The Product Desk

The Arbitrage Is Over

A staff engineer opened Cursor on a Tuesday and shipped three PRs before lunch. The cost to her employer, routed through a Claude subscription via a third-party harness, was a fraction of what the same tokens would have cost on the API. That arrangement is what teams have been buying for the last year. It is not what they were pitched. Anthropic announced that every Claude subscription now includes API credits equal to the plan's dollar amount, which sounds generous. For the cohort using Claude through third-party harnesses (Cursor, Cline, Conductor, Zed, OpenCode) at effective 70-90% discounts to API pricing, it is a price increase of roughly an order of magnitude. The change takes effect June 15. Overage bills at full API rates.

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 an order of magnitude.

The IPO Clock Explains Everything

What teams told themselves: power users were on a stable plan. What was actually happening: power users were the implicit subsidy that does not survive contact with an S-1. Anthropic hired a CFO and is likely targeting an October 2026 IPO. Expect at least one more pricing adjustment before then. ServiceNow's CDIO Kellie Romack watched her team's full-year Anthropic budget get consumed before mid-2026, and cannot tell which users drove it, because Anthropic doesn't ship enterprise-grade telemetry.

OpenAI's Counter-Move Has a Shot Clock

Sam Altman offered 2 months of free Codex to enterprise customers who switch within 30 days. This is displacement pricing, timed to the developer frustration window. Ramp data shows Anthropic at 34.4% versus OpenAI's 32.3% in business adoption. OpenAI lost the lead for the first time and is buying it back on a clock.

The Decision Matrix

The Structural Cost Problem Beneath

Separate the thing being pitched (model competition) from the thing being done (capacity hoarding). OpenAI committed $20B to Cerebras, a decade-scale lockup. Nebius reports 4+ customers competing for every GPU brought online. Inference costs are not bending down. They are being held up by buyers willing to pre-commit at enormous scale. Any roadmap that assumes cost deflation should stress-test against costs staying flat for 8 quarters.

The Convergence Signal

SAP, ServiceNow, and Salesforce all shipped autonomous agent architectures this week. They converged on the same execution layer: headless workflows callable over MCP. SAP put up a €100M partner fund and a Knowledge Graph for agent context. ServiceNow's Action Fabric decoupled workflow logic from the UI so any third-party agent can call it. Three vendors picking the same architecture in the same week is a commitment they plan to defend.

A procurement manager at a Fortune 500 opened three enterprise software demos this week and asked the same question in each: '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.

The Production Data Confirms the Shift

Vercel's AI Gateway production index is the closest thing to real usage data at scale, and it shows 59% of token volume now flows through agentic workloads. Anthropic captures 61% of spend (Opus for reasoning), Google captures 38% of volume (Flash for fast and cheap tasks), and most large teams route across multiple providers. Separately, Notion launched a full developer platform with agent tool building and plans to host Claude and Codex as 'teammates.' Workspace-as-agent-host is consolidating fast.

What This Means For Your Product

The enterprise buying committee moved from "show me the dashboard" to "can our agents orchestrate your workflows." The window before that shows up in RFPs is two to three quarters. If a product's core workflows cannot be invoked by an external agent without a human clicking through a UI by Q4, agents living in the buyer's stack will route around it and talk to the system of record directly.

The Scope Decision

Separate two decisions. First: ship an MCP server against the existing API this quarter. For most teams that is two to four weeks of build, assuming the API is not a mess. Second: restructure the core UI around the assumption that an agent is the primary first-touch user for a non-trivial share of sessions. That is a roadmap question, not a sprint question. Picking one of those is better than hedging both.

Apple Complicates the Mobile Layer

Apple is building AI agent governance into the App Store, with a possible WWDC June reveal. They are solving three problems at once: an approval process for agents, the 'agent spawns unauthorized sub-apps' problem, and making sure agents cannot route around App Store fees. If an agent roadmap involves open-ended tool use or dynamic interface creation on iOS, pre-declare the capability boundaries before Apple defines them.

The Lemkin Account: A Preview of Every Renewal

Jason Lemkin looked at his Salesforce bill and cut the seat count from 10+ humans to 2 humans plus 1 API seat. His spending went from $12K to $22K — up 83%. The vendor kept the revenue. The vendor also lost nine daily active users. Both facts live in the same account. a16z argues that pattern defines the next decade of enterprise value, and the renewal conversation it implies is one most CRM account managers are not yet rehearsing.

Switching costs are migrating from 'our data lives in Salesforce' to 'our workflows, reasoning, and institutional context live in our AI layer' — creating a new moat category.

The Moat Migration

The traditional CRM moat was install base, integrations, and muscle memory. Reps trained to put notes there. That part is still intact. What changed is where the decision gets made. The intelligence layer reads the notes, weighs the pipeline, and tells the rep which accounts to call. Increasingly that layer is a separate product owned by a different vendor. Teams use the CRM as a filing cabinet and decide somewhere else.

Anthropic's Claude for Small Business shipped with connectors to QuickBooks, PayPal, HubSpot, Google Workspace, and Microsoft 365. That is the thing being done, not the thing being pitched. It is not a model announcement dressed as a product. It is a product that happens to use a model. One has a moat. The other is a line item on an inference bill.

The TAM Reframe

Software was historically 5-10% of GTM spend. The other 90-95% was payroll. AI agents do not replace payroll. They make each payroll dollar more productive. The addressable market is a slice of the payroll line that software could never reach before. That is why Lemkin spent more, not less.

The Window Is Narrowing

Salesforce and HubSpot are building API-first AI offerings to capture the intelligence layer inside their own walls. The startups winning are clustering around narrow, high-frequency workflows with structured inputs and measurable outputs. The 2x2 to draw on a whiteboard this week: workflow breadth on one axis, input structure on the other. Breadth across many fuzzy workflows is the losing cell. Depth in one structured workflow is where retention compounds and institutional context accumulates. The forcing function for the next sprint: pick one workflow where inputs are structured and outputs are measurable, then ship deep enough that the customer's reasoning lives inside your product before the platform vendor catches up.

The Worked Example

A clinician finishes a patient visit. Instead of typing a note for 11 minutes, she reviews a draft that is already written. That is the entire user moment Abridge picked. It was boring enough that procurement said yes, which is the part most teams pitching healthcare AI skip past. The wedge produced 80-100M+ recorded medical conversations that exist in no EHR and no foundation model's training set. The valuation followed the wedge: $5.3B on $550M raised.

The Three-Act Ladder

1. Save time (documentation) — measured in minutes per encounter, earns clinician trust
2. Save money (prior authorization, billing) — measured in coding accuracy and revenue capture, earns admin trust
3. Save lives (clinical decision support) — measured in outcomes, earns clinical leadership trust

Each act unlocks a different buyer, a different contract size, and a different release cadence. Release cycles compressed from semi-annual to monthly, a 4-6x acceleration, because Act 1 earned enough trust that procurement stopped treating Abridge like a vendor and started treating it like infrastructure.

The Architecture Choices Worth Copying

Abridge runs a 'constellation of models' with thinking-fast-and-slow routing. A cheap model triages every interaction. A larger model handles the complex cases. At 80M+ conversations a year, the unit economics live on that routing table, not on the model card. Memory sits in a separate store rather than baked into weights. The team expects to swap foundation models on a regular schedule, and they built for that.

The alert design is the part most teams should steal. Over 90% of healthcare alerts are ignored because of alert fatigue. Abridge's response is ambient AI that operates 'like air conditioning': silent by default, audible only when something is contextually critical. The winning AI product is not the one surfacing the most insights. It is the one with the highest precision-to-interruption ratio.

The lesson is not the valuation. It is that the valuation followed a wedge a clinician could describe in one sentence.

The Portable Diagnostic

Two questions, in order. First: is there one user, one workflow, one moment, where the product is the default tool and not the experiment? Second: does the next thing on the roadmap extend that workflow, or does it ask the user to learn a new one? If the answer to the first is no, platform expansion is premature. If the answer to the second is 'new workflow,' the team has not built a platform. It has built two wedges and a shared logo.