The Product Desk

What the tape is de-rating, precisely

A procurement lead sorted her renewal calendar this week by contract type instead of by vendor, and the list rearranged itself into winners and casualties. That sort is the read. The market is not de-rating software. It is de-rating seat-based licenses and legacy infrastructure while paying up for consumption and AI-native models. The Information's briefing lines the evidence up cleanly enough to paste into a pricing review.

Workday is the case worth studying because the market cannot price it. Shares spiked 19% Thursday on the Reuters report that Silver Lake is in talks to acquire it, gave back nearly 4% Friday, and now sit 13% above the pre-report level. Up 55% since late June, still down 7% year to date despite $2.8B in free cash flow. Reuters carried no purchase price, so the buyout multiple everyone wants as an anchor stays unset. What teams tell themselves is that cash generation protects the price. What the tape says is that it protects the company and not the packaging, and nobody has repackaged yet.

The dated catalyst that costs nothing

Salesforce reports in roughly two weeks, with analysts polled by Refinitiv expecting revenue up more than 10% to $11 billion. Treat that call as primary research on the question being litigated in your own building: does AI arrive as incremental revenue for an incumbent, or does it only defend the base? Pull three disclosures off the transcript. Attach rate. Pricing structure. Whether AI revenue reads as incremental or substitutive. Those three go into the AI SKU business case before someone else challenges the projections with worse data.

Discount the number honestly

Two caveats belong on the same slide as the 30%. It is stated intent, not realized spend, and enterprises say this in every downturn. And KeyBanc's Jackson Ader attributes part of the recent software bounce to technical rotation, money leaving high-momentum names for the relative losers of late, which means the recovery may have no fundamental floor under it. What is different this cycle is that mainframe decline puts substitution into reported financials rather than survey responses. Plan for the intent being half true. Half of that mandate is still a repricing event across the whole renewal book.

One mechanism, three markets

This is not only a software story, which is why it should change sequencing rather than talking points. Diamandis's abundance forecast and the plant-based milk collapse describe the same mechanism from different angles. The capability a company charges for gets absorbed by a free model, by an incumbent's feature list, or by the buyer's own engineers, and the price follows within a cycle or two. The diagnostic fits on two axes: is the thing you charge for a capability or an outcome, and can the buyer obtain it without you next year. What survives in the first column is the defensible layer. Proprietary data. Integration depth. An audit and compliance surface. A measured outcome someone will sign for. Feature parity is not a defense against a buyer with a cut mandate.

A seat-based line item is the easiest thing a CFO under a cut mandate can terminate; a usage-linked contract shrinks instead of dying.

The layer that is actually copyable

An engineer at Zalando opens a chat window, picks a model, and never has to learn which vendor answered. That is the whole trick, and it is the unglamorous part. Zalando routes OpenAI, AWS Bedrock and Google Vertex models through an in-house gateway called zLLM, built on the open-source LiteLLM project, serving roughly 2,000 monthly active users and growing. One gateway produces one denominator. Active users, teams onboarded, and spend all land in the same table. Without that layer, an ROI slide is a narrative. There is a second-order effect worth naming alongside it: gateway monthly actives is simultaneously the adoption proxy and the inference-spend curve, so it grows with success rather than with budget.

The enablement machinery is equally copyable and almost never funded: weekly guild sessions, monthly trainings at 120-150 participants, hackathons, and scoped agentic pilots as the on-ramp. Scope is a political variable, not only an operational one. Once an org-wide number covering more than 250 engineering teams is public, a 5% internal pilot stops reading as progress to an exec team.

The counter-metrics are what make the story credible

Zalando published the parts that do not flatter it. That is the part to copy hardest. Cyclomatic complexity, a standard measure of how many tangled branching paths a piece of code has, is spiking as agent adoption grows. Commit messages are ballooning to 5,000 characters, which is a comprehension tax on reviewers and a leading indicator of review fatigue. There is explicit friction around tool migration and transformation too. The velocity gain is real. It is partly financed by code comprehensibility.

The forcing function is a single rule: every throughput metric on an AI-tooling dashboard needs a health counter-metric beside it, instrumented before rollout. Complexity trend, commit length, review rework rate. Debt discovered after adoption is discovered structurally.

Where the auto-approval number concentrates risk

Auto-approving a third of pull requests as low risk hands a governance decision that used to belong to a human reviewer over to a model-based risk classifier. That is a reasonable trade, and the two conditions sit in the same paragraph as the trade: name the risk classes that never route to it (authentication, payments, data-deletion paths, schema migrations), and have someone review the classifier's false-negative rate on a fixed cadence. Nobody publishes that number yet. The team holding it internally before an incident asks for it is the team that gets to keep the trade.

Throughput is the easy metric; effect size is the hard one

Set these numbers against the measurement standard emerging elsewhere. Diamandis's forecast benchmarks AI tutoring on Bloom's 2-sigma effect size, which is an outcome delta rather than an activity count. Zalando measures activity and speed. That is the correct starting point and not the finish line. The pattern running through today's material holds: a claim with a denominator survives scrutiny, and a ratio without one gets discounted the moment a finance partner reads it. Next quarter's agentic-tooling budget review will be won or lost on which of those two is in hand.

Caveat to carry with the numbers: this is a single organization's account, relayed by a practitioner who says he was personally part of the transformation. Treat it as a credible reference point, not an industry baseline.

Zalando's copyable asset is not its gateway or its models. It is that it can produce a numerator and a denominator on demand.

Three markets, three legal product names

A shopper reaches for the oat carton and scans the front panel for the word she already uses at home. Regulators in three markets disagree about whether that word is allowed to be there, which makes product nomenclature a jurisdictional compliance surface rather than a brand decision. Same carton, three legal name constraints, and naming now sits in the launch-readiness gate next to data residency.

The sequencing is the lesson. Alpro named the product "Shhh…This Is Not M*lk", a self-aware negation, and a court still ruled it potentially misleading. The category was already losing before either ruling. Dairy-free retail volumes fell more than 5% for three consecutive years on Circana data while dairy milk held, because consumers went hunting for protein. The incumbent did not win the label case first. It took the health halo, and the label case arrived afterward to remove the challenger's vocabulary.

What this turns into on a roadmap

Two exercises, both doable this week. First, delete every competitor and incumbent-category noun from the one-liner and check whether it still communicates value. If it collapses, what exists is not a position but a comparison, and comparisons are revocable by a court, a rebrand, or a search-ranking change. Second, name the single differentiator the largest incumbent could ship as a feature in one release note, then fund something they structurally cannot replicate: proprietary data, workflow embedding, distribution, or an outcome someone is accountable for. Diamandis arrives at the same idea from the other end when he concedes the skills layer goes free while judgment and accountability stay scarce. That is a moat statement dressed up as parenting advice.

Premium decoupled from trade-down, but do not over-read it

The incumbent is riding a premiumization tailwind that contradicts the trade-down narrative. US butter consumption hit a record 6.8 lbs per person in 2024 on USDA data, next to $60-per-pound Vermont farm butter, $15 ice cream pints, and third-wave frozen yogurt at up to $30. One Paris retailer sold 19 tons of butter, up 300% from 2023, largely to American tourists. That supports a test of a genuinely expensive top tier alongside a low-friction entry point, with hard scrutiny of whether the middle tier earns its existence. It does not support repricing the whole ladder: pandemic baking habits, flight from ultra-processed fats, and grocery inflation make this a lipstick-index pattern, and lipstick indices unwind when discretionary spend actually contracts.

The asset nobody could rebuild

"Got Milk?" reached an estimated 80% of US consumers on any given day and was retired in 2014. After that the category leaned on episodic virality: the 2020 #GotMilkChallenge with Katie Ledecky and Tony Hawk, then a celebrity chug at a 2025 award show. High-recall assets are cheap to reactivate and effectively impossible to rebuild from scratch. That inventory deserves an audit before a legacy feature or product name gets sunset in a rationalization pass.

If deleting your competitor's name breaks your positioning, you do not own a category — you are renting one, and the landlord can evict you.