Berlin was the first stop for our next series of Venatrix Product Room. The attendees spanned construction, marketplace, fintech, ad tech, and thirty years of product leadership in the room - different sectors, same reckoning with what AI actually changes day to day.
Two development philosophies are splitting teams right down the middle. Some are running fast, throw-it-at-the-wall builds for anything uncertain - Lovable-style tools, ship and see. Others are doing the opposite: detailed, old-school spec engineering for anything complex, because agentic coding needs guardrails or it wanders. Scrum, meanwhile, is quietly breaking down across the group. Kanban's winning out as the better fit for agentic workflows, where work doesn't arrive in tidy two-week chunks anymore.
One CPO laid out a pipeline that made the rest of the room go quiet: a structured form captures intent, ROI estimate and customer scope, AI turns that into an Amazon-style working-backwards press release, the press release becomes a product brief, the brief becomes a tech spec, and agents auto-generate Jira epics and assign them straight to a coding agent. A separate review agent, a QA agent and a monitoring agent are already live in production. The only piece not yet wired up is connecting the spec agents directly to the coding pipeline - and that's next. Another product director in the room is building toward the same end state from a different angle: business stakeholder makes a wish, agents handle research, spec, dev and testing, humans just sit at the review gates.
The speed gains are no longer theoretical. PMs across the group are now producing high-fidelity, data-connected prototypes in 2-3 days for customer testing, work that used to swallow a full sprint. But almost nobody pretended the bottleneck had actually moved - it's sat down squarely on PR review. Some engineers are now spending the majority of their week reading code rather than writing it, and there's no industry consensus on fixing that yet. Internal tools stay behind VPN until they clear penetration testing; nothing regulated gets to skip the process. "We can't just go wild and trust the output," as one leader in the room put it - EU rules were framed by the group as both a genuine safety net and a real drag on velocity.
Measurement is still catching up to the tooling, and everyone admitted it. Velocity by story points, percentage of code generated by AI, headcount-to-revenue ratio, one team doubling its data provider onboarding per quarter - the metrics exist, but story points are still being set against pre-AI complexity baselines, and stakeholder perception of "faster" is tracked almost entirely on gut feel. Two separate leaders called their own leadership's 10x productivity targets poorly defined and, bluntly, frustrating. Token cost hasn't bitten yet for most of the room, but it's on the radar as the next line item to watch.
Hiring is where the room split hardest, and neither side backed down. One camp argued domain knowledge is non-negotiable now, because context quality is what determines whether the agent's output is any good - you can't judge what you don't understand. The other argued attitude wins outright: scrappiness, ownership, communication, with domain knowledge two months and an AI research tool away for anyone motivated enough. Either way, the "silent coder" profile is on its way out, replaced by forward-deployed engineers sitting closer to the customer. PM-to-engineer ratios are already moving - one leader is now covering 2.5 teams solo where it used to be strictly one-to-one - and one innovation function in the room runs on just one PM plus two senior AI-native engineers, hacking fast and proving value before anything gets productised. Ops and business teams are increasingly building their own tools via agents too, quietly cutting their dependency on the product roadmap altogether.
The line everyone in the room agreed on: when Google reportedly used AI to analyse Gmail tone for performance reviews, that's not innovation, that's a privacy violation wearing a UI. And under all of it sat the same unspoken question nobody fully answered - is this pace actually sustainable for the people running it, or are we all just riding it until it isn't?
Berlin's done. UK and US findings on the same topic are taking place throughout October and November.
Key takeaways:
- Agentic pipelines are moving from theory to production - intent capture through to auto-generated Jira epics and assigned coding agents, with review, QA and monitoring agents already live
- Prototyping timelines have collapsed to 2-3 days, but PR review is now the real bottleneck, absorbing most engineers' time
- ROI measurement lags the tooling - story points, AI-generated code percentage, and headcount-to-revenue are in use, but 10x targets from leadership were called poorly defined
- Hiring is split between domain-knowledge-first and attitude-first camps, but both agree the "silent coder" profile is dying in favour of forward-deployed engineers
- PM-to-engineer ratios are stretching (up to 2.5 teams per PM), and lean innovation squads (1 APM + 2 engineers) are proving models fast before scaling
- EU regulation is viewed as both a safety net and a genuine velocity constraint, especially for fintech
- Privacy remains a hard line - AI-driven tone analysis of employee communications was cited as a clear red flag, not innovation
