Ship decisions should weigh how bad a failure is

Ravi Mehta gave product leaders a launch call. A checkout change lifts conversion 4%, but on older Android phones one in 500 orders gets a duplicate charge, which is refunded automatically within 24 hours. The options were to ship, exclude those devices, or hold Ravi on Product. Then he added a detail: the order is for $5,000. That customer could see $10,000 charged, maybe while rent is due Ravi on Product.

His point is that this bug was contained: the team knew the affected devices, the frequency and the fix. AI failures are harder to predict or contain. An AI travel planner's closed-restaurant recommendation might mean a short walk or a wasted 20-mile drive. A support agent only saves time if the cases it can't solve don't leave customers stuck Ravi on Product. "A product can perform well on average while creating experiences that customers find unacceptable" Ravi on Product. For launch reviews, that means scoring AI features on how bad and how contained their failures are, not only on aggregate metrics.

How OpenAI is earning enterprise trust in agents

On a16z, OpenAI's computer-use team described two layers of enterprise control:

• Deterministic sandboxes. Admins choose which apps and websites the agent can use a16z.
• Configurable human-in-the-loop rules. For example, a person must approve any download, upload or new website a16z.

An "auto approve" model reviews each action, and consumers and enterprises can set approval differently a16z. Small UX choices also build trust: a visible, animated cursor, and asking permission before using a new app or site a16z.

The team named the "boring" opportunities as the near-term targets: systems of record such as ERP and CRM, procurement spreadsheets, and month-end finance reconciliation across tools without APIs a16z. For product teams whose sites agents will visit, they flagged a business-model tension: agent ad impressions generally don't count as revenue. They also pointed to WebMCP as a standard that lets humans and agents use a site together a16z.

Prototypes over PRDs at Shopify

Shopify design lead Marvin Schwaibold says the process is still iterative, but he now uses Figma as a scratchpad and moves quickly to Codex and Claude Code to build demos he can test on his phone Y Combinator. Designers are "more the editors now than ever." AI can write code, "but it doesn't know what looks good" Y Combinator. Prototypes "are a million times more useful than any PRD" and win buy-in faster Y Combinator. They live in an admin playground, not production code Y Combinator.

A clone can copy your screens, not your users

Aakash Gupta wrote a one-page spec from Linear's public docs. In 3 minutes 28 seconds, Codex built a working issue tracker with five statuses, priorities, drag-and-drop and Cmd+K. It covered only part of Linear Aakash Gupta. What it lacked was the moat: issue history, existing users, GitHub sync, and Slack, Figma and Sentry integrations. His question for PMs: which of your features could an agent copy from a one-page spec this afternoon? Aakash Gupta

Practitioner threads

• The PM workload. A product lead with one junior PM averages 60 hours a week fielding every requirements request r/ProductManagement. One fix that worked for another PM: route all requests through a doc and triage once a day, which cut interruptions by at least half r/ProductManagement. A lead PM with eight developers works 40–45 hours, using AI for test scenarios, requirements from notes, and decks r/ProductManagement. Another PM noted the catch: if you own the results, you still have to review everything it produces r/ProductManagement.
• Diagnose activation before adding reach. A B2B founder's first week: 18 sign-ups, 7 finished setup, 5 used the product, 1 came back r/startups. Commenters pointed out that 11 of 18 never finished setup, so the biggest drop-off comes before setup r/startups. Another said to judge returns against how often users actually need the tool r/startups.
• AI search keeps old positioning alive. After a startup repositioned, AI still cited the old messaging. Fixing it meant updating owned pages, influencer content, bottom-of-funnel (BOFU) content and pages on other sites r/ProductMarketing. One PMM now audits AI snippets every sprint and found that rewritten comparison pages and integration guides were what pushed the old content down r/ProductMarketing.
• Agent security. A reply to Ryan Hoover noted that ClickFix attacks need a human to paste a command. Agents with shell access skip that step, so "a planted instruction file is the new fake prompt" @bgoutham.
• People. Peter Ellis is joining X to lead Product @petersellis.