The traditional view of knowledge work resembled a hearty hamburger: a thick patty of execution sandwiched between thin buns of planning and delivery. In that model, the bulk of effort went into writing code, debugging prototypes, or manually processing data, while deciding what to build and verifying the outcome occupied relatively little time. Recent advances in generative AI have begun to shrink that central patty to a sliver, as algorithms can now generate functional snippets, draft reports, or even design basic interfaces in seconds. The consequence is not a reduction of human effort overall, but a shift in where that effort is spent. Teams now find themselves allocating far more energy to the top bun—clarifying objectives, aligning stakeholders, and convincing themselves that a chosen direction is worth pursuing—while the bottom bun, encompassing testing, deployment, and long‑term accountability, has also grown in complexity. This transformation is not merely a curiosity; it reflects a fundamental reallocation of cognitive load that demands new habits, metrics, and organizational designs.
Arvind Narayanan’s decide‑execute‑deliver sandwich offers a useful lens for understanding this shift. The decide layer sits at the top and involves grasping the problem space, defining success criteria, and prioritizing initiatives based on strategic value. The execute layer, traditionally the meat of the work, encompasses the actual construction—coding, drafting, building prototypes. The deliver layer at the bottom covers integration, quality assurance, security reviews, and the ongoing stewardship of shipped products. AI agents excel at compressing the execute layer because they are adept at pattern‑based generation and rapid iteration. However, Narayanan emphasizes that execution was never more than about one‑third of the total workload in many knowledge‑intensive roles. As the cost of building drops, the remaining two‑thirds—decision making and delivery—become the dominant factors determining velocity and quality. Consequently, organizations that focus solely on accelerating code production risk overlooking the growing burden of thoughtful choice and reliable rollout.
Yahoo’s recent experience with AlphaSpace illustrates how quickly the execute layer can collapse when AI is embedded into the workflow. Over a two‑month span, the product team released more than a hundred new features, a pace that would have been unthinkable a few years ago. Users now spend three times longer engaging with AlphaSpace compared to the legacy Yahoo Finance experience, indicating that rapid delivery of value drives deeper adoption. The team’s ability to incorporate feedback within hours—turning a Friday evening request into a live fix by nightfall—demonstrates the power of collapsing the execute cycle. Yet this speed also surfaces a new challenge: the sheer volume of decisions that must be made, reviewed, and communicated has exploded. Product managers, designers, and engineers now spend large portions of their day in meetings, writing specifications, and confirming that AI‑generated outputs align with business intent, rather than merely waiting for a developer to finish a ticket.
The blurring of traditional role boundaries is another hallmark of this evolving landscape. At Yahoo, a principal product designer began prompting Vertex AI to produce structured JSON logic for new features, effectively creating a functional API without writing a line of conventional code. Engineers on the team noticed that this designer was now operating at a higher level of abstraction, contributing directly to the system’s core logic. Elsewhere, a designer who frequently committed code earned the informal title of “design engineer,” while an engineer with a keen eye for usability became known as an “engineering designer.” These hybrid roles signal that the value of an individual is increasingly measured by the decisions they own and the outcomes they enable, not by the specific syntax they can manually produce. Leaders emphasize that job titles are not disappearing; rather, the criteria for success are shifting toward judgment, context‑setting, and the ability to direct AI tools toward meaningful objectives.
While the acceleration of execution brings undeniable benefits, it also introduces notable risks that leaders must guard against. The ease with which AI can generate proposals, drafts, or prototypes creates a temptation to “just build whatever you want,” a phenomenon dubbed false productivity. When teams skip rigorous validation and ship AI‑generated outputs without proper scrutiny, the result is often termed AI slop or workslop—work that looks complete but lacks the necessary depth, accuracy, or alignment with goals. Studies from the Harvard Business Review suggest that such unverified outputs can erode productivity, and the Stanford Social Media Lab estimated that a company of ten thousand workers could lose roughly nine million dollars annually to rework caused by preventable errors. The danger lies not in the AI’s capability but in the human tendency to under‑invest in the decide and deliver layers when the middle becomes trivially cheap.
Counteracting this tendency requires deliberate cultural and procedural safeguards. Ryan Spoon, President of Yahoo Media Group, advises his teams to cultivate deep conviction about what they choose to build, reminding them that instant capability does not obviate the need for thoughtful selection. This conviction stems from a clear understanding of customer pain points, strategic priorities, and the potential downstream consequences of a feature. Complementing this mindset, Stephane Koenig, a Yahoo vice president, highlights the importance of decision velocity—rapidly testing many ideas while discarding failures quickly—paired with strong verification mechanisms. Creating psychological safety is essential; when team members feel comfortable questioning AI‑generated outputs and proposing alternatives, the group can catch subtle flaws before they propagate. Leaders who invest in building this safety net find that the work becomes more enjoyable, as individuals feel empowered to shape outcomes rather than merely executing predefined tasks.
Looking beyond any single company, empirical evidence from multiple sectors reinforces the notion that AI is currently acting more as a collaborator than a wholesale replacement. Radiologists, for instance, have integrated AI‑assisted imaging tools into their workflows, yet employment in the field continues to rise as professionals focus on complex case interpretation and patient communication. Lawyers are filing more suits because AI‑driven contract review and legal research lower the cost of drafting, enabling them to pursue a broader range of matters. Translators, despite years of near‑human‑parity machine translation, still enjoy steady demand because the volume of content requiring localization keeps expanding, and nuanced cultural adaptation remains a distinctly human strength. These patterns suggest that as AI handles repetitive, syntax‑heavy tasks, human labor migrates toward higher‑order judgment, creativity, and relational work—areas where contextual understanding and ethical reasoning are indispensable.
The delivery layer, often represented by the bottom bun, emerges as a persistent bottleneck even as execution speeds up. Once a decision is made and an AI model produces a candidate solution, the work of ensuring security, compliance, performance, and reliable deployment can still consume considerable time and expertise. Teams must construct robust pipelines, conduct penetration tests, verify data provenance, and monitor for drift—activities that resist full automation because they depend on evolving threat landscapes, regulatory shifts, and organizational policies. In practice, a feature that an AI can generate in minutes may require hours or days of review, staging, and rollout coordination before it reaches end users. Recognizing this reality helps set realistic expectations: accelerating the middle layer does not eliminate the need for careful, often manual, stewardship of the final product.
To capture the ideal balance, some thinkers propose a pizza metaphor rather than a hamburger. Imagine a thin, crisp crust representing the now‑minimal execute layer—just enough base to hold everything together. The toppings, varied and abundant, symbolize the decide layer: strategic choices, customer insights, experimental hypotheses, and creative directions that give the pie its distinctive flavor. Finally, the baking process—uniform heat, precise timing, and a reliable oven—embodies the deliver layer: automated testing, secure deployment pipelines, and observability that ensure the final product emerges consistently delicious and safe to consume. In this view, the human contribution is spread across the surface (decision toppings) and the underlying process (baking), while the AI‑powered crust provides a lightweight, reproducible foundation.
Translating these insights into concrete actions requires a dual focus on organizational systems and individual skill sets. Leaders should invest in decision‑making frameworks that make criteria explicit—such as weighted scoring models, opportunity‑solution trees, or predefined success metrics—to reduce reliance on gut feeling alone. Simultaneously, they must automate as much of the delivery pipeline as feasible, integrating continuous security scanning, automated compliance checks, and progressive rollout techniques that can keep pace with rapid experimentation. Cross‑training initiatives that teach designers basic prompting strategies and engineers core usability principles help blur the old silos, fostering a shared language for evaluating AI outputs. Finally, establishing regular “bun reviews”—dedicated sessions to scrutinize decisions and verify deliveries—creates a rhythm that prevents the decision and delivery layers from becoming neglected afterthoughts.
For individual contributors, the path forward centers on cultivating judgment, mastering the art of effective prompting, and honing validation techniques. Professionals should practice asking “why” before “how,” ensuring that any AI‑generated artifact ties back to a clear business objective or user need. Learning to craft precise, context‑rich prompts—leveraging few‑shot examples, specifying output formats, and defining constraints—maximizes the relevance and quality of AI assistance. Equally important is developing a habit of rigorous verification: comparing AI outputs against ground‑truth data, running edge‑case tests, and soliciting feedback from diverse stakeholders before considering work complete. By strengthening these complementary skills, knowledge workers can thrive in an environment where the middle layer is thin but the buns are substantial and consequential.
In summary, the AI‑driven compression of the execute layer is reshaping the anatomy of work, turning the classic hamburger into a model where decision making and delivery dominate. Organizations that recognize this shift and deliberately reinforce the top and bottom layers—through clearer conviction, robust verification, and supportive cultures—will be able to harness the speed of AI without falling into the traps of false productivity or unresolved technical debt. For employees, the opportunity lies in expanding their influence beyond manual execution: shaping what gets built, ensuring it is built right, and guiding its safe release. The future of work may not be a static pizza, but rather a dynamic pie where humans continually refine the toppings and perfect the bake, while AI supplies a reliable, ever‑improving crust. Embracing this mindset today will prepare teams and individuals for the evolving landscape of intelligent collaboration.