On Tuesday, July 28, 2026, roughly 2,600 Visa employees learned they no longer had jobs. The notice arrived by memo from chief executive Ryan McInerney, hours before the company reported fiscal third quarter results. The cuts amount to about seven percent of a workforce that stood at roughly 34,100 at the close of the last fiscal year, and the majority fall on product and technology teams.
Here is the part that did not make the coverage. Visa’s own earnings release, published that afternoon, records a severance charge of $563 million sitting in personnel expense for the three months ended June 30, 2026. The three-month figure and the nine-month figure are identical, which means the entire charge landed inside the June quarter. Nobody was told until July 28.
Chief financial officer Chris Suh described the charge on the call as covering changes to Visa’s workforce including the ones McInerney had just announced, so some portion of it may relate to other actions. Fine. The number is what it is, and the quarter it sits in is what it is.
Accruing severance once a plan becomes probable and estimable is ordinary accounting under an ongoing benefit arrangement. That is not the point. The point is what the accounting establishes about the calendar. By the last day of June, Visa had decided enough about who was leaving to put a number on it and defend that number to an auditor. Those people then worked for up to four more weeks. They attended standups. They filed pull requests. Some of them, as we will see, shipped the code that became the productivity statistic their chief executive would recite on the earnings call as justification for their removal.
Then, on the afternoon of July 28, Visa’s board of directors declared a quarterly cash dividend of $0.670 per share. Same day. Same filing. A few paragraphs apart.
What the quarter actually looked like
The framing on the call was efficiency. It is worth establishing what Visa was being efficient about.
Net revenue for the quarter was $11.63 billion, up fourteen percent. Payments volume crossed four trillion dollars for the first time in the company’s history. GAAP net income was $5.6 billion. Visa repurchased $4.9 billion of its own stock during those same three months at an average of $330.71 a share, paid $1.3 billion in dividends, and closed the quarter with $28.4 billion of repurchase authorization still unused. Across nine months the buybacks total $16.4 billion.
Set the severance charge against the buyback and you get the scale of the thing. The entire cost of removing 2,600 people from the company amounts to roughly eleven percent of what Visa spent buying its own shares in a single quarter. Not eleven percent of the year. Eleven percent of the quarter.
Strip the severance out of personnel expense and the underlying figure rose about eight percent year over year, roughly in line with headcount, and Suh noted that even that was inflated by a deferred compensation mark-to-market he described as neutral to earnings. There was no compensation crisis. Marketing expense went from $421 million to $649 million, a fifty-four percent increase driven largely by World Cup activation, and in fairness a good deal of that is billable to clients through marketing services, which helped push other revenue up forty-five percent. Operating expenses grew seventeen percent on a non-GAAP basis, and nineteen percent on a GAAP basis, against fourteen percent revenue growth. That gap is the genuine pressure, and it is a gap the company built itself over ten years of hiring.
Because that is the other thing the filings show. Visa employed roughly thirteen thousand people in fiscal 2016 and 34,100 by fiscal 2025. Revenue per employee never broke out of a band of roughly $1.1 million to $1.3 million across that entire run, which means the hiring was neither wildly productive nor wildly wasteful. It was a decision, made repeatedly, by the same executive committee now describing its reversal as a technological inevitability.
The metrics that were not reported
McInerney’s case for the cuts rests on a specific claim about how Visa now builds software. Product teams that used to run ten or more people are being reformed into what he called agentic squads of two to four. The supporting numbers, delivered on the call: eighty percent more code commits, requirements definition compressed from thirty days to five, better than sixty-five percent faster feature development, more than 150 AI-powered applications, and more than three hundred major product releases across twelve months.
Every one of those is a throughput measure.
The industry-standard framework for software delivery performance, the one that came out of years of DORA research, uses four metrics arranged in two pairs. Deployment frequency and lead time for changes measure how fast you move. Change failure rate and time to restore service measure whether what you shipped works. The central finding of that research is that elite teams score well on both pairs simultaneously, and that throughput gains accompanied by deteriorating stability are the signature of an organization accelerating in the wrong direction.
Visa reported the first pair. It said nothing about defect escape rate, rollback frequency, incident volume, change failure rate, or mean time to restore. For a company carrying authorization traffic across more than two hundred countries and territories, those are not footnotes. They are the entire question. Nobody on the call asked.
I have written requirements documents for systems that move money, and I want to be direct about the thirty-days-to-five number, because it is the one that should worry you rather than impress you. In payments, requirements definition is the engineering. Thirty days is not a committee. Thirty days is where somebody finally notices that the dispute window in one market runs on a different clock than the one you built for, that a settlement job breaks across a leap second, that the product marketing just added quietly invalidates a chargeback flow which has been running fine for six years.
That work does not compress because a model can draft the document faster. The document was never the hard part. The hard part was the several experienced people who had to read it, each of whom recognized exactly one thing that would have taken down production. Remove them and the document still gets written in five days. The recognition is what you lost, and you will not find out what it cost you this quarter.
So either those thirty days were genuine committee theater, which is entirely plausible inside a 34,000-person company, or the discovery work was not eliminated at all. It was relocated to production. A sixfold compression reported without any accompanying change in defect rate is an assertion. It is not a result.
And three hundred major product releases in twelve months is roughly six a week. The phrase major product release is carrying an extraordinary amount of weight in that sentence.
The eighty percent was earned by the people who were fired
There is a timing problem inside McInerney’s own argument, and once you see it the argument does not survive.
The productivity figures he recited on July 28 describe work performed before the layoffs. They have to. Visa’s fiscal third quarter ran April through June. The severance was accrued by June 30. The announcement came four weeks after that. Every commit inside that eighty percent was produced by an organization at full strength. Visa has not said which teams were cut, so I cannot tell you precisely how much of that eighty percent came from people who are now gone, but the cuts fell mainly on product and technology and the metric came out of product development. Draw your own conclusion.
What is not in dispute is the sequence. That figure is not a projection. It is a baseline, generated by a workforce that no longer exists.
A metric produced at one headcount cannot justify a different headcount. If teams of ten or more using the new tooling delivered eighty percent more commits, the defensible conclusion is that teams of ten with agentic tooling are highly productive. It does not follow that two people can now do the same work. Nobody has measured that. Visa announced the reduction and the productivity gain in the same breath, and the gain was measured before the reduction. Investors were handed a forecast dressed as a result.
There is a second reading, and it is no better. Suppose the squads had already been reformed to two-to-four when those metrics were captured. Then the 2,600 were surplus before any of this, the reorganization produced the reduction rather than the technology, and the artificial intelligence framing is decoration on a decision made earlier for other reasons. Either reading breaks the causal claim McInerney made from the podium. He cannot hold both.
1.9 trillion tokens, and the number Visa did not publish
Now the part that makes the eighty percent genuinely difficult to trust.
In April, Visa’s president of technology, Rajat Taneja, told Business Insider that the company was consuming 1.9 trillion artificial intelligence tokens a month as of March, double what it had consumed in February. He said eighty-nine percent of Visa employees were active AI users and that forty-four percent qualified as power users, a status the company defines as averaging at least twenty-five prompts per day for fifteen days a month. Claude was the most popular model internally, trading places with ChatGPT. And Visa was running a rewards program for teams that used AI to ship faster. One team that used Claude Sonnet to deliver a new API in under six days won such an award. Teams pick their own prizes, in internal company points, redeemable for things like a coffee maker.
Read that sequence with the July memo in hand. In March, Visa doubled its token consumption in a single month, defined power-user status by prompt volume, tracked it at the level of the individual, and handed out points for velocity. In June, it accrued $563 million of severance. In July, it cited productivity gains as grounds for cutting 2,600 people, most of them from the same technology and product organization that had been generating those prompts.
I want to be careful here, because the tempting accusation is not the one the evidence supports. There is no reporting that Visa threatened anyone over token consumption, and that is not the claim I am making. What the record shows is something structurally worse than coercion. It shows an incentive.
If you define a measurable target, track it per person, and attach a reward to it, you do not need to threaten anybody. This is the oldest failure in management measurement and the artificial intelligence version of it already has a name. It is called tokenmaxxing, and the critique is straightforward: more prompts do not mean better work, and once usage becomes visible it becomes a proxy for activity rather than for results. As Mercer’s Ravin Jesuthasan put it, artificial intelligence usage is a very poor proxy for productivity.
Visa is not unusual in this, which is the more troubling fact. Reporting on the practice this spring found that nearly every Fortune 500 company now tracks aggregate AI usage, and that at some firms internal platforms break consumption down to the individual with project leaders able to see it, turning adoption into a leaderboard contest. Practitioners in that same reporting pushed back, with EY’s Sameer Gupta arguing that most companies still measure at the group or role level and focus on outcomes rather than monitoring individuals. Visa’s published power-user definition, twenty-five prompts a day, fifteen days a month, is an individual-level metric. Meta has made AI-driven impact a core expectation for 2026. Across the economy McKinsey found that while sixty-four percent of companies say AI is driving innovation, only thirty-nine percent report a measurable impact on earnings.
Which brings us to the number Visa did not publish.
The company disclosed its token consumption. It disclosed the doubling. It disclosed the commit increase, the compressed requirements cycle, the count of AI-powered applications, the count of major releases, and the percentage of its workforce classified as power users. It did not disclose what 1.9 trillion tokens a month costs. It did not disclose a single stability metric. Every figure Visa volunteered is a figure that goes up. Not one of them constrains anything.
Those two omissions are the ones that determine whether any of this worked. A commit increase purchased with a doubled inference bill is not efficiency, it is substitution, and the trade cannot be evaluated without both sides of it. Visa took a $563 million charge to remove the labor while adding a recurring machine cost it declined to quantify for shareholders. That is a strange thing to leave out of an efficiency narrative delivered on an earnings call.
Others are further along in asking. Business Insider has reported companies retreating from tokenmaxxing once the bills arrived, Coinbase routing some prompts to cheaper models, and Uber executives questioning whether their spending on AI coding tools was producing enough measurable product gain to justify itself. Uber is the company usually invoked as the cautionary tale about shipping fast and sorting out the consequences later. It is now the company asking the question Visa’s board did not.
Visa’s own documentation disagrees with Visa’s chief executive
There is a document that settles the question of what Visa’s engineers actually believe about agentic output, and Visa published it.
In June the company open-sourced the Visa Vulnerability Agentic Harness under Apache 2.0, an agentic security pipeline built on what it learned running Anthropic’s Mythos model against its own payment network as part of Project Glasswing. It is good work. Four phases, eleven stages, threat modeling before analysis, multi-agent deterministic voting to suppress false positives, deterministic gates wrapped around the non-deterministic parts. It is roughly how you should build this.
Now read the documentation. Visa’s own engineers state plainly that the findings are LLM-generated triage candidates rather than confirmed vulnerabilities, that human review is required, and that runs are non-deterministic, so two scans of the same repository may differ.
That is Visa’s engineering position on agentic output, published by Visa, in June, under the same executive who had reported the token figures in April. Six weeks later the chief executive cited agentic tooling as grounds for removing 2,600 people, most of them from product and technology. The company that wrote human review is required into its own README cut the humans.
What a two-to-four person team does to a control environment
The squad math has a second problem, and it is considerably more likely to produce a regulatory finding than a bad user experience.
A ten-person team produces separation of duties as a byproduct of its size. Somebody writes, somebody independently reviews, somebody tests, somebody approves the release. Compress that to two-to-four people with agents generating the bulk of the code and you are asking one human to author while another meaningfully reviews work that neither of them wrote. PCI DSS change control assumes that separation exists. So does SOX control testing. Neither framework contains a provision for the model reviewed it.
Visa’s own risk factors list exposure to significant loss through its indemnification obligation to fund settlement losses of its clients. That is the mechanism. Not product liability in the consumer sense, but a control environment finding, then an incident, then client claims, then plaintiffs’ counsel reading the July 28 transcript back to a jury and asking why management was quantifying productivity gains rather than failure rates.
The industry has already produced the demonstration. In mid-December, according to Financial Times reporting sourced to four people familiar with the matter, engineers at Amazon Web Services handed a bug in AWS Cost Explorer to Kiro, Amazon’s internal agentic coding tool. Rather than patch the bug, the agent concluded that the efficient path was to delete and recreate the environment. It did. The result was a roughly thirteen-hour outage in one of AWS’s China regions. Kiro normally requires two human approvals before acting, but the agent inherited its operator’s permissions and that operator’s access was broader than intended, so the gate never engaged. A senior AWS employee told the paper the company had already seen at least two production outages of this kind and described them as small but entirely foreseeable.
Amazon’s public position is that this was user error rather than AI error, and that the involvement of AI tooling was incidental. Take that at face value if you like. It does not help Visa, because look at what Amazon did next. It began requiring senior engineer sign-off on all AI-assisted code changes from junior and mid-level developers. Confronted with an agent that acted autonomously inside a permissions gap, the largest cloud provider on earth added a layer of experienced human review.
Visa is removing seven percent of its people, mostly from product and technology, and compressing its remaining teams to two-to-four. It is running the opposite play, at a company whose failure mode is not a dashboard going dark but a settlement not clearing.
Uber gets invoked as the cautionary tale here and it is the wrong comparison for a different reason than the token story suggests. Uber’s liabilities came from strategy rather than sloppiness. Greyball, driver classification, the concealed 2016 breach were decisions, and the engineering underneath them was largely competent. The failure mode Visa is courting is worse in one specific respect. Control erosion inside a regulated financial entity is invisible until the incident, and when the incident arrives there is no villain available to fire. It looks like nothing right up until it looks like everything.
The question nobody asked
The earnings call ran through thirteen analyst questions. Cross-border volume. The World Cup’s effect on travel spend. The delta between cross-border volume and international transaction revenue. Pismo positioning for large banks versus small. The revenue algorithm. OpenUSD against Circle and Tether. European scheme competition. Value-added services growth. United States volume composition. Cross-border e-commerce in July. Agentic commerce. One question on AI adoption, which asked what inning we were in.
One analyst raised the 2,600 people. Sanjay Sakhrani of KBW asked whether the savings would be reinvested or dropped to the bottom line, and over what period a return would appear. That is a reasonable question from a sell-side seat. It is also, in its entirety, a question about the money.
McInerney’s answer contained no savings figure. He described the investment opportunities in front of the company as enormous, listed them across consumer payments, value-added services, commercial payments, stablecoins and agentic commerce, and observed that the track record is what the track record is. His chief financial officer added that the flywheel is working. Nobody asked what the change failure rate looks like inside the new squads. Nobody asked what the inference bill runs. Nobody asked why the accrual preceded the announcement. Nobody asked about the people. The stock closed up one and a quarter percent and handed most of it back after hours.
The offshoring the narrative omits
There is one more thread worth pulling, because it complicates the artificial intelligence story considerably.
The share of Visa’s workforce based outside the United States has risen sharply this decade, from roughly half to roughly sixty percent by 2025. Third-party workforce data has Colombia as the company’s fastest-growing location. Employees describe North American customer service functions relocating to Toronto and Bogotá, though that account comes from anonymous forum posts rather than from Visa. None of it appeared in the memo, and none of it came up on the call.
Labor arbitrage and AI substitution produce identical headcount charts. They produce very different stories. One of them makes a chief executive sound like the steward of an inevitable technological transition. The other makes him sound like every cost-cutting executive of the last thirty years. The available evidence supports both readings, and Visa elected to narrate only one.
This is also not the first performance of it. In October 2024, Visa cut approximately 1,400 employees and contractors, roughly a thousand of them in technology, announced on the day of its fiscal fourth quarter earnings. The company recorded $213 million in severance across the first nine months of fiscal 2025. The 2026 round is the same play, on the same day of the reporting calendar, at nearly twice the scale. Two rounds in twenty-one months is not a correction. It is an operating model.
The industry alibi
Visa is not alone, which is precisely the problem.
Mastercard cut roughly 1,500 people, about four percent of its workforce, in January, using near-identical language about freeing capital for reinvestment. Block cut four thousand in February amid an AI push. PayPal cut roughly 4,700 in May. Visa is the fourth major payments company to run this sequence in seven months.
The wider pattern suggests the explanation has become load-bearing well beyond what the evidence supports. Artificial intelligence is now named in a substantial share of announced United States job cuts, on the order of a fifth to a quarter of those tracked by outplacement firm Challenger, Gray and Christmas. Deutsche Bank analysts warned in January that AI layoff washing would be a defining feature of the year. A National Bureau of Economic Research working paper co-authored by Stanford’s Nicholas Bloom surveyed close to six thousand senior executives across four countries and found that the overwhelming majority reported artificial intelligence had produced no employment impact at their own firm, while readily believing it was happening elsewhere. And Gartner, surveying hundreds of large companies already deploying AI agents, found that the great majority had reduced headcount, and that the heaviest cutters showed no superior financial returns for having done so.
The people absorbing this are landing in the worst technical labor market in a generation. Tech-sector unemployment is at levels not seen since the dot-com collapse, and the median time for a laid-off technology worker to find a new role has stretched by well over a year’s worth of runway since 2024. That is the interval Visa handed 2,600 households, having already carried the cost of it on its books for four weeks.
What accountability would have looked like
None of this is an argument that Visa cannot reduce headcount. It is a public company. It over-hired for the better part of a decade and the correction was arriving with or without a language model.
The argument is about what its chief executive and its board chose to say while doing it, and what they carefully chose not to measure.
An honest version of the July 28 memo was available. We hired too aggressively from 2021 through 2024 and revenue per employee never improved. We are correcting that. We are also moving work to lower-cost geographies, and separately we are seeing gains from agentic tooling in engineering. Here is the change failure rate before and after. Here is what our inference bill runs and how it compares to the personnel cost it replaced. Here is what the savings are worth. Here is what the board intends to do with them.
That memo would have been accountable. Instead, 2,600 people were priced in June, informed in July, and told the reason was a technology whose limits their own employer had documented in a public repository six weeks earlier, measured by a statistic they themselves had produced, inflated by an incentive program their employer had built. Their removal was announced on the same afternoon the board declared a dividend and the company reported crossing four trillion dollars in quarterly payments volume for the first time in its history.
Ryan McInerney has a board that approved all of it. Boards exist to ask the question thirteen analysts did not. This one declared a dividend instead.
Code commits are not a product. Prompts are not output. Burning tokens to make things that kind of work is not a roadmap. Velocity reported without a failure rate is not an engineering result, and a productivity gain measured before a layoff is not a justification for the layoff. An inevitability that was booked to the general ledger four weeks before anyone was told is not an inevitability at all. It is a decision, made by people with names, and they should be asked to defend it in exactly those terms.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment. He has spent close to thirty years shipping production software on the Microsoft stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy and Allegiant Air, and was employee number six at Break.com. He is a Los Angeles Business Journal CTO of the Year finalist. His current work is production agentic AI — MCP servers, Microsoft Agent Framework, RAG pipelines, SQL Server 2025 vector search, and the PLogger observability framework. He is the author of the novel The Archive of Lost Suns, trains Brazilian jiu-jitsu under Sergio Penha in Las Vegas, rides motorcycles, and co-hosts Edge Grip Podcast. He writes about what actually ships, and about what does not.
A note on sources
Every Visa financial figure here comes from the company’s fiscal third quarter 2026 earnings release dated July 28, 2026, including the $563 million severance charge and its placement in the three months ended June 30, the $4.9 billion of quarterly repurchases, the $28.4 billion of remaining authorization, and the board’s declaration of a $0.670 per share dividend on July 28. Executive statements about agentic squads, code commits, requirements compression and product releases are drawn from the fiscal third quarter 2026 earnings call of the same date, as is the exchange with Sanjay Sakhrani of KBW and the characterization of the severance charge by chief financial officer Chris Suh. Token consumption, power-user definitions, the internal rewards program and the Claude Sonnet API example were reported by Business Insider in April 2026, sourced to Visa president of technology Rajat Taneja. Headcount figures are from Visa’s annual reports on Form 10-K; the international workforce share and revenue-per-employee band are calculated from those filings by third parties. The characterization of agentic output as requiring human review comes from Visa’s own public documentation for the Visa Vulnerability Agentic Harness. The AWS Cost Explorer outage was reported by the Financial Times in February 2026; Amazon disputes that AI tooling was the cause. Layoff totals for 2024 and for Mastercard, Block and PayPal are as reported by The Wall Street Journal, Bloomberg and Fast Company. Enterprise AI tracking practices are as reported by CNBC. Remaining labor market and survey figures are attributed inline to Challenger, Gray and Christmas, Deutsche Bank, Gartner, McKinsey, and the National Bureau of Economic Research.