The day in one read

The Architecture of the AI Workforce

The dominant narrative of the day centers on a fundamental shift in how artificial intelligence is being deployed within the enterprise: it is moving from a conversational interface to a persistent, agentic workforce. This transition is not merely a change in user experience but a structural realignment of value, where the ability to build and direct software is becoming a baseline competency for all knowledge workers, not just engineers. The data suggests that the gap between early adopters and laggards is widening at an exponential rate, driven by non-technical teams leveraging AI coding tools to automate complex workflows. Simultaneously, the capital markets are responding to the physical constraints of this expansion, with venture capital pivoting aggressively toward the hardware and energy infrastructure required to sustain the compute demands of these new agentic systems.

The evidence for this shift is stark. OpenAI enterprise research indicates that agentic API token consumption surpassed non-agentic ChatGPT usage by early May, signaling a decisive pivot in how businesses interact with the technology. The disparity in adoption is no longer marginal; top 10% enterprise users now consume 8.3 times more AI than average firms, a gap that has widened dramatically from 2.6 times in January. This concentration of usage is driven largely by non-engineering functions. Since February, finance and accounting usage of OpenAI’s Codex tool increased 20 times, sales and accounting usage rose 41 times, and legal usage surged 108 times. In contrast, engineering teams saw a more modest 5x growth. This inversion of traditional adoption curves suggests that the primary value of AI is no longer in augmenting code, but in empowering domain experts to build their own software solutions, effectively democratizing the creation of digital tools.

From Chatbots to Persistent Co-Workers

Product leaders at OpenAI are framing this evolution as the third era of AI products, characterized by persistent co-workers rather than transient chat sessions or simple agents. Tara Sation, who leads product for Codex and ChatGPT Work, argues that the strategic error for product managers is to build for where models are currently capable or where they are predicted to be in a year. Instead, the optimal design window is two to three months out, a timeframe that allows for rapid empirical testing without being constrained by current model limitations or speculative future capabilities. Sation describes OpenAI’s internal culture as "founders-led," where top-down direction is limited and internal strategies quickly become public, contrasting sharply with the "secret strategy" model she expected from larger technology firms. This culture prioritizes prolific, empirical loops over theoretical documentation, defining the core role of the product manager as identifying the "eigenquestion" and testing it rapidly.

The architectural distinction between coding and knowledge work is central to this new product paradigm. Sation explains that ChatGPT Work mode is powered by the same underlying Codex engine but strips away coding-specific user interface elements to serve a broader audience. The long-term goal is a unified interface where users do not consciously choose between modes. However, Sation notes a critical difference in trust mechanics: knowledge work outputs cannot be easily verified via automated tests, as is possible in software development. Consequently, products must expose reasoning, citations, and in-progress work to build user trust. Features such as "Sites" and "Visualize" are highlighted as key innovations that allow users to generate dynamic, shareable artifacts and data visualizations directly from prompts, transforming static data into interactive experiences. Sation argues that these tools expand the "range of possibilities" for individuals, enabling them to act as "auteurs" rather than just executors. The most effective users, she notes, use AI to elevate their ambition, not just to automate rote tasks.

The New Competency: Building as a Foundational Skill

The strategic framework for non-engineers using AI coding tools emphasizes that building software is now a foundational capacity for knowledge workers to compound their advantages. A KPMG and University of Texas at Austin study of over 500 early-career professionals found that top performers, termed "AI amplifiers," derive value by guiding and refining AI outputs rather than relying solely on static knowledge. This approach categorizes build patterns into three distinct types: automation, where the job and output remain the same; upgrade, where the job remains the same but the output is new; and invention, where both the job and the output are new. Delivery classes range from disposable prototypes to personal software, production-grade tools for specific teams, and market-facing products.

Recommended use cases for this new competency include converting static PDFs into interactive HTML pages, automating data extraction pipelines, and building live dashboards to replace static reports. Tools like Lovable, Replit, Codex, and CloudCode are lowering the entry barriers for these tasks. The framework argues that the real argument for deploying AI coding as part of a knowledge worker’s toolkit is not that the user will become a software engineer, but that those who are building are compounding their gains and advantages relative to other AI users. Specific projects cited include automated export pipelines, invoice processing drop zones, and agentic monitoring systems for competitor pricing or regulatory changes. This shift implies that the competitive edge in the near future will belong to those who can translate domain expertise into executable code, using AI as the bridge between intent and implementation.

Capital Flows to the Physical Layer

While the software layer is being redefined by agentic workflows, the capital markets are responding to the physical constraints of this expansion. Andreessen Horowitz launched a $1.1 billion machine-age fund on August 28 to invest in the physical infrastructure underlying AI, including chips, networking, data centers, robotics, and energy. Managing partner Jen Kha stated that the fund targets everything below the software stack, a category that was largely uninvestable for 30 years until AI’s compute intensity necessitated a rebuild. The fund aims to maximize ownership at seed and Series A stages, contrasting with growth funds that typically enter at later inflection points. Kha noted that hardware now comprises over 20% of A16Z pitches, up from a tiny fraction, driven by a groundswell of entrepreneurs rebuilding inefficient legacy infrastructure.

The fund excludes regulated industries like power, focusing instead on computer-science-guided components such as custom silicon, memory, liquid cooling, and DC-powered systems. Kha highlighted that modern data centers, like portfolio company Switch, are built by tech experts to mitigate environmental concerns, contributing power back to the grid and using minimal water. She argued that less than 2% of U.S. electricians are trained for the DC power required by next-generation chips, indicating a significant skills gap in the physical infrastructure sector. The fund’s team includes Martin Casado, Raghu Raghuram, and Guido Appenizer, leveraging deep data center experience to diligence deals in a finite talent pool. Kha’s stance is that a dedicated fund is necessary to capture early-stage winners in AI physical infrastructure, as existing software-focused investment models bias toward later-stage, high-revenue deals and fail to address the unique capital and ownership requirements of hardware startups.

Global Adoption and Sovereign AI

The expansion of AI infrastructure is not confined to the United States; governments globally are treating AI as a national priority, with varying degrees of success and strategic focus. Kha observed that countries like Singapore and the UAE are accelerating adoption faster than the U.S., potentially shifting data center supply chain buildup overseas due to domestic political headwinds in the American market. South Korea has announced premium AI as a public utility, while El Salvador implemented Grok in schools. This global competition is influencing where physical infrastructure is built, with the risk that the U.S. may cede its lead in data center capacity to nations with more streamlined regulatory environments. The divergence in adoption rates suggests that the geopolitical landscape of AI is being shaped not just by technological capability, but by the speed and scale of physical deployment.

Also Notable

Tara Sation distinguishes between "writing as thinking," which she refuses to automate, and "writing as reporting," which she automates, arguing that human value persists in accountability, artistic expression, and interpersonal care. She identifies three internal cultural memes at OpenAI: "Is this maximally accelerated?", "Are you mainlining it yet?", and elevating others' ambitions. The A16Z fund’s approach contrasts with traditional venture capital by focusing on the "machine age" rather than the software layer, with Kha noting that "what's old is new again" in the context of hardware investment. The KPMG study highlights that "AI amplifiers" are those who guide and refine outputs, suggesting that the skill of prompting and directing AI is becoming as critical as the skill of coding. The surge in legal and finance usage of AI coding tools indicates that these sectors are finding value in automating document-heavy workflows, a trend that may reshape the junior talent pipeline in these industries.