VISA

Plenty of AI programs stop at buying licenses and counting seats. What Visa did is harder to undo. It is shrinking the unit that builds a product.

On the 28 July earnings call, McInerney said Visa is "reforming our product development teams that used to be ten or more into smaller and more nimble agentic squads of two to four." Underneath it is a set of AI tools wired together so that work passes from one step to the next without a person moving it along, with people still checking what comes out.

Cutting a product team to a fraction of its size depends on more than staffing. It only works if the pipeline absorbs the work those people were doing, and if somebody is still reviewing the output. McInerney put the review part on the record. How much of the departed work the pipeline absorbs is not something Visa has said.

On the same call, Visa put its own count at "more than 150 AI-powered applications" today and "more than 300 major product releases" shipped over the previous 12 months. That second figure is a trailing-year flow rather than a running total, and the releases are described as major rather than AI, so the 300 counts output and not AI output.

McInerney gave the velocity figures in one sentence on that same call: "And the results are meaningful for those teams that are using the agentic toolchain, with 80% more code commits and an 80%-plus improvement in requirement definition from 30 days to 5 days, which has translated to 65%-plus faster feature development."

All three describe one population, the teams already on the toolchain, and Visa did not say what share of its engineers that is. Asked later about adoption, McInerney said deployment had gone well "across the company especially in product and engineering" and that other parts were getting good adoption. He put no percentage on either.

These are before-and-after numbers on a self-selected group of unknown size, and we'd still take them over a seat count, because one is a clean task measurement. Requirement definition going from 30 days to 5 names a job, a baseline and a unit of time. Code commits are just activity, and they can rise without more work getting done. CFO Christopher Suh cited faster consulting projects too, but gave no figure.

Visa's fiscal third-quarter results carry $563 million in severance charges. McInerney told investors the same day that the cuts were "the majority being in our technology and product teams," the same teams the velocity numbers come from. Visa drew no line between the two, and companies cut roles for reasons that have nothing to do with AI. What is on the record is the timing and the department.

Reading Visa's fiscal 2025 annual report, its fiscal 2026 quarterly releases and its newsroom, we found no dollar figure for what Visa's internal AI use returns. Every AI dollar figure we found is either money going out, such as the more than $3 billion on AI and data infrastructure over ten years that Rajat Taneja, Visa's president of technology, cited in 2023, or fraud stopped for clients, such as the $40 billion of fraudulent activity Visa says it helped block in fiscal 2023.

That leaves severance as the one number with a dollar sign on it anywhere near this work. Our Chevron issue found measured savings on one robotics program, which is the kind of number Visa has not yet put beside its velocity work. Full-year results land on 27 October, the next chance for one to appear.

The velocity figures live in a transcript. Visa's adoption numbers sit in an SEC filing, which makes them the most durable evidence in this issue.

The annual report for fiscal 2025 describes an internal GenAI hub for employees and, separately, Visa's own AI assistant for employees, which it says uses agentic AI for self-service. From the assistant's launch in March 2025 through the fiscal year ended 30 September 2025, the filing says, it "has been embraced by nearly 26,000 employees launching over 261,000 AI-powered chats."

That same section gives the denominator, with the workforce growing from roughly 31,600 in fiscal 2024 to roughly 34,100 in fiscal 2025. Nearly 26,000 of about 34,100 people is around three quarters of the company, inside seven months of launch. Put that fraction beside your own.

Nearly 26,000 employees and 261,000 chats works out to about ten each in seven months, which is broad reach and light use.

Our Mastercard issue makes the useful comparison. Mastercard gave a target for how many of its staff it intends to train. Visa gave a count of who actually turned up. One is a plan and the other is a result, and your board will ask which kind yours is.

Visa says it took part in Project Glasswing, the security effort Anthropic announced in April 2026, and pointed frontier models at its own systems. Anthropic doesn't name Visa among its twelve launch partners, but says more than forty further organizations got access. The result was the Visa Vulnerability Agentic Harness, which McInerney called "an orchestration layer that allows us to use models like Mythos to find and fix issues at AI speed." Visa open-sourced it in June 2026, and an August release extended it into remediation and validation, with tens of thousands of downloads. Taneja said, "AI is compressing the time between vulnerability discovery and exploitation."

The repository is public and free to use, and when we read it on 11 October it showed 2.4k stars and 323 forks. Anyone can download it and run it against their own code. One warning sits in Visa's own instructions. Left on its default setting the tool rewrites the files it is pointed at, so a first run needs the flag that tells it to report and change nothing.

Point this at one of your own AI productivity numbers before it reaches a board deck. It finds the narrowest population the number is actually true for.

Get more such prompts in the Prompting Playbook (free for you)