Most product managers believe they're "leveraging AI" because they use it to write PRDs and summarize meetings. They are not leveraging AI. They are leveraging a fast typewriter. Which is a bit like claiming you've mastered the violin because you've learned to use it as a doorstop.
Real AI leverage operates on the quality of your judgment, not the speed of your output. And here, as in so many areas of modern business, we find people optimising furiously for the thing that doesn't matter while cheerfully neglecting the thing that does.
There are four levels of AI leverage. Most PMs are stuck on the first one. Almost none have reached the fourth. And the thing that matters most isn't a level at all.
---
Level 1: Throughput Leverage. AI Does Tasks You Already Know, Faster
First-draft PRDs. Summarizing meetings. Formatting docs. Competitive research.
These are the tasks where good-enough is good enough. AI brings them from a 3 to a 7. Real time savings. No argument from me.
But Level 1 has a trap, and it's a seductive one: it feels so good that people stay there.
You save five hours a week. Wonderful. But do you reinvest those hours in deep customer work, strategic thinking, building judgment? Or do you fill them with more meetings and more documents nobody reads?
Be honest. Most people take the freed-up time and immediately spend it on more exploit: more of the known, the comfortable, the measurably productive. They never reinvest in explore.
This is the business equivalent of what happens when you build faster roads: they fill up with traffic. Economists call it induced demand. I call it the chronic inability of organizations to leave any efficiency gain uninvested in more of the same mediocrity.
---
Level 2: Capability Leverage. AI Enables Tasks You Couldn't Practically Do Before
This is the jump most people miss. Level 1 is about speed. Level 2 is about possibility.
Rapid prototyping without engineering. Testing ten positioning variants instead of two. Translating strategy docs into multiple formats simultaneously. Synthesising fifty customer interviews in minutes.
These aren't things you were doing slowly. You weren't doing them at all.
With good judgment, this is a blessing. With poor judgment, you can burn your team. And here's why.
Level 2 has a failure mode almost nobody talks about.
When you write a strategy from scratch, the blank page forces you to confront what you actually think. The struggle is the point. The struggle builds your internal model. It's the cognitive equivalent of the IKEA Effect: you value it more because you built it, but more importantly, you understand it because you built it.
When AI drafts your thinking, you skip the building and go straight to evaluating. This is the Rider and Elephant problem. Your conscious mind, the Rider, reads the AI output and says "this is coherent, well-structured, addresses the key points." The Rider is satisfied. But your product intuition, the Elephant, the pattern-matching system that actually makes good decisions, hasn't been exercised at all. It's been asleep while the Rider plays editor.
Most people cannot tell the difference between "this is good" and "this sounds good."
This creates a compounding debt:
AI drafts your thinking → you don't develop judgment to evaluate it → you depend more on AI next time → judgment atrophies further → repeat.
The PM feels more productive while becoming less capable. It's a competence debt that compounds silently, and like all the worst debts, the interest payments are invisible until the principal comes due.
The rule for Level 2: use it aggressively for execution and exploration. Use it carefully for synthesis and strategic thinking. Ask yourself: Is this task building my internal model, or substituting for it?
---
Level 3: Cognitive Leverage. AI Expands Your Reasoning Capacity
Stress-test your strategy and the assumptions behind it. Identify blind spots in your roadmap logic. Generate counter-arguments to your own pitch. Explore what happens if a key assumption is wrong. Simulate multiple futures based on emerging trends.
Here's what Level 3 looks like in practice. You're planning to launch a self-serve tier for your enterprise product. Before AI, you'd present the strategy, get questions in the review, and address them reactively. At Level 3, you prompt AI with your full strategy and ask: "What are the three most likely ways this cannibalizes our enterprise pipeline, and what early signals would I see?" Then you take those scenarios and stress-test them against your actual pipeline data. AI didn't make the decision. It expanded the surface area of what you considered before making it.
Level 3 is higher leverage than 1 or 2 because it operates on the quality of your thinking, not the efficiency of your output. When every PM has the same AI tools, and they will, the one who asks better questions and stress-tests more rigorously wins.
AI runs the simulations. You choose which simulations to run. Output quality is entirely a function of your input quality. Your questions, your framing, your product sense.
Level 3 doesn't reduce the need for judgment. It makes judgment more important.
AI as a thinking partner only works if you can identify good thinking from bad thinking.
And this is why the levels are a system, not a ladder. If you've atrophied your judgment by over-relying on AI at Level 2, your Level 3 work will be shallow. You'll ask AI to "brainstorm risks" and get a generic list instead of probing the specific, real threats that only someone with deep domain knowledge would think to test.
It's like giving a Michelin-starred kitchen to someone who's only ever microwaved ready meals. The equipment is spectacular. The operator is not.
---
Level 4: Judgement Leverage. AI Challenges and Sharpens Your Judgment
This is not "suggest a better layout," that's Level 1. Level 4 is when AI tells you your dashboard measures the wrong thing entirely. That your strategy optimises for a local maximum. That your A/B test presupposes its conclusion.
Here's what Level 4 actually looks like. You've spent two weeks on a retention strategy focused on reactivating churned users. You prompt AI with the full strategy and your churn data. It comes back: "Your strategy assumes churned users left because of missing features. But your activation data shows 40% of churned users never completed onboarding. You don't have a reactivation problem. You have an activation problem. Your entire strategy is solving the wrong stage of the funnel."
Now: is that right? Maybe. The AI is pattern-matching across its training data. It might be surfacing a blind spot you missed because you were anchored on the churn narrative. Or it might be applying a generic "check your onboarding" heuristic that doesn't apply to your specific market. Knowing the difference is taste.
Level 4 requires what I'd call taste: the ability to identify the quality of a thing, sometimes in non-consensus ways.
Without taste, you either reject every AI challenge (ego) or accept every one (no confidence in your own judgment). Both are common. Both are disastrous. The first is the person who treats every suggestion as an attack. The second is the person who has outsourced their thinking so completely that they've become a kind of human rubber stamp.
This requires the hard-won judgment that Levels 1 through 3 are meant to build.
---
Where Most PMs Actually Are
Based on my interactions with PMs, I'd say most PMs are at Level 1. A meaningful minority have reached Level 2. Almost none operate at Level 3 or 4.
The PMs most vocal about AI usage are almost always at Level 1. This should surprise nobody. The people who've barely scratched the surface are always the ones most confident they've mastered it.
---
And Now the Thing That Matters Most: Level 0. Protect.
I've saved this for last on purpose. Because after reading about four levels of increasingly sophisticated AI usage, the temptation is to race up the ladder. But the most important level isn't on the ladder. It's about what you keep off it entirely.
Reviewing raw customer feedback. Struggling through a strategy draft. Sitting with usage data until patterns emerge.
These aren't inefficiencies. The cognitive effort is the value. Don't outsource it.
Now, this is a point most productivity-obsessed people find uncomfortable, because it sounds like I'm defending waste. But consider: if a large language model were trained on Wikipedia summaries instead of raw internet text, it would be dramatically worse. Your product intuition works the same way. It's trained on raw exposure to messy, contradictory, occasionally baffling reality, not on clean AI summaries of that reality.
And the dangerous part: the highest-leverage thing most PMs do, building their internal model of the customer and market, is the very thing they're most eager to hand to AI. The hard cognitive work is uncomfortable and produces no visible output for hours. AI offers what feels like an escape. But it's a fake escape, like the man who takes a taxi to the gym so he doesn't have to walk.
---
The Meta-Framework
Use AI to do more of the work that doesn't require your judgment, so you have more time to develop and apply the judgment AI cannot replace.
---
The Audit
Tomorrow morning, review your last week of AI usage. Mark each task:
- Level 1: Throughput Leverage (did it faster)
- Level 2: Capability Leverage (couldn't have done it otherwise)
- Level 3: Cognitive Leverage (improved my thinking)
- Level 4: Judgement Leverage (challenged my assumptions)
- Level 0: Should have done it myself (the struggle was the point)
Then ask yourself one question:
"Am I using AI in ways that make my judgment stronger, or in ways that make my judgment unnecessary?"
The answer will be revealing. And if you find it uncomfortable, good. Discomfort is the feeling of your internal model being updated. Don't outsource that either.