The AI PM Is a Technical PM
No-code tools make prototyping effortless, but solving hard problems still requires systems thinking. Here's what AI product managers should actually focus on right now.
We are living through a strange moment in product building. A single prompt can spin up a working prototype. A drag-and-drop workflow can connect an LLM to a database, a UI, and a payment flow. The barrier to starting has never been lower.
But the barrier to solving hard problems is unchanged. And that is where the real work of an AI product manager begins.
The no-code illusion
No-code tools are genuinely transformative. They compress weeks of scaffolding into hours. They let you validate an idea, test a workflow, and show a customer something concrete before you commit engineering time.
The trap is mistaking that speed for depth. Prototyping is not product strategy. A demo that looks impressive can hide a system that falls apart at the edge cases: bad inputs, adversarial users, silent failures, and the thousand small degradations that happen when a real workload hits the tool.
The products that win are not the ones that demo best. They are the ones that hold up under pressure. That is why technical product managers are rising to the center of AI work.
Reframe: "AI PM" does not mean "a PM who builds LLMs." It means a technical product manager who uses AI as a material. The job is not to invent the model. The job is to design the system around it.
What an AI PM actually does
The day-to-day work of a strong AI PM is less about prompting and more about structure. If you are building in this space, these are the muscles that matter most.
Build systems, not prompts
Prompts are the surface. Systems are what survive. A good AI PM thinks in loops:
- Input handling. How do you collect, validate, and sanitize the data that enters the system?
- Reasoning. What does the model actually decide, and where is that decision visible?
- Output control. How do you format, filter, and fall back when the output is wrong?
- Feedback loops. How do you learn from failures and improve the next pass?
Your product is not the LLM call. It is the orchestration around it. The orchestration is what keeps the product from embarrassing itself in production.
Design self-improving loops
The most durable AI products get better without constant manual intervention. That means instrumentation, evaluation, and iteration are part of the product design from day one.
Ask these questions in every roadmap review:
- What signal tells us the model made a mistake?
- Who or what fixes it, and how fast?
- How does that fix make the next run better?
- Where do we draw the line between automatic improvement and human review?
A product that does not improve on its own is a product with rising operational cost. That cost eventually kills the margin that made AI attractive in the first place.
Gather and parse customer feedback
AI products produce strange user reactions. Customers may be amazed by a fluent answer, frustrated by a confident hallucination, or confused by a capability that appears and disappears. A PM must separate the emotional response from the useful signal.
Your job is to listen for:
- Valid complaints. Real failures that break trust or workflow.
- Misaligned expectations. Users who expected magic because the marketing was too polished.
- Latent needs. What users are trying to accomplish that the product is not yet designed for.
- False positives. Praise for a feature that looks good but scales poorly.
Not all feedback is equal. Learning to triage it is a career-defining skill.
Accept maintenance as a feature
AI agents hallucinate. Models drift. APIs change. Guardrails degrade. If you treat maintenance as an afterthought, your product will quietly rot.
Plan for it explicitly:
- Monitoring and alerting on model behavior, not just uptime.
- Regular evals on a representative dataset that grows over time.
- Versioning for prompts, models, and embeddings.
- A clear escalation path when the system behaves badly in front of a customer.
The teams that ship reliably are the ones that build maintenance into the roadmap, not the ones that panic after an incident.
Build systems thinking deliberately
Systems thinking is not a personality trait. It is a trainable skill. It is the ability to see second-order effects, feedback loops, bottlenecks, and unintended consequences before they happen.
If you want to strengthen it, borrow from disciplines that reward rigorous reasoning. Physics and math exercises force you to model cause and effect. Strategy games force you to manage tradeoffs under uncertainty. Puzzles and simulations build the habit of asking "what happens next?" before you act.
You do not need to become a physicist. You need to spend enough time with hard problems that pattern recognition becomes instinctive.
Practice list
- Play physics-based or logic games that reward first-principles reasoning.
- Work through math problems that require setting up a model before solving.
- Read engineering postmortems and trace the causal chain backwards.
- Map every AI feature you ship as a flowchart of inputs, decisions, outputs, and failure modes.
This moment is transitory
I believe the current hype around prompts and AI agents is temporary. Prompts will fade into the background. Agents will stop being a category and become a layer. The tools will get better, the interfaces will get simpler, and the competitive advantage will shift back to the fundamentals.
What will not go away is the need for people who understand systems, customers, and business context. That is the job. The rest is just the current interface.
From a friend: If you are deciding what to invest in right now, here is what I would do. Focus on your psychological read on people. Protect your health with meditation and movement. Study decentralization and solarpunk. Build systems for small and medium businesses. Learn to reason clearly. The tools will change; the underlying skills will compound.
What comes next
In the near future, software will become customizable by default. Voice will take over more of the interface. As the UI layer gets thinner, product design will matter less in terms of pixel polish and more in terms of intent, timing, and trust.
People are not all the same. Some are kinesthetic. Some are visual. Some are auditory. One interface does not fit everyone. The PMs who win will be the ones who design for learning styles and context, not just for the screen.
This week's practice
Pick one AI feature you are working on or using. Draw its full system diagram: inputs, model, outputs, feedback loops, failure modes, and maintenance burden. Then ask: which part is most likely to break under real load? That is your next problem to solve.
- ›No-code makes prototyping fast, but hard problems still require systems thinking.
- ›AI PMs should build systems, self-improving loops, feedback discipline, and maintenance practices.
- ›Agents hallucinate and degrade; plan for monitoring, evals, and escalation.
- ›Train systems thinking through physics, math, games, and rigorous causal mapping.
- ›The current AI hype is transitory; the underlying skills are not.
- ›Future software will be customizable and voice-driven; design for human learning styles, not just screens.
/subscribe
Get one issue like this every week.
Upskills Academy Weekly: news about product building with AI plus practical product management advice.