It’s the question doing the rounds right now: with AI collapsing development timelines from months to days, a good engineer with the right tools can go from napkin sketch to working prototype before lunch.

So why would you slow that down with customer interviews, workshops, story mapping, and all the usual product stuff?

We asked ourselves the same question, and for a moment, we thought we’d found a better way. Spoiler alert – we hadn’t.

What We Tried

Our bet was on forward-deployed engineers: senior, full-stack engineers embedded directly with clients, armed with AI-accelerated dev tools, building mind-blowing AI-powered products.

(At Propel, we partner with our clients — but if you’re in a product company or enterprise, insert “stakeholder” for "client".)

We embedded our engineers close to the problem, with the autonomy and tooling to build at pace. No handoffs or no waiting for specs, just talented engineers solving problems fast with AI driven solutions.

The theory was they could absorb the product thinking alongside the delivery.  Who needs a product manager when your engineer is this good?

What Actually Happened

Clients (Stakeholders) Will Always Want More

Here's something that shouldn't surprise anyone: when you embed with a client and start delivering visible value at speed, they want more.

Of course they do.

"Can we just add..." becomes a steady stream of enhancements, edge cases, and shiny new ideas.

But without someone (a product manager) whose job it is to hold the line on strategy and measure whether the outcome has been achieved, there's nobody consistently asking:

"is this actually the most important thing we should be building?"

Building something cool and watching a client's eyes light up is a lot of fun but doing everything because you can, fast, is no substitute for focus and prioritisation.

We found our teams doing genuinely impressive engineering work, that was, at times not focused on the core problem they'd been brought in to solve.

Engineer burnout is also a real risk when working this way.

We lost the discipline of breaking work down

Breaking down work from vision to goals, goals to epics, and epics to stories isn't process for process’s sake. It’s how you turn an idea into a sequence of thin, valuable slices that can be shipped, tested, and improved.

We didn’t do a great job of this, as it’s hard to do this without collaboration, and utlimately was deprioritised to allow more time on the tools, building.

We lost the estimation discipline that gives stakeholders (and ourselves) the ability to deliver predictability. And we lost the shared understanding that comes from a team aligning on what “done” actually means at each step.

Continuous discovery largely stopped

We also had a miss on the discipline of systematically checking whether the problems being solved were the right problems. Nobody was testing assumptions before working on solutions. The ongoing rhythm of discover, validate, ideate, validate again just wasn't happening.

The output was fast and seriously impressive. But if we were honest, we weren’t spending enough time on answering: Are we getting the most valuable pieces into users' hands first, or are we just... building?

What Should Product Management Actually Look Like Here?

Our view is that in a world of AI products (I'm talking about Gen AI), the fundamentals haven't actually changed. Discovery, slicing, collaboration, prioritisation, all matter even more when AI acclerates what you can do.

But the shape of the product manager role does need to shift.

With AI outputs being probabilistic, behaviour can drift over time, and quality is something you need to manage continuously. It's not a once off QA exercise.

The product manager’s job expands from deciding what to build to also understanding:

• what’s possible with today’s models
• what’s safe and trustworthy for users
• what needs guardrails, evals, and human oversight
• and what should never have been an AI feature in the first place

Knowing where AI actually adds value

The judgement muscle for product managers needs to be extended to know where AI genuinely helps and where it doesn't.

That means building a working understanding of what AI is good at today, what's on the horizon, and what's still mostly hype. It means recognising patterns, the common use cases where AI reliably delivers but also knowing the limits.

Not everything needs AI, and knowing the difference saves you from the "AI hammer looking for nails" trap.

Product managers need to understand the different ways to engage with AI, assess AI-specific risks (around reliability, bias, trust), and have a framework for evaluating opportunities that goes beyond "could we use AI here?" to "should we, and what's the realistic upside?"

Designing AI features that work as expected

As a Product Manager working with AI features, you need enough fluency to weigh up different models and approaches, when prompt engineering is enough, when you need retrieval-augmented generation (RAG), when you need structured workflows, when fine-tuning is worth the investment.

Answering these questions directly will shape the user experience, the cost structure, and how reliable the thing is in production.

You need to design for when it goes wrong. Traditional software either works or it breaks whereas AI features exist on a spectrum: sometimes the output is brilliant, sometimes less so... How the product handles that grey zone matters enormously. Confidence indicators “citations”, human-in-the-loop checkpoints are the difference between a feature people trust and one they abandon after a bad experience.

Shipping AI with quality

Quality means something new when AI is involved. The data going in matters as much as the code, your AI output is only as good as what you feed it. You need success metrics that go beyond usage to capture accuracy, reliability, and whether users actually trust the thing.

Evaluations (evals) are the feedback loop that drives continuous improvement. Without them, you're flying blind on whether your AI features are getting better or quietly degrading.

And bias, fairness, safety need to be baked into the product from day one, not bolted on before launch when someone remembers to check.

What We Took Away From All This

AI hasn't made product management obsolete. If anything, it's made the consequences of not doing product properly far more visible and far more immediate.

The speed AI gives you is real and it's genuinely exciting.

Forward-deployed engineers with AI superpowers can build remarkable things, but without someone anchoring that capability to the right problems, keeping discovery alive, breaking work into valuable slices, you end up with a lot of impressive output and not nearly enough meaningful outcomes.

The product managers who'll thrive in this world are the ones who pair their existing product skills of discovery, strategy, prioritisation, collaboration with genuine fluency in what AI can and can't do.

The role is a bigger one than it's ever been, so we have added an AI bootcamp for PMs to all our PM learning plans at Propel. We will be launching this for our clients in the coming weeks so if you are interested in getting access – email me at amy.johnson@propelventures.ai.