The PM Paradox: How AI Is Teaching Us to Do the Wrong Job Better

Let me start with a thought experiment.

Imagine you're a product manager for a soft drinks company. Your job: end the dominance of the market leader. You sit down, open a blank doc, and start listing the variables.

• Taste - your product needs to taste better.
• Price - competitive, probably cheaper.
• Packaging - more appealing, more variance.
• Distribution - wider.
• Brand - stronger.

You're covering the fundamentals. You're doing the work.

Now, if I asked you to guess the most popular soft drink in the world, most people would say Coca-Cola. And you’d be wrong. The answer is Red Bull.

Here's where it gets interesting. Red Bull is more expensive than Coke. It comes in a single, 250ml can, no bottles, no two-litre options, no variety. And the taste? When Red Bull first commissioned consumer research, the marketing agency came back and said it was the worst-performing taste test they had ever conducted for any soft drink. One actual specific piece of feedback was that “it tasted like p*ss”. I'm not editorialising, that was the feedback.

So how is it the most popular soft drink on the planet?

Because Red Bull didn't compete on the variables that any reasonable product manager would have identified. It didn't try to win on taste, packaging, or price. Instead, it tapped into something else entirely, something that no spreadsheet would have surfaced. The Red Bull PMs understood that when people were drinking Coke at their desks, what they actually wanted wasn’t taste? They wanted a competitive edge. They wanted fuel. They wanted an identity that wasn't "a person who drinks what everyone else drinks." They wanted the energy, the buzz and the subtle signal of ambition.

Red Bull didn't win by being logical or rational, it won by understanding an insight that was psychological, emotional, and deeply human.

Now. Imagine you took that same brief to an AI today: "I'm launching a new soft drink into a market dominated by Coca-Cola. Where should I compete?" Feed it the taste test data. Feed it the market analysis. Ask it where the opportunity is.

I will tell you now, with full confidence: no AI would have recommended building Red Bull. Because there is no substitute for getting out into the world, sitting with real human beings, and understanding the nuanced, irrational, emotional, and often embarrassing things that actually drive behaviour.

This is not an anti-AI post. But it is about something that's quietly going wrong in our industry and I think we should talk about it.

Something strange is happening to product management

AI arrived, and in a lot of ways it's been genuinely brilliant. I use it every day. This post was started by speaking stream-of-consciousness into a voice tool called SuperWhisper and then shaping the transcript into something readable and editing it. That's AI amplifying thinking and that's the right use.

But there's also a narrative that's taken hold, one that I think is doing real damage, and it goes something like this:

If AI can now write code, design interfaces, run A/B tests, and generate copy, then the PM should do those things too. Ship more. Build faster. Become a full-stack generalist. Vibe-code a prototype before the standup.

And I understand the appeal. It feels empowering and the tools are extraordinary. Cursor, Lovable, Vercel etc. allow you to go from idea to something that looks like a product in an afternoon. The principle of immediate effect, essentially seeing the product of your effort instantly, is a huge psychological pull. The dopamine hit is real.

But just because we can do something doesn't mean it's the job. And I think we've confused what is now possible for PMs with what is actually valuable from PMs.

PMs are not architects of software. We are architects of competitive advantage.

The job is to understand what to build, not to build it. That sounds obvious however, in practice, it is not. And the "PM as Builder" wave is systematically eroding the thing that makes product management actually worth having.

The first casualty: discovery

When PMs orient around building, they orient around the solution space. And the solution space is seductive. Ai allows us to quickly produce something concrete, shippable, and demonstrable. You can put it in a slide deck. You can show it to a board. You can post it on LinkedIn.

The problem space is messy, slow, invisible and hard. You cannot show a board a problem. You can show them a prototype.

So under pressure from velocity culture and the AI-enabled ability to ship something every day, we've collectively started compressing the stage that determines whether what we build actually matters. And we've got dramatically better at building things that don't matter

The shift-right trend that is happening fundamentally contradicts the very principles, and value of product management. And I get the “but we can prototype really quickly” argument. But how did we determine what to prototype? Prototypes are high fidelity artefacts that cause imprinting in our test subjects. This was the very thing sketching and wireframes were designed to prevent, yet we seem to have forgotten about the value of low-fidelity artefacts when defining the problem space.

There are five specific failure modes I keep seeing. They're not independent, they compound each other.

Failure Mode 1: We skipped straight to the solution

AI tools are great for generating great answers. They’re not so good at generating good questions.

When the instruments you use every day are optimised for output, your thinking becomes optimised for output too. The question quietly shifts from what problem should we solve? to what should I build next? They are two distinctly different trains of thought.

Ask yourself honestly: how many things have you shipped in the last six months that looked great but didn't move the metric you hoped? I’ve done it!

Real discovery, going to talk to customers, sitting with ambiguity, spending a week just trying to understand a market, doesn't produce a deliverable but it produces understanding. And in a world of daily AI-assisted shipping velocity, anything that doesn't produce a visible deliverable starts to feel like waste. Yes, you can ask Claude for a market analysis of the software test market and how it relates to manual, v automated v Agentic testing and ask for TAM, market trends, competitive landscape etc. and it will produce a really good report. But taking that as verbatim and professing that you are now a subject matter expert is where the failure mode compounds. Use this as a guide, not as a certificate that you now know everything. Because when the rubber meets the road, and you’re in the deep discussions on PMF, a deep understanding of the market, not a superficial one, will help you win.

Failure Mode 2: We started treating customers like data points

There's a kind of neo-liberal economic view of the customer that's quietly crept into product thinking. This is the idea that customers have no stable preference and every buying and usage decision they make is based on rationality and maximizing utility. This is the idea that if you measure enough, you can model the human completely. That behaviour is reducible to quantifiable signals.

This view is useful. It is also dangerously incomplete.

Humans are not rational agents optimising for utility. They are emotional, social, aspirational, often irrational creatures who make decisions based on how things make them feel about themselves, about how others perceive them, and about who they want to be.

Metrics capture what people do. They rarely capture why. And the why is where competitive advantage lives.

The data-only PM is building a map of human behaviour that is accurate but useless.

Failure Mode 3: We lost our feel for the thing that can't be measured

Some of the most consequential product decisions in history were not justified by data. They were justified by taste, by judgment, by a designer or PM who understood something about human experience that didn't fit a metric.

• The iMac G3 came in colours. Not because research said "users want colourful computers." Steve Jobs and Jony Ive perceived the PC as an expression of identity. They noticed that many beige IBM PCs were hidden away in people’s homes. This was a time when owning a PC was a sense of identity. So, making it cool changed the behaviour of hiding the PC away to showcasing it in your living space. Imagine going to your finance head and saying you want to make the commodity you ship more expensive by adding blue, transparent, plastic.
• Dyson made the vacuum cleaner transparent so you could see the dirt being collected. Functionally irrelevant. Emotionally enormous. Cleaning felt satisfying. The technology felt trustworthy. Again, imagine saying that we’re going to revolutionize the vacuum market by making a cool vacuum cleaner that costs way more
• American Express prints "Member Since" on the card. It costs nothing. It completely changes how cardholders feel about their relationship with the brand. This creates an identity of loyalty. AMEx, despite having higher fees and not being accepted in many countries outside of the US, has a really high retention rate. Nobody wants to cancel their card as their member since status will reset
• Buc-ee's, the American truck-stop chain, became famous for having the cleanest women's bathrooms in America. Not because of a metric. Because someone paid attention to an experience that nobody else was, and it became a defining differentiator. Now, imagine asking a consultancy firm to analyze the business. The first thing they would do is ask why you are spending so much on the female toilets when they are free? They have no attributable value.

None of these decisions seem rational but all of them are architecting competitive advantage. All of them required understanding something true about human experience. LLMs are based on historical data that is mostly post-rationalised. Therefore, they are designed to give logical, rational answers where irrationality pervades. That's where PMs add value, that’s discovery and framing doing its job.

Failure Mode 4: We confused efficiency with improvement

There is a generation of product decisions that made processes measurably more efficient and experientially much worse. The self-checkout is the canonical example.

Self-checkouts are faster on paper. They reduce labour costs. Every metric you'd think to measure says they're better. And millions of people hate them! Many people feel patronised by them, feel abandoned by them, experience them as the moment the shop stopped caring.

The same pattern plays out in IVR phone trees, in AI support chatbots that can't route to a human, in ticketing systems that replace a one-minute conversation with a ten-field form. In every case, an efficiency lens approved the decision and a human experience lens would have caught the problem.

The efficiency lens dominates when PMs are spending their time building. The discovery lens of actually watching real humans struggle with something gets compressed. And the decision that looked good in the model turns out to be quietly corrosive in the world.

Efficiency is about cutting cost, becoming leaner. And this is a valuable thing for any business. But anybody can increase efficiency, but increasing efficiency whilst increasing value is real genius! And that’s the job of a PM.

When we only look through the lens of efficiency we can jump to the wrong solution. A classic example of this was the case of Eurostar looking to increase the speed of the London to Paris route. The problem was that the tracks on the British side of the channel couldn’t support the speed of the TGV trains, as they were shared with standard British trains. To upgrade the tracks would cost billions and save 40 mins on the journey. The rush to be efficient (quicker) forced a multi-billion decision. Was it worth it? Did people really care about saving an extra 40 mins? The argument was that people could get to work quicker. However, for a fraction of the cost they could have installed the best high-speed wifi and given it away for free, allowing people to work better when commuting, and for around £50 million they could have hired the world’s top male and female supermodels to serve champagne to guests for free and people would have actually asked for the train to slow down! The rush for efficiency masked the optimal user experience solution

Failure Mode 5: The lamppost problem

PMs use data the way a drunk uses a lamppost, for support and not illumination. I’ve stolen a saying that Marketing uses data the way a drunk uses a lamppost, for support and not illumination and replaced Marketing with PMs. I believe the same thing is happening with our adoption of AI; we are using it to support our biases and not as a learning assistant

The most important product questions such as what should we build that nobody is yet asking for? What experience are we destroying without knowing it? What does this customer actually value? etc. exist in the dark. They require discovery, not dashboards.

There's also a combinatorial problem. A single data point is thin. A single customer insight is thin. But when you put five customer insights, a competitive signal, a market trend, and a team's technical capability together, you get something qualitatively different. You get understanding. That synthesis is a PM skill. It is irreducibly human. It cannot be automated.

The James Watt problem (which is actually the solution)

James Watt invented the steam engine. But the thing that made the steam engine a commercial success wasn't just the invention itself, it was Watt's invention of a unit of measurement to make its value legible: Horsepower.

In the beginning, despite the obvious benefits, Watt struggled to sell the steam engine as it was expensive. It was difficult to articulate the value of it to factory owners. A factory owner knew how many horses they had and they knew what horses cost. When Watt said "this engine is equivalent to ten horses," the value was suddenly tangible, comparable, and purchasable. He didn't just build something extraordinary. He made it possible for ordinary people to understand the value of it and why they needed it.

That is framing and positioning. That is product management.

We are living in a moment of extraordinary technical capability. The question is never "can we build it?" The question is always: should we build it, for whom, and can we make them understand why it matters?

This is not anti-AI. I promise.

Let me be clear, because this is the part people misread.

AI is remarkable. I use it aggressively every day and so should you. The question is where it belongs in the PM workflow.

Where AI excels:

• Synthesis: a 90-minute customer interview transcript turned into themes in two minutes
• Competitive intelligence: Agents that digest competitor updates, pricing changes, new features, so you don't have to
• Feedback aggregation: NPS verbatims, support tickets, community posts, all surfaced as signal rather than noise
• Communication: First drafts of PRDs, specs, narratives, from structured inputs you provide

The heuristic I keep coming back to is: AI on the inside. Humans on the outside. AI is powerful inside the building. But, the discovery work such as going outside, talking to customers, sitting with the problem, and being in the world, is where human judgment is irreplaceable.

The PM who uses AI to synthesise ten customer interviews is better than the PM who doesn't use AI. But, the PM who uses AI instead of doing ten customer interviews has a faster process and a worse understanding.

Those two things look identical on a sprint board but they produce very different outcomes.

"No" is still the most powerful thing you can say

The PM who can't say no is not a PM, they're a backlog manager.

"No" is not an obstruction. It is the primary mechanism by which PMs allocate the scarcest resource in any organisation, the attention and energy of their team. Every feature built is a feature that didn't get built. Every sprint spent on a stakeholder request is a sprint not spent on discovery.

The PM who has done the actual discovery work, who knows what problems matter and why, is the PM who can say no credibly. Without discovery, you can't say no. You can only say: "I don't know either, so let's build it and see."

Which is, increasingly, what AI-accelerated teams are doing. Building more, discovering less, and accumulating a faster-growing pile of things that didn't matter.

What we owe the craft

Product management as a profession exists because building the wrong thing, even today, is expensive, and figuring out the right thing is hard.

That is still true. It is more true now, not less, because the cost of building has collapsed and the volume of things that could be built has exploded. The constraint is no longer engineering capacity, it is judgment.

Red Bull didn't win because they developed a better soft drink. It won because someone understood something true about human beings that the data didn't show and had the courage to act on it.

In a world where anyone can build, the scarce skill is knowing what's worth building.

That is the PM job. It always was. Don't let the tools of the moment make you forget it.

Glen Holmes is VP of Product at Qase. He has been doing this for longer than he'd like to admit and still occasionally ships things that don't move the metric.