I recently listened to a talk by Prof. Rubingsheng Ru on the growth path of AI-native PMs. A framework in his talk resonated with me: in the AI era, product managers tend to move through four stages: tool use, capability transfer, system design, and finally value judgment.
Stage 1: PMs start using AI tools, but the product has not changed yet
This is the most common and easiest to understand stage. We use AI to draft PRDs, summarize meeting notes, analyze competitors, organize user feedback, or generate first drafts of ideas. But what matters most is not "how much efficiency I gained," it is whether the work style shifts from consumption to investment.
In the past, many tasks ended when the delivery was done. Now, a high-quality AI conversation can be distilled into skills, standards, knowledge cards, and methods. So the key at this stage is not whether you can use the tools, but whether you can turn each delivery into a long-term asset.
The stronger AI becomes, the more PMs need to train basic abilities: asking, thinking, judging, hypothesizing, predicting, attributing, experimenting, and innovating. AI can generate a lot, but quality still depends on whether you ask the right questions, judge the answers, and design verification paths.
Stage 2: AI is integrated into products, but mostly as point enhancements
At this stage, products start to embed AI: smart summaries, auto generation, recommendations, AI customer service, auto correction, image enhancement. The product form itself does not fundamentally change yet. It is still "AI + product."
A key reminder: good AI features do not need to stand in the foreground and say "I am AI." Many times AI should be like photo auto-enhance, keyboard auto-correct, or smart music in editing tools, quietly helping users do things better. The user goal is not to "use AI"; the user goal is to complete a task.
That is why many AI products fall into the trap of turning everything into a chatbot. But do users really need a chat box? Not always. Sometimes auto-complete, smart suggestions, background detection, or a well-timed one-click generation feels more natural.
What PMs must do at this stage is understand the boundaries of LLMs. When can we trust it? When must we question it? What can be generated, and what must be controlled by rules, workflows, reviews, and safeguards?
Generative models are inherently uncertain and hallucinations are hard to eliminate. The PM's value is not blind trust, but designing product systems that wrap uncertain model capabilities into stable, controllable, and trustworthy processes.
Stage 3: PMs become builders of AI-native products
At the third stage, change is no longer about adding AI features. The product form itself is reshaped by AI. AI-native products are not traditional software, nor fully improvised software. They sit between: stable product systems with internal dynamic decision-making and task orchestration based on user intent.
They usually have at least three layers. The first is the intent layer, where users do not just click buttons or fill forms, but express goals. For example:
- "Summarize this user interview into product opportunity points."
- "Monitor how this batch of KOL videos performs after publishing."
- "Identify the main risks behind this project's delay."
The second is the orchestration layer, where agents actually work. It decides how to break tasks down, which steps are done by the model, which steps call tools, which require user confirmation, which must use deterministic logic, how to retry on failure, and whether to ask follow-up questions when information is missing.
The third is the execution layer: querying databases, calling APIs, creating tasks, updating CRM, payments, reservations, and messaging. The more deterministic this layer is, the more reliable the system becomes.
So AI-native products are not "let AI do everything." It is the opposite. PMs must decide which steps need AI understanding and generation, which should rely on deterministic systems, and which require human confirmation.
This raises a real issue: human-in-the-loop is not only an experience choice, it is also a responsibility boundary. AI can draft an email, but a human should confirm before sending. AI can plan a trip, but payment and booking should require confirmation. AI can assist medical judgment, but cannot replace a doctor's diagnosis.
PRDs also change at this stage. In the past, PRDs documented requirements, flows, interactions, and edge cases. AI-native PRDs must also define evaluation criteria, fallback logic, and test data. What is a good output? What happens when the model is uncertain? How do we recover from tool failures? Should we enforce confirmation for high-risk actions? Do we have task sets, boundary cases, high-risk cases, and multi-turn examples? Without these, AI products risk being "looks okay this time," without knowing whether they are truly improving.
Stage 4: PMs become value definers in the post-AI era
This is the most important layer for me. In the AI era, the winning PMs are not necessarily the ones who understand AI the most, but the ones who understand users the most. Model capabilities will become common, tool use will become common, and agent architectures will become common. What becomes scarce is:
- Do you understand the user's real context?
- Do you know which problems are worth solving with AI?
- Do you know where AI should not be overused?
- Can you judge whether a feature truly helps users or only creates a sense of "smartness"?
Not every problem fits AI. Some problems do:
- Information overload that needs understanding and filtering.
- Complex workflows that need task decomposition.
- High-cost content creation that benefits from assistance.
- Ambiguous user intent that needs interpretation.
- Strong personalization that rules cannot cover well.
But many problems may not fit AI:
- Users need certainty and control.
- Error costs are extremely high.
- Users can already complete the task quickly.
- AI makes the workflow more complex.
- Users do not want to be auto-judged, auto-recommended, or auto-guided.
Mature AI PMs are not eager to AI-enable everything. They stay restrained. They know what should be AI-driven and what should remain simple, transparent, and controllable.
Closing
A final question every PM should ask in the AI era: Is your AI making users feel respected, manipulated, or exploited?
Good AI products do not constantly remind users "this uses AI." They make users feel: this product understands me, this flow is smooth, this suggestion arrives at the right moment. It saves effort without taking away control. It helps users reach their goals instead of being dragged by algorithms.
The growth of AI-native PMs may not mean becoming more technical, but becoming better at value judgment. Being able to use AI is the starting point. Designing AI features is the next step. Building AI-native systems is capability. But the real height comes from understanding users, having taste, judgment, empathy, and responsibility.
The most important question in the AI era might not be "Will AI replace me?" but rather "Can I define the problems truly worth solving with AI?"