The Product Desk

What Changed This Week

A developer opened Cursor on Monday and shipped his usual amount of Claude-assisted code. Starting June 15, that same workday will cost roughly ten times more. Anthropic is moving third-party harnesses (Cursor, Cline, Aider, OpenCode, Zed, T3 Code) into a separate credit pool equal to the plan's dollar value, with overages at full API rates. Developers who were effectively getting 70-90% discounts on Claude inference through these tools see per-developer costs jump by an order of magnitude.

The pitch is "aligning pricing with usage." What is being done is closing an arbitrage ahead of Anthropic's likely October 2026 IPO, because subsidized power users do not produce the revenue-per-user metrics public markets price on. Model at least one more adjustment before October.

ServiceNow Is the Cautionary Tale

ServiceNow's CDIO Kellie Romack confirmed her team burned the entire full-year Anthropic budget by May 2026. She cannot say which users or workloads drove it, because Anthropic does not ship the telemetry to answer that question. PagerDuty and National Life Group report the same gap. National Life's Nimesh Mehta called Anthropic "great for consumer usage but not great for companies."

AI costs are structurally unpredictable, and the model providers have not built the instrumentation customers need to govern them.

OpenAI's Counter-Move Has a 30-Day Clock

Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. This is displacement pricing dressed as a goodwill gesture. The Ramp data tells you why the clock is 30 days and not 90: Anthropic at 34.4% versus OpenAI at 32.3% in business adoption means OpenAI lost the business adoption lead for the first time, and the cheapest moment to win an account back is the week the incumbent raises prices.

The Decision Framework

The Infrastructure Gap You Need to Close

ServiceNow built an internal AI Control Tower with a dedicated person watching Anthropic consumption. Most teams cannot attribute inference cost per customer, per feature, per use case. That instrumentation sprint is prerequisite work before shipping the next AI capability, not a follow-up ticket. Skip it and the team becomes ServiceNow in January: confident the budget is fine until May.

The Headless Enterprise Stack Lands in Production

SAP launched a €100M partner fund for its Autonomous Enterprise initiative plus a Knowledge Graph for agent context. ServiceNow released Action Fabric, decoupling workflow logic from UI and exposing it via MCP servers for third-party AI agent execution. Notion launched a Developer Platform with markdown API, external data sync, agent tool building, and plans to host Claude/Codex as 'teammates.'

The bet underneath all three announcements is the same: workflows callable by an agent without a human clicking through the UI. Hundred-million-euro partner funds don't get stood up for features. They get stood up for platform commitments the vendor plans to defend for years.

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

What the Production Data Actually Shows

Vercel's AI Gateway data across 200,000+ teams shows 59% of all token volume now flows through agentic workloads. Anthropic captures 61% of AI spend (Opus for heavy reasoning) while Google captures 38% of token volume (Flash for cheap, fast tasks). Most large teams route across multiple providers. Buying criteria lag the interaction model by a quarter or two, which is the window product teams have before RFP language catches up.

Separate the Pitch From the Work

The pitch is "agent-ready platform." The work is making the core workflows of the product invokable by an agent without a human in the seat. The window before this shows up in RFPs is 2-3 quarters. If the buyer's agent can't call the product directly by Q4, the agent living in their stack will route around it to a vendor whose workflows it can call. The build is smaller than most teams estimate: a week of scoping, 2-4 weeks of build for an MCP-compatible headless layer, assuming the underlying API isn't a mess.

The Forcing Function

Pull the last twenty support tickets and feature requests from the top decile of accounts. Count how many assume a human in the seat versus how many assume an agent or integration is doing the work. If that ratio has moved even ten points toward agents in the last two quarters, the headless layer is not a platform bet. It is a retention bet, and the clock is the next renewal cycle.

Duolingo Reversed Its AI Mandate

A team lead at Duolingo logged into the AI tool every morning for a quarter because the mandate told her to. She pasted in tasks she could already do, accepted outputs she would have written herself, and shipped on the same cadence as before. Duolingo's CEO eventually said the quiet part out loud: the mandate produced performative adoption without productivity gains. The harder number is the one that matters for capacity planning. AI content at scale yields roughly 20% unusable output that requires human quality control. They reversed the policy. The lesson for anyone setting team AI goals is to measure cycle time and output quality, not tool logins.

Persona Drift Is a Cliff, Not a Slope

Research from Li et al. (COLM 2024) quantifies what most teams have not measured: significant drift within 8 rounds of dialogue. The mechanism is attention decay. As context grows, the system prompt's influence weakens proportionally. Demos run 3-4 turns and look great. Power users with 15+ turn sessions are using a different product.

The cheap detection pattern is to embed a distinctive behavioral marker in the system prompt, a signature opener or framing device, and monitor for its disappearance in production logs. When it vanishes, the persona has drifted. That signal is the canary before a formal eval suite is worth the investment.

Amazon's AI Search Reveals the Trust Trap

Amazon merged Rufus into core search and rebranded it Alexa. Real testing shows results are "somewhat hit and miss." When AI is right 60% and wrong 40%, users cannot tell which case they are in without doing the work the AI was supposed to save. So they do the work anyway. The feature does not reduce effort. It adds a step.

The failure mode that matters is silent wrongness. Users who do not notice until after they have acted are the ones who churn. A feature that replaces an existing step and fails invisibly erodes trust, which is the budget AI features spend fastest.

The Counter-Signal: GPT-5.5 Crossed a Bar

GPT-5.5 Instant reduced hallucinations by 52.5% versus GPT-5.3 Instant and is now the free-tier ChatGPT default. Many teams have specific internal thresholds. Above that bar, AI features get blocked by legal or brand review. A 52.5% reduction in one generation potentially moves features from 'parked' to 'acceptable with guardrails.' The interesting question is which features in the shelved pile were blocked specifically on hallucination rate versus other reliability concerns.

Design Implications

• Use 20% as the human-in-the-loop capacity planning assumption until internal data exists
• Test multi-turn AI interactions at 8+ rounds and add it to acceptance criteria
• Ship AI in the companion slot next to the existing workflow, not the replacement slot, until confidence signals are reliable
• Default to the 'ambient over alert' pattern: AI that stays silent with high-precision interruption beats a notification machine

One Person Shipped What Used to Take a Team

Elena Verna — former head of growth at Amplitude, Miro, and Dropbox — now spends 90% of her time building at Lovable in a pure IC role. She pushed Lovable's enterprise pricing page to production herself. In a traditional org, that ticket pulls in a PM, a designer, a couple of engineers, and about a week of calendar time. Lovable has zero product managers on staff.

The company is hiring Growth PMs in parallel to Verna, not under her. That is not a growth team. It is a roster of autonomous operators who happen to share a Slack.

Why This Isn't Just a Startup Story

The PM job decomposes into three things people actually pay for: cross-functional coordination, customer and market judgment, and strategic prioritization. Pillar one is what AI-enabled flat orgs are eliminating. When one operator can design the page and ship the page, the coordinator stops being enablement and starts being latency.

Ravi Mehta's 'average intelligence' framing is the useful lens here. AI does not turn a PM into a world-class designer or a world-class engineer. It makes them average-to-good at everything at once. For a PM who already thinks across functions, that is leverage. Only if the reclaimed hours go into shipping instead of standups.

The PMs who survive this shift look less like project managers and more like mini-GMs who prototype and iterate directly. The question is not whether to unbundle the role — it's which of the four jobs goes missing the day the title goes away.

Claude Code's Autonomous Mode Makes This Concrete

Claude Code's new /goal command lets an engineer type a measurable objective, walk away, and come back to a finished task or a stopped run. A separate Haiku model grades completion against user-defined criteria. That evaluator-as-judge pattern — a lightweight model deciding whether the worker model actually succeeded — is the architectural primitive that makes unattended work tractable.

The sprint exercise is concrete. Open the current backlog and mark every ticket with clear input, verifiable output, and bounded scope. Add up the engineering hours sitting in that pile. That is the capacity autonomous mode unlocks. It is also the pile where the PM's coordination role was already close to zero.

The Diagnostic

Two questions for Monday. First: of the work shipped last quarter, how much of the PM contribution was judgment about what to build versus coordination of the people building it? Second: if the coordination half went to zero tomorrow, does the judgment half justify the role? If the answer to the second is yes, the job gets better. If no, the job gets done by someone who looks like Verna.