap compounds.
A team at Level 4 onboards a new PM in days, not weeks. They run competitive analysis in hours, not days. They produce consistent output across the whole team.
A team at Level 1 is still copying and pasting into ChatGPT. Each PM reinventing the wheel. Output quality depending on who wrote the prompt.
Every week you wait is a week of compounding you don't get back.
Your first moves at each level.
Pick one PM deliverable your team does weekly. Create a shared prompt template for it. Just one. See what happens.
Write a one-page company context doc. Who you serve, how you win, top 3 competitors. Paste it into your AI tool's project instructions. The difference in output quality will be immediate.
Stop letting every PM build their own setup. Create one shared context folder the whole team loads. Assign someone to own freshness.
Identify one recurring workflow — competitive scans, metric summaries, research digests — and automate it end to end. Run it for 2 weeks. Compare to manual.
Add memory. Start logging what performs and what doesn't. Feed that data back into your agents. Build a shared context layer that compounds over time.
“Marty Cagan doesn't write every PRD.
Teresa Torres doesn't run every interview.
Their entire value is judgment.”
The moat in product management was always judgment.
Knowing what to build.
Knowing what to cut.
Knowing when to say no.
AI doesn't replace that.
But teams at Level 4+ have something the others don't:
time to exercise judgment.
When your PMs aren't spending half their week producing artifacts,
they spend it thinking.
And thinking PMs build better products.
The best time to start building your PM operating system was six months ago.
The second best time is today.