For most of the past four years, the question about AI at work was what the models could do. That question is slowly giving way to a quieter and more consequential one: whether organizations know how to hand work over. The technology has moved from answering questions to taking actions, from conversation to delegation, and the evidence so far suggests that the pace of that shift depends far less on the tools themselves than on the institutions that use them.

The clearest illustration comes from OpenAI's own usage data. The paper The Shift to Agentic AI: Evidence from Codex tracks how AI use moved from conversational products to agentic ones across consumers, enterprise customers, and OpenAI itself. Inside OpenAI, in roughly two months, almost all AI use became agentic in every department, from engineering to legal and recruiting. At other companies, engineering and data science led a similarly fast transition, while in legal and recruiting roles only around 2 percent of AI use was agentic as of June 2026. The models were the same and so was the product, yet the pattern of work could hardly be more different, which is why the gap is better understood as an organizational one than as a question of capability.

From conversation to delegation

The distinction between conversational and agentic AI is less tidy than the labels suggest, since even chat products now search the web and use tools on our behalf. What matters is the change in the unit of work. Let’s break this down a bit. In a conversation, a person asks and reads; in delegation, a person specifies, waits, reviews, and specifies again, so the worker gradually stops being the one who performs the task and becomes the one who assigns and checks it.

Outside OpenAI, a typical user tops out at two or three, whereas inside OpenAI many workers at some point have eight, nine, or ten agents working in parallel, one drafting emails, another writing code, a third building a presentation. Someone in what used to be an individual-contributor role is, in practice, managing a small team. Running ten agents at once is not the same job done faster but a different job altogether, built on assigning and reviewing rather than performing, and most organizations have not yet written its description.

The factory floor, again

Economic historians have a familiar story for moments like this. When factories were first electrified, the new power source did surprisingly little for productivity, because plants were still laid out around the constraints of steam, with machines clustered around a central shaft and work sequenced by where power could reach. The real gains came only when managers realized they could rearrange the floor entirely, and that realization took the better part of a generation.

Knowledge work is still organized around the constraints of a single human attention span: we write the email, then the code, then go to the meeting, then finish the slides. Agents loosen that constraint much as electricity loosened the shaft, but the floor plan has barely changed. There is a further point that deserves more attention than it gets. OpenAI itself was built before agents existed and carries some of that path dependence, so firms designed from day zero around delegation may end up looking different from anything visible in today's data, including the most AI-forward company in it.

One early sign of the redesign is the spread of skills, the short files that give an agent a codified and repeatable way of doing a task. Mechanically they are just markdown, but economically they are something more interesting. Every organization runs on tacit knowledge, from how the data is structured to how a table should be formatted to which mistakes the senior people always catch, and that knowledge has traditionally lived in people and travelled through onboarding and apprenticeship. Writing it down as a skill turns it into organizational capital that can be reused, audited, and scaled across as many agents as the firm is able to supervise. The firms building that library now are making an investment that will not show up in any productivity statistic for some time, which is precisely what intangible investment tends to look like in its early years.

The capability overhang

OpenAI's chief economist, Ronnie Chatterji, has a name for the distance between what AI can do and what most people actually use it for: the capability overhang. Frontier users exploit a large share of what these systems offer, while the median user barely scratches the surface and often confines even agentic tools to a narrow set of writing tasks. If the transformation of work depends on frontier capabilities becoming ordinary, then the overhang is the variable that matters most, and nobody yet knows how to close it. Training helps in some settings and structured exploration in others, but there is no single lever.

Early results from micro1's survey of experts across its global community complicate the picture in a useful way. Heavy users do behave differently: they delegate tasks end to end, report more confidence, and describe agents enabling work they would not previously have attempted on their own. The population, however, does not split neatly into two camps. Usage is spread along a continuum, with many people delegating 20 or 30 percent of their work and roughly one in ten relying on agents for 80 percent or more. Two findings are worth considering. More than half of the heavy users work outside technical fields, which suggests the overhang is not simply occupational, and for respondents outside the United States cost appears to be a real constraint, which suggests it is not simply a matter of willingness either.

Read together, the two sources point to a more precise diagnosis, worth considering.  The gap between frontier and median users is shaped by access, by cost, and by organizational permission to experiment at least as much as by the nature of the work, and all three are things institutions can change or at least build the necessary clarity and ambition to do so.

Measuring productivity is the hard part

Knowledge-work productivity has always been difficult to observe. Software engineering is the partial exception, which is why so much early research concentrates there, since commits, pull requests, and closed tickets are imperfect proxies but proxies nonetheless. For a lawyer, a recruiter, or a marketer there is often nothing comparable. Counting slide decks tells us very little worth knowing, and hours worked say even less about what an afternoon of delegation actually produced.

Two further problems make this harder. The first is that individual gains do not automatically aggregate: a worker who becomes five times faster contributes little if the final deliverable still waits on a colleague whose part of the process has not changed, so the relevant unit is the workflow rather than the worker, and bottlenecks tend to migrate. The second is that the data needed to see the whole picture sits in different places. AI labs observe usage in great detail but rarely see outcomes inside the workplace, while firms see outcomes without the texture of how the tools were used. Neither can answer the productivity question alone, and the credible path runs through collaboration between labs, firms, and academic researchers, combining datasets that no single party holds.

Supervision belongs in that accounting as well, because delegation is not free. Agent output has to be reviewed, corrected, or redone, and any estimate that ignores that time will overstate the gains precisely in the roles where review is hardest. Measuring net delegation, the work handed over minus the work taken back, is one of the more tractable contributions economists can make right now.

Measuring the transition while it is happening

None of this settles the question that hangs over every conversation about AI and labor, which is displacement. The historical record is reasonably clear that general purpose technologies create new kinds of work that nobody anticipated. It is equally clear that the path from the old equilibrium to the new one is rarely smooth, and that the costs of the transition fall on specific people in specific places. The role of the economist here is neither optimism nor pessimism, but producing evidence, putting it in front of workers, firms, and policymakers, and helping them reason about it quickly enough for it to matter.

There are good grounds for thinking this transition can be managed better than earlier ones. Governments now hold fiscal and monetary tools that were not available during previous technological shifts, and they can use them to cushion short-run disruption while new work takes shape. The measurement environment has also changed beyond recognition. When personal computers entered the workplace, the best available data told us roughly how many employees used one; today, aggregate and privacy-preserving workplace telemetry can show which tasks people work on and how their time is being reallocated, and AI agents themselves can run adaptive, conversational surveys at a scale and depth that fixed questionnaires never reached. A policy apparatus that reads those signals in close to real time could respond to labor-market shifts in months rather than years.

The harder task is to reason ahead of the data, which is what AGI economics sets out to do by applying the tools of the discipline to questions that sound speculative but are quickly becoming practical. How fast will capabilities improve, and what would we need to measure to detect AI systems accelerating their own research? How should teams of agents be organized, and does hierarchy serve them better than a flat structure? And what happens to markets built on the assumption that searching, booking, and negotiating take effort, once personal agents make that effort nearly free? These are questions about the near future, and economics is far better placed to answer them now than after the fact.

A zero-principle problem

In the 1930s, economists built national accounts because the existing statistics could not describe the economy that had just collapsed in front of them, and GDP was in that sense an answer to a measurement crisis. The agentic transition is creating another one, because the categories we use to define a job, a unit of output, and even a working day were designed for a world in which output could be traced to human work hours.

That is why the moment calls for more than first-principles reasoning about familiar quantities. First-principles reasoning starts from what we already count (jobs, hours, output) and works forward, whereas a zero principle starts one step earlier, by deciding what should be counted at all.

The capability is already here. The overhang will close at the speed at which organizations redesign themselves around delegation, and at which our measures learn to see what delegation produces.

This piece draws on a conversation hosted by the micro1 Forum, featuring David Holtz, Head of AGI Economics Research at OpenAI, and Liu Zhang, Economist and Member of Technical Staff at micro1. The paper The Shift to Agentic AI: Evidence from Codex is available here.