The Product Desk

The precedent chain is what makes this one different

The safety lead who pulled her app's under-18 numbers this week did it because of a docket, not a roadmap review. Design-defect claims against consumer software are now three for three. A Los Angeles jury found Meta and Google liable in March for one young woman's depression and anxiety, awarding $6M on a defective-product-design theory. A New Mexico judge then ordered $942M in total penalties, $567M on top of an earlier $375M, called Meta a public nuisance comparable to air pollution, and mandated in-state safety features. Thousands more suits are pending. Meta's CFO Susan Li told investors in July that this year's trials "may ultimately result in a material loss." CFOs do not use that phrase about nuisance litigation.

The damages figure is the part that gets briefed upward and the part that matters least. The presiding judge called the states' $1.4 trillion demand "unreasonable" at a pretrial hearing, then called Meta's own $4M estimate "also unreasonable" in the same breath. Reuters puts the states' internal figure closer to $200B. What actually ships out of this trial is injunctive relief against named mechanics, and unlike New Mexico's in-state order, this one would potentially apply nationwide. That collapses the state-by-state compliance arbitrage most consumer teams have quietly relied on.

Two of the remedies are builds, not deletions

Separate what the remedy list is called from what it requires. Plaintiffs want Meta forced to detect minors running multiple accounts and to build parental verification for underage users. Those are staffed engineering programs with schedules, not config flips, and if granted they become de facto national requirements. Meta's courtroom posture makes the exposure map worse for everyone else: its lawyer argued Facebook is overwhelmingly used by adults while conceding Instagram skews younger, and the narrative pushes Snapchat forward as the app better known for appealing to kids. That is risk transfer. Whoever skews youngest without defensible age assurance inherits the target, and no one can make the counterargument without knowing under-18 and unknown-age share by surface.

MIT Technology Review's coverage of an unrelated case supplies the checklist that generalizes. Flock Safety's roughly 120,000 license plate readers, where the harm traced back to four choices: what to collect, who can search it, how long it is retained, how widely it is shared. Misuse controls arrived retroactively. A fifth row belongs on that list, which is abuse cases reviewed before GA.

Section 230 covers what your users publish. It does not cover how you engineered the loop that keeps them publishing.

Five independent reports read the same trial and reach the identical conclusion, which is unusual: the money is theater and the injunction is the product event. The live disagreement is scope. Appeals could narrow a nationwide order, and Meta's demographic defense may genuinely undercut intent. The sort worth running in the next planning cycle has two axes. One is cost to build proactively. The other is cost under subpoena. Everything sitting in the cheap-now, expensive-later cell goes into the sprint, and the scope uncertainty does not move a single item out of it.

The mechanism behind the cheaper number

An admin opened the model-selection panel, read the list, and left it on automatic. Glean says customers do that overwhelmingly, "mostly because of cost." That choice is the tell for what is actually being sold. The claimed advantage is routing sequence, not model choice. Waldo, the agentic search model launched in April, decides how to break down a question, which tools to use, what to read next, and when it has enough evidence to hand off to a frontier model, so model selection happens after the context is assembled. Arvind Jain's framing: "We're able to assemble the raw materials needed to do the work without burning LLM tokens." Claimed payoff: 50% lower latency and 25% fewer tokens. Two adjacent patterns are becoming table stakes. One is three-tier model control, where end users pick a model, admins restrict models and set usage limits, and an automatic mode routes per task. The other is that refusing to call an LLM at all now ships as a feature; Jain's example is users asking the product to multiply two numbers.

Separate the thing being pitched from the thing being measured. The $0.45-versus-$1.84 comparison, the 4x claim against Claude Code, and Waldo's 50% and 25% figures are vendor-reported without disclosed task mix or methodology. They are good evidence that the competitive axis moved. They are not numbers that survive a slide review. The valuation carries its own caution: $300M ARR at $7.2B is roughly 24x, which holds only if the budget compression Jain describes doesn't bite.

The sources disagree on the direction of token prices

The Information reads Alibaba as commoditizing the low end: a free 27B model that ties OpenAI's cheapest flagship variant and satisfies air-gapped deployment. Latent Space reports the same inflection from the buyer side, with open-weight usage moving from "minuscule" to essential inside roughly three months, at about an order of magnitude cheaper per task. Both arrows point down.

Paul Smalera's reporting points up, and it is the less comfortable read. Nvidia licensed Groq's technology and hired founder Jonathan Ross with much of his team, then invested in what was left at $3.5B, roughly half its $6.9B September 2025 mark. Groq now resells inference capacity instead of designing chips. The credible price competitor in inference silicon was neutralized without an acquisition. The planning assumption is flat-to-rising token costs on the frontier tier while the open-weight floor keeps dropping, because that is what the two stories describe together.

TheSequence sharpens the latency half: a feature needing sixteen samples plus a majority vote pays roughly 16x per query forever, and its p95 disqualifies in-editor, voice and real-time surfaces entirely. Compounding Quality supplies the investor frame boards already use, which treats personalization as a moat only while unit cost stays flat as engagement rises. Engagement is a weak stand-in for value, but it is the axis the chart gets drawn on anyway, so chart cost-per-personalized-session by engagement decile before someone external does. The forcing function for this sprint is narrow. Every shipped feature names its cost-per-task and its p95. Anything that cannot produce both numbers is a prototype.

Enterprise AI stopped competing on what the model can do and started competing on what the task costs. A team that cannot state its cost-per-task cannot defend its roadmap.

Churn one: a power user, three named defects

Casey Newton had thousands of saved links in Notion and gave up on its agentic search. He replaced it with a prompt written by Claude Fable 5, pasted into a terminal, writing Markdown into Obsidian, seeded with his entire archive since 2020. One month later the stack holds 1,440+ pages, including a single Meta page running 12,000+ words with 1,300+ internal links, auto-generating concept pages and refreshing a home page every morning. His job title contains no engineering.

He never filed a model complaint. The three reasons he left read like a PRD checklist: in-database search was keyword-only, the agent lived on a separate app surface from the data, and it failed to cite sources without extra prompting. The second failure in the same account travels further. His manual concept-tracking system died of proliferation. Nothing was removed; the marginal cost of maintenance outran the marginal value. Products that ask the user to be the librarian have already priced that decay into the retention curve. He calls the replacement the most useful thing he tried this year and too clunky to recommend, which is the textbook precondition for a commercial winner. Score the managed-LLM-wiki category (Town, CEO Jean-Denis Greze) before funding an internal build.

Churn two: nobody's dashboard moved

The SaaStr cancellation came with no competitive loss event, and the analysis says plainly that traditional usage and customer-health metrics stop predicting anything under agent absorption. The detectable fingerprint is a ratio, not a level: programmatic, API and MCP read volume rising per account while human seat sessions fall. Retro-test it against the last four quarters of churn. If it would have flagged half of them a quarter earlier, it has paid for itself.

The same reporting supplies the vocabulary buyers bring to renewals. Google shipped least-privilege agent identities, flow-level access controls, audit trails and human approvals into the Workspace admin console. Docker began logging AI governance decisions and streaming them to SIEM. Teleport shipped short-lived database certificates for agents touching production data. Dynatrace paid ~$915M for Arize to join model output and agent behavior to app traces. Four vendors converging on one control list means the list is a procurement checklist now, not architecture advice.

The buyer may not be a person

Vercel shipped an unpromoted CLI capability and saw an immediate usage uptick. Agents reach new products faster than its own employees absorb them. It holds data on agents creating accounts and buying third-party products through a Marketplace that Guillermo Rauch describes as "almost like an Amazon.com for the cloud." Alipay launched a full agentic commerce foundation plus the AHA cross-device protocol suite, turning authorization, fulfillment and settlement into a protocol across thousands of already-adapted services.

Read the magnitude carefully. Vercel's usage claims are unquantified, "an uptick," "data showing," and a16z discloses that third-party information was not independently verified. Direction is high-confidence, size is unproven. What the sources agree on is narrower and worse. The failure is instrumentation, not capability. None of these accounts left because a model underperformed.

An agent evaluated, signed up for, and bought software. Analytics that cannot tell it apart from a human are optimizing the funnel for the wrong user.