The Great AI Reversal: End of the PRD and the New Rules of Discovery

Milos Belcevic at UXDX Community: The Builder Mindset: Ship to Learn. Video: https://youtu.be/OHt1ZoYdLms

Readable transcript: edited from the recording's captions for readability (fillers and false starts removed, punctuation and section headings added). Wording is the speaker's own. Timestamps are positions in the video. Names marked [?] could not be verified against the audio.

AI has flipped the equation upside down

[00:00:08] Thanks so much, Rory, for the intro. It's really a pleasure being here, and as mentioned we will discuss how AI is changing the landscape of product management, and in particular discovery and documentation.

[00:00:22] So for decades we used to do discovery and documentation and then build, ideally in smaller increments. Discovery and writing documents was cheaper. Writing code was expensive. However, nowadays things are changing, and they're changing fast. AI has totally flipped the equation upside down. Show is today not just better than tell. Now it's faster and cheaper as well.

[00:00:52] So we will cover how and why product management is changing. We'll do some math, I promise you just simple math, on why the PRD is actually dead today. Then we'll discuss the new discovery loop, rapid prototyping, and rapid prototyping at scale, and then share some final thoughts, examples, and have a few minutes for a Q&A session.

[00:01:20] Before I continue, just briefly to introduce myself. As you heard, I'm Milos. I'm based in Belgrade in Serbia. I work as a product expert at Toptal, and recently I also joined a small startup, Rewardful, which is an affiliate network with clients like Clay and Lovable. I worked for employers and clients ranging from startups to global S&P 500 corporations.

[00:01:49] Overall, I have around 10 years in product. I love meeting new people and speaking at conferences, so I spoke at TEDx and more than 17 conferences in over five countries. I wrote a book on applying product management to personal life called Build Your Way. I'm also a PhD student. I'm exploring optimization, analytics, how we can apply AI and algorithms to make better product decisions. And I'm a lifelong learner and also a mentor, so I really like this whole idea of having a product-related community and sharing knowledge and exchanging experiences.

Why product management is changing

[00:02:36] So how and why is product management changing today? Well, we have very high expectations but fewer and fewer resources. So we need to achieve more with less, more and more. We are transitioning from a coordination role, where you have a PM who is more orchestrating the whole scenery, over to a PM who is a builder and also takes care of execution and ownership. As mentioned in the beginning, it's becoming much cheaper, easier and faster to build than ever before.

[00:03:18] However, the market requires speed. So nowadays it's almost hours, not even days or weeks, until you need to put something out there in production. And it also requires leaner teams. Organizations in this new alignment of things need PMs who execute. And we also have, which is super interesting to me, the rise of solo builders, indie hackers, solopreneurs, and people who are now empowered by AI and vibe coding in general to launch their own ideas and grow meaningful and impactful businesses much faster and, I would say, smarter than we could do before this awesome technology.

The old math versus the new math: why the PRD is dead

[00:04:10] Now is the part of the old math versus the new math, in general why the PRD is dead. So very simply put, I thought let's have a measurable way of saying this gives value, or this is cheap, this is expensive. And I called it the value efficiency score. Basically it's value, utility and impact divided by cost, so resources, expenditure, and time, which is duration to create the value.

[00:04:44] So in the old math, which was the high-friction era, code was high value but it was also expensive and it was slow. Documents were medium value, but they were low cost and fast time. This means they were 2.5 times higher on this value efficiency score. But as mentioned, AI is totally changing this math. Nowadays we have a new math in the Gen AI powered era. So code is now high value still, but it's also low cost and fast time. In fact it is two times higher on value efficiency score than doing documentation. Two times.

[00:05:33] We can pretend that nothing has changed. We can act as before and do things in a status quo, I could say even doing things blindly. Or we can act accordingly and adjust to this new reality we are living in.

The new discovery loop

[00:05:55] So speaking of adjusting to the new reality, I would go on to discuss product discovery and this new discovery loop that's happening. Some of you, or most of you, are likely familiar with this book and author. It's Teresa Torres. And if you haven't read the book, I highly recommend it. It's like the product management bible, or one of the bibles, which uncovers how we can conduct user research and product discovery, find out about different opportunities, and then run experiments, build prototypes and test these assumptions.

[00:06:40] There's a whole theory around it and I will not go into nitty-gritty details, but basically you have a product team that is talking to customers and then it identifies opportunities. These opportunities are aligned with outcomes that are again aligned with company strategy and the overall direction we are going to. Again, the product team finds small solutions, multiple solutions per opportunity, to test them out and see what would be the most meaningful outcome, or what we should build. As you have opportunities and solutions, it's called OST, opportunity solution tree. You can find out more in her book and on Product Talk.

[00:07:29] But now the robots are coming. So if you look at this flow of the continuous discovery, where we do regular discovery interviews, where we map out opportunities, where we run experiments, AI can be mixed into everything. So when it comes to conducting regular or weekly interviews, AI is now being used for note taking, note summarizing, and also even other data gathering at scale. We can go beyond just surveys. We can have our agents or bots actually do the interviews. Now, I know nowadays, or in the beginning, it might be a bit awkward user experience, so it's not very alive and thriving today. But a PM is super busy in any organization, and you maybe don't have time for five or 10 interviews per week, but AI can do thousands while you sleep. So it's interesting to see how things evolve there.

[00:08:43] When it comes to discovering opportunities, AI is great at identifying patterns and trends. So now, when we have all our meeting notes, when we have all this data, we can point AI and then see patterns and trends. How is our user base evolving over time? What are the churns, what are the most wanted or needed features, and so on. We can brainstorm in general, or brainstorm based on our data. So like, give me five feature ideas, or give me five solutions to this new trend we are seeing, or this new user need.

[00:09:31] We can do scenario playing out, and also customer voice discussion simulation. We can literally ask AI to pretend they are our user, such and such user, based on these meeting notes, and then instead of just analyzing the historical data we can literally have a natural talk with our users, or a digital twin of our user, or whatever you call it.

[00:10:02] Beyond that, again, for running experiments, we can brainstorm what experiments to run. We can do rapid prototyping, vibe coding, analytics. It's crazy. It's also becoming very conversational. Tools like PostHog now have their own AI agents where you can ask like, you know, what's the average lifetime of a user of our product, or more complex things. You don't even need to know SQL or Python or any sort of coding skills now. AI covers it pretty much. And then again, all of this can be done at scale. In reality the things are changing very drastically, and AI gives us a lot of possibilities and a lot of power when it comes to product discovery.

[00:11:04] Teresa Torres herself also acknowledged this, and beyond acknowledging, promotes using AI in day-to-day discovery. So this is the new model. We had a few slides ago the original opportunity solution tree with different stakeholders and so on. But now, in particular, she's also collaborating with Vistaly AI, which is a tool for this purpose. You can record the interviews and then have AI process it, analyze it and prefill a draft opportunity solution tree for you. So it's not manual, it's automated. It's not generic, it's relevant, because it's based on your discussions, your interviews and data you have.

[00:11:56] I would note that you don't necessarily need to use the tool Vistaly. Nowadays even Google Drive is launching their agents, so you can point an agent to your Google Drive notes and then have them populate a relevant OST. So beyond the tool itself, it's just an illustration of what you could achieve and how you can improve your daily or weekly discovery processes. And then one last side note is, if you see NotebookLM, that's a product by Google, and this whole new setup was also generated with AI. So AI is alive and thriving everywhere.

Rapid prototyping at scale

[00:12:51] The next section is about rapid prototyping at scale. So I just wanted to cover a few key moments here. We are seeing, from indie hackers to PMs in startups, PMs in large organizations, to huge old sort of organizations also changing now, and they're all doing AI-enabled development.

[00:13:29] So some of the things that are changing is this shorter loop. I mentioned this in the beginning, and I really believe in some cases now it's a matter of hours, or it can even be live. You don't even need to finish the call with your stakeholder. You can type in, wait a few minutes and get a new user flow or a new functionality to test out on the spot. Of course it's not production grade, but still it's a very, very rapid prototype. So we are having these immediate high-fidelity experiments, which is a lot better than using a proxy. So instead of describing to the user, how would you feel if we let you sign in with this or that, you actually just do it on the spot.

[00:14:26] There is another point, which is blurred role profiles. So in terms of creativity and building stuff, now you are not a coordinator, you're a builder. So you don't need to rely on historically scarce resources such as engineering, for at least the early iterations. But a lot of companies are now pushing for having vibe code ready for production, or like enterprise-grade quality. I don't think we are there yet, but AI is just getting better day by day, so soon enough we will be even there.

[00:15:06] We can scale our assumption testing. Previously you could have an assumption, write a scenario of a small experiment, then get engineers to develop it, then track it, then have a data analyst to sort of analyze it for you. Now you can do everything yourself in a few hours or less. And then I mentioned this outcome of a proxy when we discussed shorter loops. Prototypes are something live and valuable that we can test right away. So we don't need to discuss the potential new feature or whatever we are introducing to the product. We can play with it and work with our users, even co-build stuff and test it live.

Examples: Google, Duolingo, Coinbase

[00:16:04] Lastly, I have quite a few thoughts and examples I wanted to share with you. So in terms of examples: July, so summer last year, Google started asking PMs to vibe code at their job interviews, and they said, we're moving from writing to building-first culture. It's also interesting that the term vibe coding, or vibe code, was coined less than half a year before Google asked people to start using it on interviews. It was coined by Ilya Sutskever. I butchered his last name probably, I'm sorry. Who was one of the founding members of OpenAI and is a very big authority in the area. So we have this new trend of vibe coding, and then less than half a year later a FAANG company is asking for it from their potential PMs.

[00:17:09] One other company which is really awesome, and you maybe know this green owl that chases people to do their daily lesson, is Duolingo, the largest language learning and other skills learning app in the world. And also in the summer of last year their CEO asked all employees to vibe code a project, and within months closely 100% of the company did that. And it's interesting that they soon released a chess course which actually started as a vibe code project by employees. So it went beyond familiarization with the technology. It actually influenced the real product out there in the world.

[00:17:58] Again, summer last year, Coinbase's CEO fires engineers who didn't adopt AI. This one is a bit rough, and I would say not an ideal situation, but it's a strong signal of the changes that are happening. They say not a lot of people were affected. The CEO said, okay, I was maybe heavy-handed, and so on. But it's just showing how big this change is, how important it is, how crucial it is. And they also, on the positive side, host regular meetings with teams that share and learn together about using AI in new creative ways.

Merging roles, and some cautionary tales

[00:18:50] There's also more recent, from early 2026. There's another great resource, Lenny's Podcast, for all things product, another product bible. If you haven't checked it out, do check it out, there are many good episodes. There was a discussion with Marc Andreessen, from the Andreessen Horowitz company, about what's the future of the roles of product manager, engineer and designer. And he said we sort of have this Mexican standoff where the designer thinks he can do PMing and engineering, the engineer thinks he can do the designing, or she can do the designing and PMing, and the PM thinks they could do designing and engineering.

[00:19:42] And I talk about this at some other conferences, and I also think we will see these roles become closer and merge into something like product maker, product builder. So it will become a bit more similar. We might still need highly specialized people, someone who is training the models or optimizing algorithms, or some chief product strategist or something like that. But the vast majority of PMs, designers, developers will at least overlap, but I would bet on the roles merging.

[00:20:27] Next, again early this year, Google pushing for AI usage. So we need to be aware that Google is a company developing LLMs and other foundational models, so it's in their interest to have this marketing of having this story around AI being so transformative and great, et cetera, et cetera. However, it's really used in reality, and Google started pushing it on non-tech roles as well. And in some cases they even say that AI usage will factor into their performance review. So in and of itself, it's another strong signal that AI is very important and it's here to stay and it's changing all our roles.

[00:21:20] There are even more gruesome and not that ideal situations, but it's also an illustration of our current reality and the tech landscape. So in February, Jack Dorsey, who is the founder of Twitter and now founder and CEO of Block, he said in a memo that leaked that nearly half of the workforce will be made redundant. And it's not because the company is performing not well. The company is in fact profitable, it's growing well, everything is fine. But they see that with AI they need two times less people than they would need. And then after this announcement, stocks went 25% up. So it's yet another signal about the things.

[00:22:23] On the not so nice side, at least, there's AWS, which is like the core infrastructural network or platform we rely on heavily. They had two outages that were caused by AI, so early this year, or at least the rumors are saying so. But it's like we're living in a Silicon Valley episode where an AI agent deletes all the code or messes up the databases. Another rumor is that now at AWS junior developers are not allowed to push to production anything AI generated unless it's authorized by a senior peer. So they are adding some guardrails around it, obviously for security reasons, stability and performance.

[00:23:24] And I think this is the last fun fact. Ironically, one user of OpenClaw got her inbox totally deleted, nuked, by OpenClaw. And this is an AI security researcher from MIT. But basically, even though she prompted like, confirm before acting, even though she said don't do that, stop, abort, whatever, it so happened that Claw emptied her total inbox. So we do need to be enthusiastic about the new technology but also be a bit careful.

Final thoughts

[00:24:09] So, some final thoughts, and then we will wrap up. AI is the central driver of technological change we are seeing, especially Gen AI, agentic AI, digital twins, real-time analytics, and what's interesting to me, hyperpersonalization. So personal tutors, precision medicine, even custom silicon and so on. AI is deeply integrated into product development and into products themselves.

[00:24:40] The human role is shifting from execution more into strategy, oversight, framing the problems, which is good for product managers because those are some of our core skills. And AI will only get better with time, as it learns, as it's being developed, trained. But we are still very early and it will be messy, so yeah, hang on in there. That's all from me. Now, I'm not sure if we have some questions, we can discuss more, and also feel free to reach out on any social media and connect.

Q&A

[00:25:21] Rory: Excellent, thank you very much. That's kind of very topical and I think it's something that everybody is kind of trying to discover at the moment. Yeah, if anybody has any questions for Milos, please write them in on whichever platform you're watching. But one thing that I wanted to ask you, with the Mexican standoff, because it always reminds me a few years back where Tesla were making their cars, they weren't great quality, and there was this big thing, could Tesla figure out how to make cars quicker than Ford and the existing manufacturers could figure out tech? And it turned out that Tesla was better at making the cars. So in the Mexican standoff, do you think there's one role, if they're going to merge, that somebody has an advantage over the others?

[00:26:08] Milos: Yeah, that's a very good question. I'm a bit biased, but without joking, honestly I believe PMs are well placed, or they have this core skill of framing the problems, of strategy, which also of course developers have and designers have, but I think this core competency lies a bit more on the product side. So I think they will be better positioned in the standoff, but I feel like it's not a huge differentiator.

[00:26:44] Rory: And do you see, because as you were saying they're going to merge, do you think team sizes are going to shrink then, or do you think teams are going to stay the same but they'll do a lot more?

[00:26:56] Milos: Yeah, I talked about this on another opportunity, and I would also bet that teams will be smaller. So smaller teams made of high agency individuals. 70 to 80% of the team would be generalists, maybe like 10% would be someone highly specialized, if you need someone to tweak the algorithm, if you need someone to do really, really specialized work. Otherwise it's high agency generalists, and there will be a lot of AI agents. So we will have agents, and our maybe future overlords, in teams with us.

[00:27:38] Rory: Yeah. And I guess the last question then: when do the agents replace us? What is left for the people, now that you have OpenClaw, you have all of these things where the agents are just taking on more and more?

[00:27:52] Milos: I'm not even sure where it ends, to be honest. I would say hopefully strategy. Or maybe, if everything goes well, we go like retire early.

[00:28:05] Rory: Excellent. But on that happy note of everybody retiring early, that brings us to time. So I just want to thank you, Milos, for sharing, and I hope everybody out there enjoyed it as much as I did.