The Product Desk

The take rate is the actual dispute

Per Alex Heath's Sources reporting, cited in The Information's briefing, Zuckerberg believes Muse monetizes by "taking a small cut of transactions." Hold that against the counterparty: Amazon's retail business is not known for high margins. Andy Jassy has said on earnings calls that Amazon is "having conversations" with companies wanting to run commerce agents, and Martin Peers reports those talks continue. His verdict is the part to carry into planning: "Don't hold your breath for a quick agreement," because how money changes hands still has to be negotiated.

That explains the escalation pattern better than any technical read does. Amazon is not reacting to crawler traffic; it is pricing monetization intent.

The demand side is thinner than the share price

Meta jumped 11% Monday to $741, its highest close in a year, on Muse enthusiasm, and a WSJ-cited analyst projected Muse could add $28.5 billion by 2030 against Meta's $200 billion prior-year revenue base — roughly 14%. Treat that as a third-party analyst projection, not company guidance. Amazon rose nearly 2% the same session, so investors are not reading this as zero-sum.

The hands-on evidence points the other way, and it is more useful to you than the tape. Peers used Muse on Monday to buy a pen and a notebook from small retailers: the flow was "surprisingly smooth," and it was still faster to go directly to Amazon. A separate hands-on review reported by Lenny's found Muse returned wrong results on a specific New Balance 9060 colorway and opened the wrong movie before buying tickets, with browser use called "the weak link for consumer agents" rather than a Meta defect.

Where the sources diverge

Morning Brew reads the assistant layer as a proven-demand platform land grab, citing Muse at No. 1 free on the US iPhone chart and Instinct reportedly raising at a $10 billion valuation. The Information's reporting says consumers "aren't battering down the door." Both can be true, and the reconciliation is the product lesson: distribution is proven, transaction completion is not. Agents won on fragmented long-tail inventory where no aggregator exists, and lost precisely where Amazon already compressed search cost to two clicks.

Agents win where aggregation doesn't exist — and the category's best-reviewed agent still can't complete a purchase.

Two clocks matter. Perplexity v. Amazon is the only proceeding that will tell you whether agent access gets negotiated or litigated. And OpenAI is still only "mulling" a Muse response while it builds features to counter Grok Bot, with Apple and Google shipping nothing in consumer agentic commerce — which means merchant-side integration terms are cheap to set today and will be dictated to you once a device-level default agent exists.

Google made the admin plane the product

The capability half of Workspace Studio — custom triggers, Apps Script-powered steps, webhooks, native connectors — is the easy half. The half that resets your competitive position is that Google shipped the governance layer alongside it: every one of those capabilities is off by default, gated behind admin controls for integrations, webhooks, approvals and URL allowlists, per TLDR IT. Enterprise buyers normally have to demand, negotiate or build that. Here it arrives at zero marginal price inside a license they already hold.

Score the damage by buyer, not by feature. Zapier and Make lose long-tail Google-centric workflows and keep breadth of non-Google sources. Workato and Tray sell strong governance as a premium tier, which is exactly the budget justification that gets harder for "simple" internal automations. Microsoft's Power Automate is already bundled, so split-suite accounts become the contested ground. What survives is run-level observability, error handling, cost attribution, and policy granularity that an on/off toggle cannot express.

OpenAI took the identity layer without asking you

Separately, OpenAI enabled multi-account support across most ChatGPT plugins. A user links personal, work and side-project accounts for the same service, and ChatGPT — not your integration — picks the right account per request and respects that account's permissions. The critical detail from Simplifying AI's reporting: it works on existing plugins without any developer changes. Your product's behavior changed inside one of the largest distribution surfaces in software, you did not test it, and you cannot opt out.

Read the risk before the opportunity. Automatic account selection is a wrong-tenant-write vector you inherited without shipping a release. The one lever you control is a profile tool on your MCP server so each linked account renders a human-readable label — "Acme workspace" instead of a generic entry. Hours of engineering, in front of a surface millions of people see, on a change most of the plugin directory will receive passively and never optimize for.

Perception is now priced per thousand

Meta put SAM 3.1 on its Model API: one call from a short text phrase returns bounding boxes, pixel-precise masks and identity-preserving video tracks, with up to 16 targets per pass and nothing to host. Pricing is $2.50 per 1,000 images and $0.20 per 1,000 video frames — roughly $0.36 per processed video minute at 30fps. It accepts calls through the OpenAI SDK, which makes switching into Meta a config change rather than a project. Note the asymmetry the open weights carry: they support visual prompts only, with no text-prompted segmentation, so the hosted path is the capable path.

Any epic that still reads "build detection, segmentation or background removal" is now an integration ticket with a price list attached.

The pattern repeats a fourth time in the same window: Microsoft added Drive Sanitization and remote deployment via Recovery CSP to Cloud Rebuild in a Windows 11 Insider build, pulling endpoint sanitization and imaging into the OS. Across three independent sources the mechanism is identical — the hard part moves into the platform, the price goes to zero or near it, and what remains yours is workflow depth, policy expressiveness and licensing discipline.

What the disclosures actually license a team to claim

A product lead spent Friday afternoon reading earnings transcripts, hunting for one number she could borrow for her own board deck. The Exponential View research desk did that work at scale and found that 33% of companies holding a June 2026 call made a quantified AI statement, with 35% of quarter-to-date calls doing the same. That is roughly 10 percentage points above the same point last year. Composition matters more than the count. Cost and productivity claims came from 24% of the cohort, averaging 47% improvement inside a relatively tight 20–60% band. Revenue and demand claims came from 18%, averaging 40% across a 20–83% range. The authors note that the two averages sit inside each other's interquartile ranges, so the 47-versus-40 gap is noise.

The usable instruction is about sequencing. Efficiency effects land before top-line effects, so lead with the cost claim and hold the revenue claim until instrumentation backs it.

The contradiction worth naming in the planning meeting

The same dataset says two opposite things. Deployment language rose from 7% to 12% of calls while pilot language never cracked 4%, so enterprises are shipping and declining to call it experimentation in public. Only 15% of the index produced a quantified claim about AI's business impact, up from 9%. They can deploy; they just aren't measuring, and that measurement problem is what the 15% is showing. Every figure here is self-reported, unaudited and selectively disclosed. Treat them as a floor for internal targets. Don't put them in a customer contract.

The same audit fails on internal velocity claims

Warp's own factory data, reported by Lenny's, shows where the internal version of this gap sits. Its agent system carries a request from Slack through triage, ticket creation, implementation, PR and merge at roughly 2,000 PRs per month, averaging 35 minutes from kickoff to PR. The first human review lands 3.5 hours later. The CEO names human code review as the biggest bottleneck in his own engineering process. The metric Warp says most teams aren't tracking is human interactions per pull request: every Slack follow-up, ticket clarification and review correction.

Kickoff to PR is 35 minutes; the first human review is 3.5 hours after that.

Refactoring is the dependency underneath both problems. Machine-readable quality gates serve juniors, seniors and agents identically, and without that harness the AI-driven velocity in a quarterly plan is booked but unfunded. The shortfall shows up in a feature's defect rate, weeks after the plan was signed. The forcing function for this sprint: take the two line items in the plan that assume agent throughput, write down the human interactions per pull request each one requires, and if that figure isn't instrumented yet, the claim belongs in the internal target column and nowhere else.