ClarityBeta

Leadership & Executive

The Board Room

9 min read · 1,871 words · 3 moves to watch
The Signal

Four states asked a federal judge to strip infinite scroll from every consumer feed.

Meta has failed repeatedly to get the design-defect framing dismissed under Section 230, which matters more than the damages arithmetic. Both numbers on the table have already been called unreasonable: $1.4T from the plaintiffs, $4M from Meta. So the live exposure is no longer a check. It is a nationwide mandate covering autoplay, ephemeral content and filters, which makes the engagement features already shipping in your product a compliance question rather than a design one.

In Play

  1. Engagement Design Becomes Product Liability

    California, Colorado, Kentucky and New Jersey opened trial against Meta in Oakland federal court, asking the judge to order removal of infinite scroll, autoplaying video, Stories-style ephemeral content and beauty filters, per Morning Brew's reporting. For any consumer surface with retention mechanics, this is now a product-defect question rather than a speech question. The judge has called both the states' $1.4T demand and Meta's own $4M estimate unreasonable, which moves the real risk to the remedy.

    Ask Clarity
    Try

  2. Switching Costs Reprice as Migration Labor Automates

    Rillet raised $100M at a $1B valuation, double its November mark, and Campfire is fielding offers at $1B against $375M last October, per The Information's dealmaking coverage. Both sell almost exclusively into companies with no ledger history to move — the absence of switching cost is the whole position. Google Cloud separately called the embedded-engineer model structurally uneconomic and shipped software to automate the data-context work, which prices the labor that makes leaving expensive.

    Ask Clarity
    Try

  3. Agent Absorption Is a Churn Vector With No Signal

    SaaStr canceled Notion after seven years because an internal agent gradually absorbed the last workflow the tool served, with no drop in usage metrics beforehand. Dynatrace agreed to pay roughly $915M for Arize, the first public comparable for AI and agent telemetry. If your revenue is seat-based and serves a narrow workflow, the churn signal now arrives at the renewal call rather than in the health dashboard.

    Ask Clarity
    Try

  4. AI Capex Is Now Setting Your Cost of Capital

    A Wall Street Journal footnote review found roughly $3T of off-balance-sheet AI commitments across nine large technology companies — about $1.2T of unstarted leases plus $1.9T of purchase commitments — against roughly $600B of reported capital spending. Bloomberg separately reported that AI capital demand is lifting US Treasury yields, with the 30-year at its highest level since June 2007. Any 2027 case approved at last year's rates is worth less, and rival capacity read from cash flow statements runs five times light.

    Ask Clarity
    Try

  5. Board Interlocks Became a Legal Surface

    The Justice Department has spent nearly a year probing Andreessen Horowitz under the Clayton Act's ban on interlocking directorates, citing partners holding seats at competing companies including Databricks, dbt Labs and Fivetran. Your investors' board seats are now your regulatory overhang, and the probe began alongside a merger review — meaning deal diligence is the discovery gateway. Sequoia's 2020 exit from the Finix board already gives regulators and acquirers a benchmark for what remediation looks like.

    Ask Clarity
    Try

Deep Dives

The Injunction, Not the Check

Damages in Oakland will be negotiated down; the nationwide product mandate the states actually want is the exposure most consumer surfaces have never priced per feature.

Why the shield stopped holding

Section 230 answers a publishing question: who is liable for what a user posts. The states in Oakland decline to ask it. Their claim is that Meta engineered a product to addict minors, which routes the argument into product law, and Meta has repeatedly failed to get design-defect claims dismissed on 230 grounds. One appeals court already declined to reverse. Appellate rescue is now a binary bet rather than a planning assumption.

The migration inside the defense is the more useful signal. The lead argument in the opening was not immunity but demographics: Facebook is overwhelmingly used by adults, Instagram skews younger but is "no Snapchat," per The Information's account of the proceedings. That is a factual defense, and companies fall back on facts when the structural argument stops carrying weight. Note the direction it points. Every defendant in this category will now gesture at whoever skews younger, which makes products with young user bases the designated next target, and Meta's CFO has already used the word "material" with investors.

Damages get negotiated; remedies get imposed

Our sources disagree productively about the money. One reads the headline demand as theater against a roughly $200B actual claim. Another treats New Mexico's $942M — $567M on top of $375M — as the honest calibration. A third argues thousands of parallel cases compound into a portfolio-reserving problem rather than a single check. All three converge on the part that survives settlement: plaintiffs are also seeking operational changes to the products themselves, plus mandatory parental verification and enforcement against minors holding multiple accounts.

The precedent sits in the structure, not the sum. New Mexico's judgment was absorbable partly because its safety requirements were state-specific, which let the arbitrage stand. A California federal court ordering a nationwide design mandate ends the arbitrage. In March, a Los Angeles jury extended the same liability theory to Google, so the exposure is not one company's peculiar problem.

A fine is a charge against earnings. A court-ordered redesign of retention mechanics is a revenue-model event.

The capability market forming around the remedy

Age assurance, parental verification and multi-account detection are moving from compliance theater to table stakes, and the capability prices cheapest while the category leader is still on trial. Shipping it voluntarily also shapes a consent decree rather than receiving one. The harder internal problem is attribution: most firms cannot tie revenue or engagement to a specific per retention feature mechanic — autoplay versus streaks versus ephemerality — which means they would negotiate a redesign blind.

The regulator changed address

Four states are driving this trial. Twelve, California-led, have stalled Paramount Skydance's $110B takeover of Warner Bros. Discovery, and Paramount has answered by demanding the states post a $1.88B bond covering delay fees owed to Warner, payable to Paramount if it wins. A skeptic would call that a litigation stunt, and it may fail. If it survives, every acquirer copies it inside a year and state intervention gets repriced. Either outcome leaves a deal model gated on federal clearance alone under-scoped.

One reflexive detail worth holding at board level: Meta's AI buildout is funded by advertising revenue tied to the exact mechanics under injunction. Anywhere a single revenue loop funds the strategic roadmap, litigation aimed at the loop is litigation aimed at the roadmap.

What to do

  1. Commission a design-liability audit this month that maps every retention mechanic — autoplay, infinite scroll, streaks, ephemerality, filters — to per-feature revenue contribution and measured under-18 usage.

  2. Price an age-assurance and parental-verification vendor shortlist this quarter, with a shippable milestone inside two quarters.

  3. Add an explicit state-AG delay-cost line item to every transaction above $500M before the next deal reaches your committee.

Your Moat Was Migration Labor

Two accounting startups and one hyperscaler's cost-structure attack point at the same conclusion: the price of leaving your product is falling, and almost nobody has priced their own exit.

The tell is who they sell to

Rillet and Campfire are not prying accounts out of NetSuite. They are landing where there is nothing to migrate: three-year-old, fast-growing companies with no ledger history. Campfire has not spent the $65M it raised in its prior round, and a16z sits in Rillet's cap table. Capital is being pushed onto these companies rather than pulled by need. The marks are not evidence of product-market fit against an incumbent. They are evidence that sophisticated capital now believes switching costs in enterprise software are becoming payable. That belief is a claim about the incumbent's customers, not about the challengers.

The labor that priced the moat

Google Cloud spent the week arguing the same thing from the supply side. Andi Gutmans, the VP running its database products, called the embedded-engineer model structurally uneconomic: "If you want to move to activating 100% of your enterprise data, you're not going to be able to hire enough people to make that happen." That is aimed at the billions OpenAI, Anthropic, Microsoft and Amazon are putting into forward-deployed engineers, following the playbook Palantir invented. Knowledge Catalog carries the argument: agents crawl a customer's data, infer how it maps to sales, inventory and operations, and emit knowledge graphs and semantic layers. Virgin Media O2 connecting 20,000 data sets is the flagship proof point. That figure comes from a Google spokesperson, not from the customer.

Palantir shipping its own automated FDE product earlier this year is the more reliable signal. When the company that invented a category races to automate its own differentiator, the category's pricing power has peaked. Every rival's services headcount becomes a stranded cost the moment automation looks credible.

Where the sources diverge

A reasonable skeptic would point out that Google concedes its context agents are not foolproof and that customers must still staff humans to vet output. The skeptic is correct. The promised "fully agentic" end state carries no timeline, no accuracy benchmark and no independent validation, which makes the automation directional rather than shipped. What the skeptic misses is that the failure mode is nastier than slow deployment. A human who misreads a revenue definition produces a visibly wrong dashboard. An agent that encodes the same error into a semantic layer produces thousands of confidently wrong decisions across invoice routing and onboarding, with no obvious point of detection. The verification tier is a permanent line item, not a transition cost.

Two numbers to run before the automation is real

First, whoever stores the knowledge graph and metric definitions owns the renewal conversation. Artifacts generated inside a vendor platform with no portable export are a switching cost accepted without ever being priced, in dollars and in months.

Second, the revenue-quality question sitting inside the growth story. Rillet's marquee customer Mercor booked $614M of gross revenue in the first half, up 70% on all of 2025, and roughly $205M net of contractor payments. An entire vendor layer is benchmarking its growth off cohort numbers carrying that gross-to-net gap. Both challengers are now loudly disclosing healthcare, nonprofit and distribution logos, which is pre-emptive defense against a concentration question they know is coming.

The tradeoff on the services line is a timing tradeoff, not a strategic one. Modeling implementation revenue under 30% price compression, and moving toward outcome-based contracts while the timing is still a choice rather than a response, costs less than repricing after a competitor's automation claim lands in a customer's procurement deck. This quarter's decision sets the price of next year's renewal.

Nobody attacks a moat built on inertia. They make leaving cheap, then charge for the move.

What to do

  1. Commission a semantic-layer portability audit this quarter that names who will own your knowledge graph, metric definitions and lineage, and prices the exit in dollars and months.

  2. Re-segment ARR by customer age and funding dependency before the next board meeting, underwritten on net rather than headline gross customer revenue.

  3. Move migration and onboarding into the product organization this quarter, with median time-to-cutover as its single owned metric.

The Churn Your Dashboard Cannot See

Agent absorption removes revenue without moving a single usage metric, and the first public price for agent telemetry has now been set — both decide where you sit in the stack.

The churn had no precursor

The cancellation is worth reading as a mechanism rather than an anecdote. Seats stayed provisioned. Logins continued. The job-to-be-done migrated to an internal agent one workflow at a time, and the whole apparatus of modern retention forecasting is structurally blind to that: usage telemetry, customer-health scoring, expansion signals. Renewal becomes the detection event, which makes it the most expensive possible place to learn.

A leading indicator exists and it costs an instrumentation ticket. Measure the ratio of programmatic and agent-mediated access to human session depth, per account. Rising programmatic volume against falling human engagement is what absorption looks like while it is still in progress.

What Dynatrace actually bought

The Arize price is the first hard public comparable for AI and agent telemetry, and it settles build-versus-buy for observability incumbents by answering the question with cash. The asset is a join, not a dashboard: model outputs and tool-use traces correlated with application and infrastructure telemetry. That comparable resets upward every quarter the remaining independents stay independent, so waiting is a priced decision rather than a neutral one. There is a second read worth ten minutes of corp dev's time. Arize's customers chose best-of-breed specifically to avoid full-suite absorption, and they are in play for the next several quarters.

The defensibility test

The useful frame is narrower than "AI is eating SaaS." Agents are eating interfaces, not systems of record. If an agent with direct access to a customer's data could reproduce 80% of the value, the product is an interface. If the agent needs the data, the write path or the enforcement authority, the product is infrastructure, and agents make it more valuable, because every agent action needs a verified system of record to write against. No ERP, ledger or identity provider is getting absorbed by an internal agent this year.

A skeptic would say that shipping an agent buys an exemption. Notion shipped one and still lost the power user, because database search remained keyword-only, agentic search lived on a different surface, and the agent would not cite sources without extra prompting. The product spec that falls out of that is one surface where search and agent both live, with citations by default.

Control is the new competitive axis

Three vendors converged on the same vocabulary inside a single week. Google added least-privilege agent identities to Workspace Studio. Docker began streaming agent policy decisions to SIEM. Teleport retired shared database credentials. None of them is competing on agent capability any more. They are competing on agent control, and that vocabulary is on its way into procurement questionnaires, where deals will be lost for its absence without anyone saying so out loud.

The distribution mirror image is already measurable. Vercel shipped a CLI capability with no promotion and saw an immediate usage uptick, because agents reach new surfaces before employees know they exist, and it reports agents creating accounts and making purchases on behalf of their humans. It also told the whole company to automate its own jobs, then consolidated hundreds of internal agents into one, because the scarce asset turned out to be shared context, identity and permissions rather than agents. Caveat: Vercel sells agent infrastructure and its usage claims are unaudited, so treat the direction as credible and the magnitudes as marketing. A channel nobody has tagged cannot be defended or funded.

A seat is an interface, and interfaces are what agents absorb first.

What to do

  1. Instrument programmatic-versus-human access ratio per account this month and review it monthly as the churn leading indicator.

  2. Force a funded build-or-buy decision on agent telemetry this quarter, with one path chosen and a milestone that holds.

  3. Tag agent-originated sessions, signups and purchases as a first-class channel within 45 days and publish a baseline.

The bottom line

Read together, today's items describe the expiry of protections nobody ever paid for: a legal immunity that arrived free with the internet, the cost your customers face to leave, the months a rival needed to copy you, and the assumption that a human signature counts as review. Each was a subsidy, and subsidies get withdrawn on someone else's schedule. What replaces them is only what you can evidence and enforce — contract terms, authoritative data, a documented decision trail. Inventory every protection your three-year plan treats as structural, name its owner, and price its replacement this week.