Robert Ta shows how to use AI agents to extract product requirements from customer call transcripts, then validate them against real codebase data. A requirement synthesizer pulls candidate features from recordings, the PM reviews and annotates each one, and engineering gets requirements grounded in actual performance metrics and operational constraints.
A really interesting thing that Jonathan and I have a very opinionated take on is for any product management function, there's this like, what do you do as a product manager? You spit out what is the requirement in a prioritized list of what is most economically valuable to the business at the intersection of product strategy and solving for the customer's jobs to be done. Like, that's kind of like your output. If you're thinking of like, I am an agent, what's my input? Fucking everything. Output, strategy, strategic prioritized cutthroat list of stuff that is yes and stuff that's like, fuck no, we're not doing that. That's the whole PM job. And it's a lot, it's a lot to make sense of on the left. And then on the right is a very succinct, what are we handing off to engineering? And how are we aligning the teams around this? It's a tough job. So how do you compound that? Me and Jonathan have an opinion to take where as a product manager making decisions, and now with AI, we're able to make more sense of things. And we want to drive higher quality decisions through the pipeline of like prioritization and then work of getting the software out the door. We think of like, Okay, a PM's gotta judge the artifacts. They gotta judge the requirements that come through from the sensemaking in particular. So we have this requirement synthesizer agent that we work with that takes in any transcript on the left. And this in particular, you're seeing this requirement, server-side reading lifecycle instrumentation and person state flags for retention automation. We could see like what part of a customer transcript this came from and who said that so that we know the stakeholder asked for this. This is our synthesized requirement and I, as the PM, judge if I have like 100 of these going on, right? I wanna easily make sense of this, see the agent's reasoning chain, see its provenance, and then I could actually go and annotate that and, or I could approve it. I could like add things to this, like we should include more requirements, specific requirements around the API. So this server work against the Implicit server performance requirements that are not named, but we have ground truth on from the codebase. So yeah, for the requirement for that, I'm just like, okay, I understand that that's a gamification or server-side instrumentation thing. Cool. Like I get that. Okay. I, I think when I pass it downstream to engineering or a coding agent to execute the work, we need to ground it in a good level of performance requirements based on the volume and load that we typically see. So that's one way that I'm thinking about this from a product management function of I'm using coding agents. I need to know what agents are doing. I need to be able to annotate their runs and I need the runs to make it so next runs are better.
