George Leimer said he recognized the image immediately.

During a recent call with the Yahoo Finance product chief, a slide showed two hamburgers side by side as a metaphor for work before and after artificial intelligence. Leimer leaned toward the screen and said, “It’s totally that.”

The slide came from Arvind Narayanan, a Princeton computer scientist who recently delivered a keynote at the International Conference on Machine Learning in Seoul. In Narayanan’s illustration, the burger labeled “Traditional” has a thick patty representing the execute layer of work — coding, debugging, and doing the task — between relatively small buns. The burger labeled “With AI” shows the patty reduced to a thin layer, while the upper bun, labeled “decide,” and the lower bun, labeled “deliver,” have grown much larger.

The metaphor describes a shift in knowledge work. The core task, once the largest part of the job, is being compressed by AI. But the human work above and below that execution layer is not disappearing. Instead, the work of deciding what should be built, and of validating, integrating, and delivering it, is expanding. In Leimer’s words, work is increasingly becoming a conversation about work.

“We come in and talk to each other and we make sure that we really have real confirmation and conviction about, like, these are the things we’re going to do. And this is how we’re going to unleash the agents,” Leimer said.

Leimer’s team led the development of AlphaSpace, a new investing product that had been live for about two months when Fortune first reported on it. In that period, the team shipped more than 100 new features, a pace Leimer said illustrates how quickly the execution layer has accelerated. Users are spending 3x more time in AlphaSpace than the average Yahoo Finance user, he told Fortune, and the team has been able to respond rapidly to feedback, including by adding real-time options data through a partnership with Unusual Whales.

Leimer said sprints that once lasted two weeks now take 24 hours: planning in the morning, shipping by the end of the day, and reviewing the work at night. In one case, user feedback arrived at 4:45 p.m. on a Friday; by 8:30 that evening, the issue had been fixed, reviewed, and deployed.

In software teams, that kind of cycle changes the meaning of an agile sprint. If AI tools make implementation faster, the scarce resource becomes not only engineering time but the team’s ability to choose the right problem, inspect the output, and decide whether it is safe and useful enough to ship.

“In the old days,” Leimer said, “that would have been, ‘We can’t launch it next week.’”

The decide-execute-deliver sandwich

Narayanan calls the framework the “decide-execute-deliver sandwich.” In his view, most knowledge work has three layers. The top layer is deciding what needs to be built and why. The middle layer is execution, or the actual implementation. The bottom layer is delivery, which includes integration, testing, verification, and long-term accountability.

AI agents are compressing the middle layer. Narayanan’s argument is that the execution layer was only ever about one-third of the overall job. The two outer layers are not merely remaining in place; they are growing.

“The first and third layers are arguably expanding as AI compresses the middle layer,” Narayanan told the ICML audience. Once building becomes cheaper and faster, it becomes easier to start projects, but harder to keep pace with deciding which projects matter and verifying that what has been produced is reliable.

As AI lowers technical barriers, roles begin to blur

For years, product teams at large technology companies were often organized around what each person could not do. Designers created static mockups and waited for engineers to translate them into working code. Product managers wrote requirements but needed developers to test whether an idea could work. Analysts understood data but did not build production systems.

Yahoo teams say those boundaries are becoming less rigid as AI automates more of the “syntax” involved in building software.

On the team behind Yahoo Scout, the company’s AI answer engine, principal product designer Nick Lockington began prompting Google’s Vertex AI directly, asking it to generate structured JSON logic for new features. An engineer later observed that Lockington had effectively created a functional API through prompting, bypassing the traditional build phase.

“It was a total ‘aha’ moment,” David Grandinetti, a distinguished software apps engineer on the project, said on Yahoo’s blog. “Nick became immediately the most leveraged engineer on our team, because he was working at the highest level of abstraction.”

Other members of the team followed similar paths. One designer who became prolific in code commits was internally described as a “design engineer.” An engineer with a strong sense of usability was called an “engineering designer.”

Yahoo executives say the point is not that job titles are disappearing. Rather, roles are increasingly being defined by the decisions people own, not only by the code they can manually produce. Designers still make judgments about user behavior and aesthetics. Engineers still own architecture and safety. But both groups can now build and iterate directly with AI tools.

“This is about increasing decision velocity and cutting the cost of experimentation, so teams can test a hundred ideas instead of five and quickly discard failures without the traditional burden of technical debt,” said Stephane Koenig, vice president at Yahoo. “It allows us to reserve human expertise for the hardest decisions where you need context AI doesn’t have.”

Leimer made a similar point in his conversation with Fortune: work is now largely a conversation about work. “We come in and talk to each other and we make sure that we really have real confirmation and conviction about like, these are the things we’re going to do,” he said.

The risk, Leimer added, is that teams now need to be especially “really thoughtful” about the top layer of work, because AI makes it easy to “just build whatever you want.”

Leimer cited advice from his manager, Yahoo Media Group President Ryan Spoon: “It’s important that you’re really convicted about what you build,” because teams can build almost anything, almost instantly. In the past, he said, organizations spent significant time on the top bun because engineering resources were scarce. “You didn’t want to waste one second of an engineer’s time.” Now, the risk is building unnecessary things instantly — a pattern McKinsey’s global technology and AI leader Kate Smaje previously described to Fortune as the “false productivity” trap.

Leimer said that when he sees sloppy AI-generated work in the world — often called “AI slop” — it reflects “where there’s a really skinny bun, basically.” In those cases, humans have not been sufficiently deliberate about what they are releasing as work product, because the execution part of the hamburger has been compressed almost to nothing.

There are measurable costs to that dynamic. Professionals have begun using the term “workslop” to describe AI-generated emails or work products sent for review without proper verification. Experts writing in the Harvard Business Review have argued that the problem is not merely irritating but destructive to productivity. The Stanford Social Media Lab estimated the cost at $9 million per year for a company with 10,000 workers, according to a report cited by Techdirt.

That concern is especially relevant as companies move from AI demos to everyday workflows. A draft, prototype, or generated feature may look complete before it has been checked against user needs, security requirements, data accuracy, and the organization’s own standards.

Decision fatigue becomes a new constraint

Together, these examples support Narayanan’s view that AI is currently more of a collaboration technology than a full automation engine. He argues that most AI agents still struggle with reliability in high-stakes environments, leading companies to favor human-in-the-loop systems rather than fully automated processes.

At Yahoo, leaders say the result is showing up as decision fatigue rather than immediate job loss. Because AI enables teams to test and build many more features, the bottleneck becomes the number of decisions that product managers, designers, and engineers can thoughtfully make, review, and guide to completion.

Lockington said AI does not eliminate the complexity of overlapping workplace roles. “Traditionally, you might have someone ask, ‘Why is this engineer doing design work?’ or ‘Why is this designer doing my job?’” he said. He credited Koenig with creating the psychological safety needed for the team to navigate those questions. “We’ve been able to break that down.”

Leimer said that once that safety exists, the work becomes “really fun.” After roughly 20 years in management, which moved him away from the building and coding work where he started, he said the shift has changed how people experience product development. “Now, everybody’s got the cool job,” he said, because everyone can feel they are contributing directly to building. “It’s also been fun watching people get energized around the problem and the thing they’re building as opposed to spending more of their time on the process to get it built.”

The future of the hamburger — and the case for pizza

Narayanan cautions against assuming jobs will remain unchanged. He has said he expects decades of structural adaptation as companies reorganize around AI, comparing it to the 40-year transition factories underwent after electricity enabled assembly lines.

So far, however, the pattern in fields such as software, law, medicine, and content translation is that demand for human work shifts and grows as AI spreads. Radiologists have adopted AI while employment has risen. Lawyers are filing more suits because drafting is easier. Translators continue to have steady work years after computer translation approached human parity, because there is no fixed ceiling on what can be translated or into how many languages.

Yahoo’s experience suggests white-collar work may follow a similar, if messier, path. As technical barriers decline, company leaders say personal agency and judgment are becoming more important differentiators: not who can write code, but who can decide which problems are worth solving and how to verify what AI produces.

Leimer was asked to extend the food metaphor further, including a reference to his Philadelphia roots and the city’s famous cheesesteak. If the defining feature of a cheesesteak is not only the meat but the Italian hoagie bread, the question was whether that also applies to the future of work.

Leimer said the bigger bottleneck is the “bottom bun,” or the delivery layer. After a team decides what to do, the work may be produced almost instantly, but delivery remains difficult. “You’ve got to work through all the other stuff, you’ve got to review it,” he said, referring to security risks and the need to build a deployment pipeline, even when the work appears ready to release with the press of a button.

For products tied to financial information, that delivery layer can include data quality, permissions, reliability, and user trust as well as software deployment. Those checks do not vanish just because the code or interface can be generated faster.

Asked whether the destination is something closer to pizza or flatbread — where human work is concentrated in decisions, tasks shrink to the scale of toppings, and delivery becomes seamless — Leimer laughed and said, “Something like that.”

In Narayanan’s hamburger metaphor, AI has already thinned the patty. The expanding buns are where the jobs, responsibilities, and risks now sit. Whether companies can redesign work into something flatter and more sustainable remains an open question for the coming years, and perhaps decades.

This story was originally featured on Fortune.com.