The new game of product with AI
AI crashes the price of the mechanical work of discovery, PRDs, and prioritization. What starts being worth more is your judgment about which problem to solve. And there's a catch that fools even experienced people: synthetic users don't replace real users.
Friday afternoon, and next quarter's roadmap needs to be locked by Monday morning. Before, that used to be your whole weekend: rereading the feedback board item by item, tabulating research, building the PRD, drawing the priority chart. Today AI does the first draft of all of that in minutes: it groups the tickets, drafts the PRD, and even suggests a "synthetic persona" meant to represent your user. The question that separates the expensive PM from the cheap one is no longer "do you know how to document?". It's: "do you know which problem is worth solving now, and does this persona AI made up actually look like your user, or is it the average of a generic user who doesn't exist?".
Friday, six in the evening. You need to close the month's variance analysis and still write the summary for Monday's board meeting. Before, that used to be your whole night: pull the numbers, check them, build the table, format it, write it up. Today AI does the first draft of all of that in minutes. The question that separates the expensive professional from the cheap one is no longer "do you know how to build it?". It's: "do you know what this number means, and what to do with it?".
You throw a sixty-page contract at AI and ask for an opinion on the termination clause's risk. In thirty seconds it comes back with an impeccable text, with the tone of someone who's certain, citing a statute and a precedent. You almost sign off on it. Except that precedent doesn't exist, and the cited statute is about something else.
Monday morning, and the week's calendar needs the five posts ready by noon. You ask AI for the draft, it delivers pretty texts in minutes. The problem shows up when you check the engagement rate it cited to justify the campaign's tone: a number that looks like a benchmark and came from nowhere.
You need to close the screening for a competitive opening and still prepare feedback for three performance reviews. Before, that used to eat your whole day. Today AI does the first cut and drafts the feedback in minutes, except it also hands you, with the same confidence as always, an offer-acceptance rate that "seems reasonable" for the market, without ever having looked at your real funnel, which is a lot tighter.
Monday morning, and the squad wants to know which feature makes the next cycle. You throw the whole feedback board at AI and ask for a prioritization. In minutes it hands back a pretty RICE score for each item, along with a "synthetic user" quote saying how much they'd miss feature X. It looks like research. It isn't. That quote never came out of anyone's mouth: it's a statistical average dressed up in interview quotation marks. The question that separates the PM who understands AI from the one who just uses the tool is no longer "do you know how to run RICE?". It's: "do you know how to tell the real pattern that came from your user base apart from the average pattern AI made up to fill the gap?".
Friday afternoon, and you need to update the forecast and send three pending proposals. Before, that used to be your whole night. Today AI generates the draft of all of that in minutes, but it also applies, with the same confidence as always, a proposal-to-close conversion rate that's a market average, not your CRM's actual funnel.
Start of shift, and you need to understand why the SLA blew up yesterday and adjust the operation for today. Before, that used to be hours cross-referencing delivery spreadsheets, building the dashboard. Today AI pulls the data and drafts the report in minutes, except it also estimates, with the same air of certainty, a damage rate that doesn't match your operation, which is running worse.
The day before an audit, and you need to close the risk map and review whether the data-privacy policies match actual practice. Before, that used to be days reading control after control. Today AI scans the documents and drafts the matrix in minutes, but it also assumes, with the same confidence as always, an enforcement probability that "seems moderate" for a generic case, without knowing about the notices you already received this year.
Friday night, and there's a stuck deploy and a production incident to postmortem. Before, that used to be the whole small hours reading logs. Today AI pulls the logs and suggests the patch in minutes, but it also points to, with the same air of certainty, a root cause that "seems plausible" for any similar stack, without having cross-checked your real deploy history.
End of sprint, and you need to synthesize the research interviews before usability testing. Before, that used to be hours transcribing and grouping findings. Today AI summarizes the interviews in minutes, except it also assumes, with the same confidence as always, a drop-off point that "seems obvious" for any generic flow, without having seen your latest real user research.
Sunday night, and the board wants the expansion analysis ready for Tuesday. Before, that used to be the whole weekend cross-referencing competitor data, running scenarios, building the deck. Today AI drafts a first version in minutes, TAM, SAM, and adoption curve all ready. The problem is it also hands you, with the same air of certainty, a penetration rate it made up because it "seems reasonable" for a generic market, without ever having seen that your last launch converted a lot less.
Man, let me start with a somewhat uncomfortable truth about product. A good chunk of what made you proud your whole career, reading all the research, tabulating feedback, writing the PRD from scratch, building the priority chart, AI now does in minutes. Think about it: does that scare you or free you? The right answer depends on you understanding which part of your work became a commodity and which part became worth more. This lesson settles that account.
The core idea of this lesson. AI crashes the price of the mechanical part of product discovery: grouping feedback, drafting the PRD, calculating the priority score, formatting the document. That's good news, because more of you is left over for the part AI doesn't do: deciding which problem is worth solving now, understanding what the user really means, and signing off on the roadmap decision. But product has a catch that fools even experienced people: AI loves offering you a "synthetic user", a persona or a quote that sounds like research and isn't. Synthetic users don't replace real users. This track teaches you to operate in this new game.
01What AI commoditizes (and why that's good news)
Let's call it what it is. These things, which used to take up your hours, AI already does fast and cheap:
- Grouping raw feedback, tickets, NPS responses, review comments, all together in a first pass.
- The first read: "what keeps repeating in this batch of feedback?".
- The PRD draft: structure, sections, the document's skeleton.
- The priority calculation: running RICE or whichever framework you choose on top of numbers you supply.
The big issue here is simple: when mechanical work gets cheap, it stops being your edge. The PM who only knew how to organize the board loses value. That sounds harsh, but there's another side, and it's the good side.
What shifts in your day:
Here's the good news: the time you used to spend on the mechanical doesn't vanish, it shifts to the part that's worth more. Whoever understands this stops fearing AI and starts using it to climb altitude. Fair enough?
02What's still yours (and got more expensive)
Here's the part AI doesn't do for you, and precisely because of that, it's now worth more:
- Which problem is worth solving. AI groups the feedback, but choosing which pain point makes this cycle's roadmap is a reading of strategy, and it's yours.
- The "so what?". The metric dropped, the feature didn't catch on. So what? What does that mean for the product, what gets prioritized because of it. AI describes; you interpret and decide.
- The roadmap decision. AI recommends, but whoever signs off on what makes the next cycle, and answers for it if it goes wrong, is a person. Always.
- The narrative for whoever approves. Translating the problem into a story the team and leadership understand and act on. That's influence, and influence doesn't commoditize.
Notice that all of this is judgment, not organizing. And judgment is exactly what gets expensive when organizing gets cheap. From here on, your value lives much more in "which problem to solve and why" than in "I managed to build the document".
03Product's catch: synthetic users aren't real users
Now the part I can't let you forget, because in product it costs dearly in a subtle way. Asking AI to simulate a user, or generate a "synthetic persona" to test an idea fast, sounds like a brilliant shortcut. And it sometimes even helps organize your thinking. The problem is confusing that exercise with real research.
A synthetic persona is, in practice, the statistical average of what millions of people wrote on the internet, predicted word by word. It tends to hand you the consensus opinion, the most probable answer, the most common case. What you most need in product discovery is usually exactly the opposite: the exception nobody had noticed, the specific pain point of your niche, the odd phrase a real user uses that reveals the problem in a way no average captures.
Learn more: why the synthetic persona converges on the average
It's worth understanding the mechanism, because it changes how you use the tool. A language model predicts the most probable answer given everything it has seen. When you ask it to "simulate a user of my product reacting to this feature", it doesn't have access to your real user: it builds a plausible answer from the average pattern of similar texts that exist in the world. That's great for drafting a hypothesis or practicing how you'll run an interview. It's dangerous when you treat that answer as if it were research data, because it systematically pushes toward the center, toward common sense, and erases exactly the divergence that's usually the insight worth money.
That's why the golden rule of this track, which you'll see repeated in every choreography: synthetic users are for exploring a hypothesis and rehearsing a question; product decisions rest on real users, always. It's not distrust of the tool, it's understanding what it's for. AI speeds up the research draft; you make sure the voice that decides the roadmap is a voice that actually exists.
04The map of this track
This lesson was the framing. The rest of the module is hands-on, installing systems into your real work, one at a time:
- Connecting AI to your product's context: backlog, analytics, and research, each in its own way.
- From raw feedback to the insight that becomes a decision.
- The PRD that comes out written and ready for the agent to execute.
- Prioritization with real evidence, not with AI's guesswork.
- From draft to testable prototype, in hours.
- The metric audit, for when AI makes up an adoption number.
- And, at the end, your product OS: the library of prompts, templates, and agents that keeps working for you.
Each one takes a task you already do and redesigns it. By the end, you won't have learned about AI, you'll have installed AI into your way of working. Shall we?
Do it now
Take a product deliverable you did this week, your real task or another one (a PRD, a prioritization, a research synthesis). Grab a sheet of paper and split it into two columns:
- Mechanical (what AI would do for you): grouping feedback, drafting the document, calculating the score.
- Judgment (what's still yours): which problem you chose to solve, what the pattern meant, the decision, how you communicated it.
Now look at the proportion. How much of your time went to the left column? That's exactly the time this track is going to give you back, for you to spend on the right column, which is the one that pays your salary.
Practice
1. In the new game of product, what loses the most value as a professional differentiator?
2. AI generated a 'synthetic persona' that quotes, in interview-style quotation marks, how much they'd miss a feature. What's the correct read?
For the board
On what got cheapreading the research, tabulating feedback, drafting the PRD, building the priority chart.
On what got expensivedeciding what gets in, defending the why, and answering for the roadmap.
On the synthetic userit is for exploring hypotheses and rehearsing questions. Product decisions rest on real users, always.
Thanks for the feedback. It helps sharpen the next lesson.