The Product Desk

What Changed This Week

A developer opens Zed on June 15 and runs the same prompt she ran on June 14. The output is identical. The bill is not. Anthropic has announced that starting June 15, Claude usage through third-party tools (Conductor, Zed, Openclaw, T3 Code, Cline) gets a separate credit pool equal to the plan's dollar amount, with overage billed at full API rates. First-party Claude app and Claude Code usage is untouched. The 50% rate limit bump for two months is a grace period, not a permanent concession.

What this actually closes is the 70-90% implicit discount that many development teams quietly built into their unit economics. A developer who was effectively paying $20/month for Claude through a third-party harness now faces $200/month-equivalent metered usage for the same workload.

The era of subsidized AI inference through integrations is ending. If a product's AI capabilities route through third-party tools, model the cost impact this week, not after the June 15 switch.

ServiceNow Is the Public Version of Your Problem

ServiceNow's CDIO Kellie Romack confirmed her team burned through the entire full-year Anthropic budget before May 2026. She cannot say which users drove it or which workloads consumed it, because Anthropic does not ship the telemetry enterprise buyers expect. PagerDuty and National Life Group describe the same problem. National Life Group's Nimesh Mehta calls Anthropic 'great for consumer usage but not great for companies.'

Here is what the data actually shows. AI costs are unpredictable because usage scales with customer success. A feature pitched as saving two hours gets run eleven times a day instead of the three the pricing model assumed. Retention improves. Margin goes sideways, then down.

The Competitive Window OpenAI Just Opened

Sam Altman offered 2 months free Codex to enterprise customers who switch within 30 days, timed to Anthropic's moment of developer frustration. Criticism from Theo, Jeremy Howard, Matt Pocock, and Omar Sanseviero landed within hours of the Anthropic announcement. OpenAI lost the business adoption lead for the first time (32.3% vs 34.4% per Ramp) and is now buying it back with displacement pricing.

The Duolingo Calibration Point

Duolingo publicly acknowledged their blanket AI mandate failed. 20% of AI-generated content is unusable and requires human QC, and they reversed the policy. The planning implication is that human-in-the-loop capacity has to assume one in five AI outputs needs intervention. That is a permanent COGS line, not a quality issue that gets prompted away.

The Decision Framework

Three Platform Bets Landed the Same Week

SAP shipped a Knowledge Graph for agent context and stood up a €100M partner fund for Autonomous Enterprise. ServiceNow launched Action Fabric, which decouples workflow logic from the UI and exposes it through MCP servers so third-party AI agents can execute it. Notion released a Developer Platform with a markdown API, external data sync, agent tool building, and code execution, hosting Claude and Codex as 'teammates.'

Vendors do not stand up hundred-million-euro funds for features. They stand them up for platform bets they intend to defend for years. Three of the largest enterprise vendors landing in the same place in the same week is a commitment signal, not a coincidence.

A procurement manager at a Fortune 500 asked three vendors the same question this week: 'Can our agents call this directly, or do my people have to click through your UI?' Two didn't have an answer. The third moved to the next stage.

The Data Readiness Gap Creates a GTM Problem

Only 15% of organizations have the data foundation for agentic AI. They are spending millions anyway. Nearly half cite data quality and lineage as the primary blocker. The failure pattern is what teams tell themselves will not happen to them: buy, fail to activate, churn, blame vendor. Fivetran is positioning on the 'foundation' narrative. The counter-play is helping customers succeed across data maturity tiers instead of shipping capabilities into an unprepared market.

Microsoft's published agent memory architecture is the first real benchmark a PM can spec against: 400-500 memories, 97.2% retention precision, with consolidation, forgetting, and delayed maturation. Put those numbers in the PRD. The 'forgetting' mechanism is the counterintuitive one. Infinite context sounds like a feature and kills relevance on a predictable timeline.

59% of Token Volume Is Already Agentic

Vercel's AI Gateway data across 200,000+ production teams confirms agentic workloads carry the majority of token volume. Anthropic captures 61% of spend, with Opus doing the reasoning. Google captures 38% of volume, with Flash doing fast and cheap. Most large teams route across multiple providers by workload type. An architecture that still assumes request-response as the primary pattern is built for the minority consumption model.

What This Means for Your Product

The product decision is not 'build an agent.' It is whether the workflows the product owns can be invoked by an agent that is not yours, without a human clicking through the UI, by Q4. If the answer is no, the agent living in the buyer's stack will route around the product to reach the system of record directly. The product becomes a reporting surface on top of someone else's execution layer.

The work is smaller than the deck will suggest for most teams. A week of scoping, 2-4 weeks of build, assuming the underlying API is not already a mess. The window before this shows up in RFPs is 2-3 quarters.

The Worked Example

Elena Verna — former head of growth at Amplitude, Miro, Dropbox, and SurveyMonkey — now spends 90% of her time building at Lovable, almost no meetings. She shipped Lovable's enterprise pricing page to production herself. In a traditional org, that work needed a PM, a designer, engineers, and roughly a week of calendar time. She did it in hours.

Lovable has zero product managers. Engineers talk to users, write specs, ship code, and read the feedback themselves. The company is growing fast enough that the absence is not an oversight. It is the operating model. They are hiring Growth PMs in parallel to Verna, not reporting to her. That is not a growth team. It is a roster of autonomous operators.

The PM value proposition decomposes into three pillars: cross-functional coordination, customer and market judgment, and strategic prioritization. Pillar one is what fills a PM's calendar. It is also what AI-enabled flat orgs are eliminating.

The 'Average Intelligence' Reframe

AI does not make a PM world-class at design or engineering. It makes them average-to-good at everything at once. For a PM who already thinks across functions, that is an opening, but only if the saved hours go into shipping rather than coordinating other people who are shipping. The PMs who come through this shift look less like project managers and more like mini-GMs who prototype and iterate in the artifact directly.

Abridge Shows the Wedge-to-Platform Path

Abridge's three-act sequencing in healthcare answers the order-of-operations question: save time (documentation), then save money (prior auth and billing), then save lives (clinical decision support). Each act unlocks a different buyer, contract size, and release cadence. They compressed health system releases from quarterly/semi-annual to monthly, a 4-6x acceleration, by embedding 'clinician scientists' (MDs who are full-stack engineers) inside product teams.

The lesson is not the $5.3B valuation. The lesson is that the valuation followed a wedge a clinician could describe in one sentence: 'I edit two lines and sign instead of typing for forty minutes.' The platform conversation came after the wedge earned distribution.

The Diagnostic for Monday

The real call is not whether to unbundle the PM role. It is which of the four jobs — user research, prioritization, spec-writing, cross-team coordination — quietly goes missing the day the title goes away. Write that down before the reorg meeting, not after.