The week in three numbers: $563M Visa severance, in a beat-and-raise quarter · 2–4 people per Visa agentic squad, down from ten or more · No % — Visa never quantified AI's share of the cut.

Asked where the savings go, McInerney named a long list — acceptance, affluent propositions, cross-border, risk and security, marketing services, Pismo, Featurespace, B2B, Visa Direct — and then "stablecoins, agentic, our brand and our advertising... and obviously the Visa as a Service stack." Our Analysis: the headcount is not the story. A company beating and raising shrank the function that builds its product, and put part of the proceeds into rails that could route around it. (Filing · corrected transcript)

This week's question: when AI frees up budget, what is it actually being spent on?

Our standing scorecard: named companies' AI value claims vs. what shows in the P&L. None of the companies we track has reached Stage 4.

The evidence ladder: 0 · Narrative (management mentions AI) · 1 · Operational (quantified activity) · 2 · Financial claim (a dollar figure asserted) · 3 · P&L-attributed (AI isolated in reported results) · 4 · Sustained (Stage 3 holds four straight quarters).

Adjudicated this week. Issue 006 said S&P Global and Ecolab would earn Stage 3 on 28 July or come off the board. Both reported; neither earned it. S&P Global — its filed exhibit contains no AI-attributed figure of any kind. Ecolab — real AI-linked revenue ("29% growth in Global High-Tech," approaching "$1.5 billion in annualized sales"), but from selling into the build-out, not from using AI internally. That is an AI-infrastructure demand indicator, not isolated AI revenue. Both retired; rows are never deleted and both stay in the archive with the original call verbatim. (Ecolab primary)

On Klarna and JPMorgan. We searched for a primary written source for each and could not find one; both trace only to spoken remarks reported by third parties. Leaving that visible is more useful than quietly dropping the rows.

A note on this instrument. Two rows above carry a label rather than a number, because the ladder could not honestly represent them: Visa has a solid workforce-cost figure with the AI share unquantified, and Duolingo has a written P&L attribution expressed in basis points rather than dollars. Rather than force either into a score the definition does not support, we are reworking the methodology over the next few weeks — the likely direction is scoring number quality and attribution quality on separate axes. We will publish the revision in full when it is ready.

Method, so you can re-run it. EDGAR full-text search (efts.sec.gov), all form types, exact phrase "AI capital expenditures": 0 results for 19–25 July and 0 for 26 July–2 August. The companion query, 8-Ks containing both "artificial intelligence" and "restructuring plan", returned 1 filing in the first window and 2 in the second. Two is not a trend and we are not calling it one. The zero applies to that exact phrase, not to every possible way of describing AI-related capital spending. Note: EDGAR's ciks filter requires 10-digit zero-padded CIKs and silently returns zero hits otherwise.

Most AI-and-jobs coverage stops at the headcount number. That is the least interesting part, and for a capital allocator it is nearly useless — because a reduction tells you almost nothing on its own. Visa is worth studying precisely because it did the unusual thing and said what the money is for.

Look at the shape of it. The eliminations fell mostly on technology and product — not the functions that dominate AI-displacement narratives. They came while revenue grew 14% and guidance went up. And the reinvestment list is long and mostly conventional: acceptance, cross-border, risk, marketing services, B2B, Visa Direct. Two entries are not conventional. Stablecoins and agentic commerce are mechanisms by which a transaction could one day settle without touching a card network at all — or become new layers Visa orchestrates before they do. That is the actual bet: not defending old rails, but staying in the transaction path as the rails change.

Why this decides who keeps the rents. Rent is an excess return that persists after competition catches up. AI-driven efficiency inside a product organisation is a poor candidate — model access is diffusing fast, which is why Mastercard, Amex and Block are all running versions of the same reduction. What may not diffuse at the same speed is the implementation system around it: proprietary data, workflow integration, distribution, trust, acceptance. If Visa's advantage survives, it will not be because it wrote software more cheaply. The AI is the multiplier; the moat is whatever scarce thing it multiplied.

The strongest case against this reading. Three objections. First, Visa never said AI caused the reductions — it tied agentic tooling to the team redesign and left the causal share unquantified, and a prior cut of roughly 1,400 roles in October 2024 predates most of this capability. Second, "reinvesting in growth areas" is what every restructuring says; the list is cheap to publish and expensive to verify, and CFO Chris Suh separately told investors Visa expects "to continue to be able to deliver strong margins" — so this is not savings redirected wholesale away from margin. Third, we cannot confirm from outside whether the redeployment is real. The honest test is not this quarter's language but next year's capital-expenditure and hiring mix in the named areas.

The decision test. For any AI-enabled reduction: what freed budget or remaining capacity is being redirected, toward which named threat or growth bet, who owns it, and by when? If the answer is only margin, you have an efficiency case — real, possibly recurring, but not a strategic option. If it is a named bet with an owner, a date and a return test, track it like an option.

How we source the Visa story: the severance figure and financial statements come from Visa's Q3 earnings release, which is unaudited and was furnished under Item 2.02 rather than filed. The squad sizes, productivity figures, the "majority in technology and product" scope and the reinvestment list come from Visa's own corrected transcript, linked above — Primary source. The ~2,600-role and ~7% figures, and the staff-memo language, appear in no Visa-published document and remain Reported.

[How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis.] · Written & edited by Mario Suarez · Independent analysis · Every link and date verified before send.

AI Above the Cut is a weekly decision brief for executives — VP-and-up leaders in business who want signal over noise. It covers the outcomes and impact of AI rather than its engineering, and asks a standing question of every development: who keeps the rents? Each Sunday we read a fixed spine of the field's highest-signal voices — operators, researchers, and independent skeptics like Andrew Ng, Ethan Mollick, Simon Willison, Nathan Lambert, the AI Snake Oil team, Erik Brynjolfsson, and Cassie Kozyrkov — plus a rotating edge of specialists (Chip Huyen, Jack Clark, Ben Thompson, Benedict Evans, and others) and the primary research, regulator, and lab feeds. We tag every source — vendor, researcher, operator, investor, regulator, economist, or skeptic — and check strong claims across categories, so we curate evidence, implementation, and disagreement rather than celebrity.

The brief comes in two speeds: a fast skim — the single most important development, three concrete moves, and the week's decision-relevant signals plus one Below the Cut counter-signal — then the Margin-Proof Tracker, our standing scorecard of AI-value claims against reported P&L evidence, and a longer Synthesis that connects the moves, takes a position, and names the tests we're watching. We optimize for quality over influence, link to the source rather than the hype around it, and flag anything unconfirmed. No "10 AI tools you need today."