We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Big Ideas
Treat AI spend as a team-level investment, not a token score. A new product-management essay argues that “return on tokens” repeats the old hours/capacity mistake: teams focus on measurable inputs, favor short-term attributable use cases, and confuse ease of measurement with value. The better unit is the team: define its cost, causal model, demand, durable lane, lifecycle expectations, and validated leading proxies; use flow metrics for improvement, not investment attribution. For each AI bet, state what bottleneck it reduces, what downstream cost it may create, what durable asset it builds, and what evidence would change funding. ROI should be the output of those hypotheses, not the hypothesis itself.
Tactical Playbook
When CTR is healthy but demos are zero, recruit conversations before running a survey. One B2B SaaS campaign had decent CTR and CPC but no demos; the operator wanted 10–20 conversations with the exact buyer before spending more. The recommended sequence is 8–15 incentivized calls—prioritizing people who clicked but did not book—then a survey to measure how common the repeated objections are. Zero demos may indicate that the ad and landing-page promises do not match.
For a first or only PM, map the decision system before writing a roadmap. In a Series B onboarding thread, the useful first move was to talk to everyone, learn workflows and political hierarchies, and build trust that you can discover the “money maker.” Otherwise, the role can collapse into forwarding decisions to engineering.
Case Studies & Lessons
Fin’s AI rollout shows why productivity metrics are not the outcome. The company reported tripling PRs per person per month across R&D; 94% of PRs were Claude-authored, 19% auto-approved, product changes doubled, shipping was 39% faster, and nine major launches shipped in two months. The speaker called PR throughput crude; the strategic shift is that agents own the middle, moving PMs further upstream and making problem framing, decision quality, and product cohesion the new constraints. Apply the lesson with clear briefs, success criteria, deliberate pause points, and shared quality standards before scaling output.
Stripe channels agent capacity into customer value. Stripe says Minion-created PRs rose from roughly 1,200 per week to about 7,000, or 30% of PRs; Stripe Projects was led by a PM and senior engineer in a few weeks, and an eight-person team was doing three times more. Its stated alternative to AI-driven cost cutting is to work through unmet user asks faster. To scale quality, teams are told to use PII-free simulated accounts with realistic disputes, refunds, and seasonality—not just rely on review after shipping.
Payroll is a compliance product, not a feature. Practitioners responding to an in-house payroll question warn that US payroll spans 50 states and is either compliant or not; the inherited burden is filings, support, and maintenance after customers depend on it. First verify that customers need the capability; if embedded UX matters, one proposed route is an API partner that owns filings, compliance, and payroll operations.
Career Corner
Positive performance feedback is not the same as promotion availability. A promotion thread describes capped slots and forced curves; one candidate was one of three people deemed ready for a single opening. HR-safe feedback may require naming development areas rather than revealing that the constraint was simply slot availability. Separate what you must improve from whether a promotion slot exists, and ask what evidence and timing would actually change the decision.
Tools & Resources
Use graph engineering selectively. A 100M-token experiment tested four graph designs on 40 PM tasks: graphs won 25, but a single-node prompt won 15, including TLDR and email reply. The useful patterns were task-shaped: an assembly line for launch kits, a backward chain from metrics for instrumentation, and parallel validity checks followed by a checker that re-derives experiment numbers. The linked 40-design library is free to founding newsletter subscribers.