One of the most awesome perks you get in A8c is to work with soooo interesting people, clever and always ready to help and to share their thoughts. Like George Jen. He said in a recent post about the Product Manager job that the work is shifting from PRD to implementation, while judgment remains stubbornly human. I think that is right, but I’d add a nuance: judgment only helps if it is grounded in a real understanding of the problem.
I’m at this moment finishing the delivery of a data artifact on the occasion of the Radical Speed Month (in a few words, engineers, designers, and product managers pair up and pick a challenge and try to fix it in one month with the help of AI, no limits, no teams, no team leaders…). I decided to go further and be bold, working in a domain that was new to me, and very strategic for my company, WooPayments, So I converted the challenge into two: being able to deliver something that really moves the needle from scratch, and doing it in a very sophisticated environment where I started with absolutely no business knowledge other than as a user.
Agents release time you can use to explore and think
So, I’ve spent a Radical Speed Month working in a domain that was new to me. AI absolutely helped. Models handled a lot of the heavy lifting: data profiling, SQL and code generation, signal exploration, and documentation drafts. But the biggest benefit was not just speed. It was that this freed up something more valuable than compute: time to think and time for meaningful conversations with the key persons. And thanks to the AI, I was able to build a corpus of knowledge that was disseminated among lots of different actors and fairly undocumented systems (I guess this is the curse of big startups)
A significant part of what made this phase work was not the code, but the growing map of people across teams who gave us context that no model could have surfaced alone (and I tried). The time also went into conversations with experts, stakeholders, and, as I called them, key adopters to make sure what we build is aligned with what they actually need — because there is no impact without adoption. The relationship map we built to understand where each piece of knowledge lives ended up feeling almost bigger than the model itself. Honestly, Yep, you guessed it. That map is on an Excalidraw over Obsidian drawing that is connected with data sources, people’s profiles, reflections, and pieces of data that the AI agents can explore, use, and enrich.
You want to think beforehand
This mattered because one of the key lessons from the project was deceptively simple: before building a model, you need to be sure of the what and the why, and the agents will fill in the gaps. That line is funny because it is true. In data work, labels often look more solid than they really are. A field name, a status, a threshold, a category: they can look objective while actually being rough operational shortcuts. Useful shortcuts, sometimes. But shortcuts nonetheless. And if you automate on top of them before understanding what they mean in context, you do not just risk being wrong. You risk becoming wrong at scale, with excellent formatting.
That is why George’s point landed so strongly for me. Yes, judgment is the work that stays. But judgment is not just a trait you bring to the keyboard. It is something you earn through contact with the domain: through conversations, edge cases, awkward definitions, and the slow discovery that reality is always messier than the schema.
Morale: You appropriate the problem and then you put agents to work
So I’ve come away from this RSM thinking that “prompting well” is a downstream skill. The upstream skill is understanding well. The better the tools get at producing, the more important it becomes to know what should be produced, what should not, and which assumptions are carrying more weight than they appear to.
The agents can do more of the work. But you still own the problem.
Leave a Reply