A new OpenAI working paper on ChatGPT Enterprise is useful because it moves the enterprise AI debate away from demo theology and toward observed behavior. Aaron Chatterji, David Holtz, Neel Rakholia, Prasanna Tambe, and Gawesha Weeratunga link ChatGPT Enterprise account records to usage, worker roles, task classifications, and public-company financial data through March 2026. At the six-month adoption horizon, their worker-level sample covers more than 1,500 organizations and more than 17 million messages.
That is still a slice of one vendor's enterprise product, with all the selection issues that implies. It is not a census of AI adoption. But it is a much better object than another survey where executives mark "AI transformation" on a five-point scale and then go back to approving pilots.
The headline finding is not just that usage is growing. Aggregate output tokens from ChatGPT Enterprise customers grew roughly sevenfold between June 2025 and March 2026, and almost fourfold inside a consistent cohort of firms that adopted between January 2024 and June 2025. The authors' more important result is about who adopts early. Among U.S. public companies, adoption is concentrated in larger, more valuable, more R&D-intensive, and more SG&A-intensive firms.
That pattern should make the equalization story a little more conditional.
The steelman for democratization is strong. A general-purpose model is cheap to try. A small firm can rent capabilities that used to require a larger staff, a better software budget, or a consultant. Workers without deep technical training can use the same interface as a programmer. The frontier model arrives through a browser instead of a multi-year IT project. That matters. AI really does lower the first cost of experimentation.
Experimentation is not the same thing as capture.
A firm captures value from AI when it can move the tool into actual work without breaking compliance, confusing ownership, leaking data, or adding review costs that eat the time saved. That takes clean internal processes, data access, managerial tolerance for workflow change, and enough slack to let workers discover where the tool is useful. Those are not evenly distributed assets. They are close cousins of the intangible capital economists have been talking about since the computer age.
This is why the OpenAI paper reads less like a revolution against organizational advantage and more like another chapter in the productivity J-curve. Erik Brynjolfsson, Daniel Rock, and Chad Syverson argued that general-purpose technologies often require complementary investment before the gains show up in measured productivity. The early spend can look disappointing because the asset is not only the machine. It is the redesign around the machine.
ChatGPT Enterprise has an unusually low on-ramp, but the same logic is visible. The paper reports use across job functions and seniority levels, with especially high usage intensity among early-career workers. Marketing and communications workers send more messages than executives. The common tasks include writing, communication, information synthesis, research, planning, data analysis, legal and regulatory work, and finance. That breadth is the point. This is not one workflow that gets automated cleanly. It is a layer of small assists scattered across the knowledge-work stack.
Small assists can be valuable. They are also hard to monetize cleanly inside the firm. If an analyst writes a memo faster, a lawyer drafts a first-pass clause, or a marketer iterates on copy, the P&L does not immediately print an AI line item. The gain arrives as throughput, quality, fewer bottlenecks, or a decision made with slightly more evidence. Those gains are real, but they need managerial conversion. Someone has to change workload, staffing, review standards, cycle time, or customer response. Otherwise the saved minutes become nicer days and cleaner drafts rather than economic surplus.
That is not a complaint. Better work and less drudgery are good outcomes. The investment question is narrower. Who gets paid?
The vendor gets paid first because the subscription is explicit. The worker may get paid indirectly if the tool raises output enough to protect bargaining power or move them into higher-value tasks. The firm gets paid if it converts dispersed productivity into margin, growth, or faster learning. Customers get paid if competition forces the firm to pass along lower cost or better service. The distribution is open. The mere fact of enterprise usage does not settle it.
The early-career usage result is especially interesting. One reading is labor substitution. Junior workers do much of the drafting, searching, cleaning, summarizing, and first-pass analysis that generative AI can touch. If the tool makes those tasks faster, firms need fewer entry-level workers for the same output. That is the bleak read.
Another reading is worker reach. Junior workers are closer to the messy work and have fewer established routines to defend. They may adopt faster because the tool raises their effective scope. A first-year analyst who can summarize filings, test a spreadsheet, draft client language, and ask better questions becomes useful in more places. The scarce complement then shifts from raw drafting speed to judgment, supervision, and domain context.
Both states can coexist. The same tool can compress the number of routine junior seats while making the remaining seats more productive. The distributional question is whether the firm treats AI as a way to cut the bottom rung or as a way to move workers up the learning curve faster. Incentives will differ by industry, labor market, regulation, and error cost. A bank does not face the same loss function as a media team or a software company.
The adoption concentration among larger and more intangible-heavy firms gives a clue about the first-round winners. Large firms have more bureaucracy, so it is tempting to assume they should be worse at AI. In practice, they also have more repeated workflows, more data, larger internal markets for a successful use case, and more budget for governance. A small improvement in contract review, customer support, finance analysis, or sales enablement can be worth a lot when it repeats across thousands of employees.
That scale cuts both ways. A large firm can waste more money on theater. It can also amortize the cost of finding one good workflow. If the first ten pilots are mediocre and the eleventh changes a recurring process, the large firm has a portfolio. The small firm often has a few shots and less capacity to formalize what worked.
This matters for the macro story. A world where AI tools are cheap and accessible does not automatically produce broad productivity convergence. It may produce a wedge. The interface democratizes access, while the complements determine capture. If larger, more productive firms adopt earlier and integrate better, AI can widen the gap between frontier firms and the median firm even while almost everyone can open a chat window.
That would not be unusual. Computers were eventually everywhere, but the returns did not arrive evenly. The firms that reorganized around software did better than firms that merely bought software. AI may have a faster diffusion curve because the user interface is simpler. I would still expect the returns to be sorted by organizational quality.
The useful test over the next few years is not whether employees keep sending more messages. They probably will. The test is whether message growth maps into measurable changes in revenue per employee, cycle time, error rates, customer retention, or product velocity. If those links appear mostly in firms that already had strong data and management systems, the economic story will look less like mass automation and more like capital deepening for the firms best prepared to absorb it.
There is a policy version of this too. Subsidizing adoption alone may be weak medicine. If smaller firms lack clean data, training, managerial capacity, or trust around deployment, a cheap AI subscription does not close the gap. It rents a piece of frontier capability from an upstream provider and leaves the harder complement problem untouched.
My prior is that enterprise AI will be genuinely useful and still less egalitarian than the interface suggests. The browser makes the tool feel flat. The payoff is not flat. The rents go first to the organizations that know what work is worth changing.
Sources: Chatterji, Holtz, Rakholia, Tambe, and Weeratunga, "How Organizations Use AI: Evidence from ChatGPT"; OpenAI business usage guide; Brynjolfsson, Rock, and Syverson on the productivity J-curve; BIS and CEPR evidence on AI adoption, productivity, and firm complements.
Originally published at deanlee.info.
Top comments (0)