The Product Desk

The Arbitrage Window Closes June 15

A staff engineer opened Cursor this morning and ran the same agent loop she has run every day for six months. Tomorrow that loop costs her team roughly ten times more. Anthropic announced that every Claude subscription now includes API credits equal to the plan's dollar amount, which sounds generous until you separate the pitch from the thing being done. Teams using Claude through third-party harnesses (Cursor, Cline, OpenCode, Zed) had been getting effective 70-90% discounts to API pricing. That subsidy is what is ending. Anthropic hired a CFO and is likely targeting an October 2026 IPO, and the old per-user economics do not produce numbers a public market wants to underwrite.

If a team adopted Claude through any non-Anthropic harness in the last year, the per-developer cost assumption in the budget deck is now wrong by roughly 10x.

ServiceNow Is the Cautionary Tale

ServiceNow's CDIO Kellie Romack watched her team's full-year Anthropic budget get consumed before mid-2026. She cannot tell which users drove it, or which workloads, because the telemetry to answer those questions does not ship from the provider. PagerDuty and National Life Group report the same gap. National Life's Nimesh Mehta puts it plainly: Anthropic is 'great for consumer usage but not great for companies.' ServiceNow's response was to build an internal AI Control Tower and staff it with a dedicated person. Most teams do not have that headcount to spare.

OpenAI's Counter-Play Has a 30-Day Shot Clock

Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. That is displacement pricing aimed at a specific user moment, namely the week the Cursor invoice changes. Ramp data has Anthropic at 34.4% of business AI spend against OpenAI's 32.3%. OpenAI lost the lead and is buying the window back.

The Decision Framework

The thing being done here is not a pricing change. It is the end of a period when AI costs are structurally unpredictable and providers have not shipped the instrumentation customers need to govern them. A team without per-customer, per-feature inference cost telemetry is one successful adoption cycle away from ServiceNow's outcome.

The Lemkin Account Tells the Whole Story

Jason Lemkin opened his Salesforce bill, cut the seat count from 10+ humans to 2 humans plus 1 API seat, and ended up spending 83% more money. The new bill is $22K. The old bill was $12K. That is what a GTM leader actually does when they rewire around agents. Fewer humans logging in. More budget on the line item. Salesforce kept the revenue. Salesforce also lost nine of its daily active users. Both facts live in the same account, on the same renewal.

If the pricing page is still per-seat, it is optimizing for a shrinking metric.

a16z Calls It: System of Record → System of Intelligence

The a16z GTM thesis (May 2026) argues that most of the next decade's enterprise value accrues to the intelligence layer rather than the data layer. The frame they use: switching costs are migrating from 'our data lives in Salesforce' to 'our workflows, reasoning, and institutional context live in our AI layer.' Software historically captured only 5-10% of GTM spend. The other 90-95% was payroll. Agents do not replace payroll. They make each payroll dollar more productive, which means the addressable market is a slice of payroll software could not previously bill against.

Where Sources Disagree

a16z says CRM usage has actually risen since AI tooling spread, because the system of record is being consumed as API infrastructure. Lemkin's account shows the opposite signal at the seat layer: 9 of 10 daily active users disappeared. Both are true. Separate the thing being pitched from the thing being done. The data gets consumed more, through APIs the agents call. The interface gets consumed less, because humans stopped opening the tab. Products whose value lives in the interface are exposed. Products whose value lives in the data are defensible only if they own the orchestration above it.

The Timing Window

Startups are deploying agentic workflows faster than Salesforce or ServiceNow can ship them, but both incumbents are building API-first AI offerings on the same clock. The window for orchestration-layer lock-in is open and narrowing. The forcing function this quarter: pick one high-frequency workflow, instrument retention and usage depth rather than engagement, and ship depth there before the incumbent's API surface closes the gap. One workflow owned end-to-end beats three workflows half-integrated. That is the decision on the desk Monday.

The Convergence Is Not a Coincidence

SAP, ServiceNow, and Salesforce all shipped autonomous agent architectures this week, and they all landed on the same execution layer: headless workflows callable over MCP. SAP added a €100M partner fund and a Knowledge Graph that feeds business context to agents. ServiceNow's Action Fabric decoupled workflow logic from UI and exposed it so any third-party agent can call it. Companies do not stand up hundred-million-euro funds for features. They stand them up for platform bets they plan to defend for years.

The Procurement Conversation Already Changed

A Fortune 500 procurement manager asked three vendors the same question in demos this week: 'Can our agents call this directly, or do my people have to click through your UI?' Two vendors did not have an answer. The third did. The third advanced. Meanwhile, Vercel's production data shows 59% of AI token volume is now agentic workloads. Agents are the primary consumption pattern, not the experimental one.

The enterprise buying committee has moved from 'show me the dashboard' to 'can our agents orchestrate your workflows' — and the window before this shows up in RFPs is two to three quarters.

The Build Scope Is Smaller Than the Deck Suggests

For most teams, shipping an MCP server against the existing API is a week of scoping, two to four weeks of build, assuming the underlying API is not already a mess. The harder decision is whether the product's core UI should be restructured around the assumption that an agent, not a human, is the primary first-touch user for a non-trivial share of sessions. That is a roadmap question, not a sprint question.

The Diagnostic

Pull the last twenty support tickets and feature requests from the top decile of accounts. Count how many assume a human in the seat and how many assume an agent or integration is doing the work. If that ratio has moved 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 on it is the next renewal cycle.

One Person Shipped What Took a Team a Week

Elena Verna — former head of growth at Amplitude, Miro, and Dropbox — now spends 90% of her time building at Lovable with near-zero meetings. Last week she pushed Lovable's enterprise pricing page to production by herself. The same scope in a traditional org would have meant a PM, a designer, engineers, and about a week of calendar time. She is not a curious executive opening Figma for the first time. She is a senior operator using AI tools to hold the whole stack in her head at once.

Lovable Has No Product Managers

The company is growing fast enough that the absence is not an oversight. It is the design. Engineers talk to users, write the specs, ship the code, read the feedback. The bet underneath is specific: when the engineer is the one in the user conversation, three of four PM jobs (user research, prioritization, spec-writing) collapse into the engineer. The fourth, cross-team coordination, disappears when the org is small enough that coordination happens in a shared channel.

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

The Threat Model, Not the Career Story

Ravi Mehta's framing is the useful one: AI makes one person 'average-to-good at everything at once.' For a PM who already thinks across functions, that is an opening, but only if the reclaimed time goes into shipping rather than coordinating. The PMs who survive this shift look less like project managers and more like mini-GMs who prototype and iterate directly. Companies that ungate information access will collect disproportionate talent density. Companies that protect management layers will end up with coordinators and no builders.

Your Diagnostic This Week

Two questions, in order. First: of what 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 still justify the role? If yes, the job gets better. If no, the job gets done by someone like Verna.

Separately, Duolingo's CEO publicly acknowledged that the blanket 'evaluate all employees on AI usage' mandate produced performative adoption with ~20% unusable output. They reversed the policy. Mandate AI outcomes, not AI activity.