We’ve spent the last three years hearing about corporate AI adoption.
Every company is AI-first. Every CEO has an AI strategy. Every employee has Copilot, Claude or ChatGPT. There are committees. There are transformation programs. There are PowerPoints containing diagrams of glowing brains.
Wonderful.
The problem is that buying 5,000 ChatGPT licenses tells us approximately nothing about whether anyone is actually using them.
So I was very happy to find a new working paper from researchers at OpenAI, Columbia Business School and Wharton that does something considerably more useful: it looks at what people actually do.
The researchers linked privacy-preserving ChatGPT Enterprise usage data with employee roles, automated task classifications and, for U.S. public companies, financial data from Compustat.
The scale is unusually good: the worker-characteristics sample covers 1,764 organizations and 17.4 million messages, while the task-classification subset contains 973 organizations and 8.7 million classified messages across 60 task categories. ArXivSignals
For once, we’re not asking executives whether AI is important to their strategy. We’re looking at the receipts. Read the OpenAI research overview
Enterprise AI Usage Is Growing Ridiculously Fast
The first thing that jumps out is simply how quickly usage is expanding. Enterprise token consumption increased roughly 7× between June 2025 and March 2026. But here’s the more interesting bit.
If you only look at companies that were already customers by June 2025, usage still increased roughly 4×.
So this isn’t merely OpenAI signing more Enterprise customers. Existing customers are using substantially more AI. And early 2026 appears to contain another acceleration across adoption cohorts at roughly the same time.
That suggests something changed across the ecosystem rather than companies simply becoming more comfortable with ChatGPT as they aged.
Which feels about right. The models and surrounding products became dramatically more capable around that period. Once AI stops being something you ask questions and starts becoming something to which you can delegate work, token consumption has a tendency to become... enthusiastic.
OpenAI’s newer enterprise data already shows the same direction becoming even stronger: agentic usage is expanding beyond engineering, and its most intensive enterprise customers now generate 8.3× more output tokens per active user than typical firms. OpenAI
We’re moving from: “Ask AI something.” toward: “Give AI something to do.”
Those are very different consumption curves.
The Companies Adopting First Aren’t Random
Here’s where the paper becomes more interesting economically. Enterprise adopters are dramatically larger than non-adopters.
Among the public companies in the analysis, adopters have much higher revenue, market value, employee counts, assets and R&D expenditure.
This isn’t simply: big companies have more money, therefore they buy ChatGPT.
The researchers find more structure than that.
Companies in the top quartile by revenue are about 6.9 percentage points more likely to adopt, while those in the top 5% are about 9.8 points more likely. Looking relative to firms inside the same industry makes the relationship even stronger, reaching roughly 11 percentage points. Revenue per employee is also positively associated with adoption. Capital intensity, measured through PP&E per employee, goes the other way.
In wonderfully simplified terms:
companies whose economics depend heavily on knowledge workers are adopting faster than companies whose economics depend heavily on giant physical things.
Not exactly shocking. A steel mill cannot prompt-engineer its blast furnace. At least not yet.
AI May Reward Companies That Were Already Good Companies
This is probably my favorite part of the paper. The companies adopting early tend to be the ones that were already investing heavily in R&D, software, processes and organizational capabilities.
This fits an old pattern with general-purpose technologies. Buying the technology isn’t enough. Electricity didn’t magically make every factory equally productive. Computers didn’t make every corporation equally productive. The internet didn’t make every retailer Amazon.
New general-purpose technologies tend to produce the greatest gains when organizations make complementary investments around them.
Processes change. Skills change. Information architecture changes. Management changes. Eventually the company itself changes. AI looks increasingly similar.
Which is why I remain deeply skeptical whenever I hear:
“We’ve rolled out Claude to all 2,000 employees. We’re AI-first now.”
No. You have purchased 2,000 subscriptions. Congratulations.
Big Companies Buy More AI. That Doesn’t Mean Employees Use More AI.
There’s another lovely contradiction in the data. Large companies are more likely to adopt ChatGPT Enterprise. But usage intensity per employee tends to be lower in larger organizations.
This makes perfect sense. Signing an enterprise contract is centralized. Changing how thousands of humans work is not. Procurement can buy AI for everyone on Tuesday.
By Wednesday, perhaps 50 enthusiasts are doing extraordinary things with it. Another 400 occasionally summarize documents. And somebody in accounting is still printing emails.
Technology diffusion inside organizations is messy. Which is why I think AI access is becoming an increasingly useless metric.
The interesting metrics are things like: usage intensity, workflow penetration, task delegation, agent autonomy, measured output quality, and eventually business outcomes.
A license is not transformation. It’s permission to begin one.
Who Actually Uses It?
ChatGPT usage isn’t confined to engineering. In the average organization studied, software engineers account for about 11% of weekly active users. Executives and partners are around 9%. Finance and marketing are each roughly 5%. But headcount share and usage intensity aren’t the same thing.
Analysts and marketing roles tend to generate particularly high message volumes. Executives and finance employees send fewer. And then we get one of my favorite findings: junior employees use AI more.
Early-career workers send substantially more messages per week than senior employees in the same organizations. OpenAI’s newer enterprise analysis finds the same gradient: six months after adoption, early-career employees were sending around 13 more messages per week than executives. OpenAI
There are at least two ways to interpret this.
The optimistic one: young employees adapt faster to a new technology.
The less optimistic one: young employees have discovered an extraordinarily convenient way to outsource everything they haven’t learned yet.
I suspect both are true. And this creates an interesting management problem.
AI can dramatically accelerate learning if you use it to interrogate a problem. It can also allow you to skip learning entirely if you use it to make the problem disappear. Those behaviors can look identical in usage statistics. Until something breaks.
What Are Employees Actually Doing?
Now we get to the part I was most curious about. What do millions of enterprise ChatGPT conversations actually contain? The answer is surprisingly mundane.
Which is precisely why it’s important.
Among weekly active users:
56.3% use ChatGPT for documentation and technical writing.
49.6% for technical, digital and electronic work.
41.1% for interpersonal messages.
38.9% for topic overviews.
Then come facts and figures, professional or academic work, business research, sales and marketing, planning, legal work, data analysis and finance.
This belongs directly after the percentages above. It’s the cleanest visual explanation of what enterprise ChatGPT actually does.
There isn’t one magical killer application. There is a broad layer of cognitive assistance spreading across ordinary knowledge work.
Writing. Technical work. Communication. Research. Planning. Analysis. Exactly the sort of work that fills most people’s days without appearing anywhere in the company’s official process diagrams.
By message volume, activity is somewhat more concentrated: documentation accounts for around 18.3%, technical work 12.7%, and interpersonal communication 11.9%. But there’s also a substantial long tail: roughly 27.3% of messages sit outside the twelve largest task categories. That long tail matters.
General-purpose technologies become powerful precisely because people discover uses nobody designed centrally.
Industries Are More Similar Than You Might Expect
There are obviously differences between industries. Finance produces more finance and tax work. Retail and media lean more heavily toward sales and marketing. Fine. But the differences aren’t enormous.
Across industries, the core remains remarkably similar:
writing + technical work + communication + information synthesis.
That’s interesting because we often discuss enterprise AI as though every industry requires an entirely bespoke AI revolution. At the model level, perhaps not.
A surprisingly large amount of corporate cognition consists of manipulating information.
The specialization increasingly appears around the model: company data, tools, permissions, workflows, domain rules, memory, verification, and organizational context.
The intelligence layer can be increasingly generic. The harness cannot.
Executives Use AI Differently
Senior leaders also show a distinct usage profile. They disproportionately use ChatGPT for topic overviews, legal questions and financial work, while doing less of the production-oriented activity found further down the hierarchy. Again, unsurprising. But useful.
AI doesn’t simply distribute one new capability uniformly throughout an organization. It amplifies different parts of different jobs. A junior analyst may use it to produce. A marketer may use it to generate and research. An engineer may use it to build. An executive may use it to compress information and explore decisions.
Same model. Different leverage.
Now Look at Productivity
This brings us to the most economically provocative figure. The paper compares revenue per employee and market value per employee among non-adopters, low-intensity adopters and high-token-intensity adopters.
The distributions shift. High-intensity adopters are disproportionately found toward the higher-productivity end.
And here we need to be careful.
This does not prove:
ChatGPT makes companies more productive.
The study is observational. More productive, innovative and valuable firms may simply adopt AI earlier and use it more intensely.
In fact, that’s probably part of what we’re seeing. The paper itself documents associations, not a clean causal estimate of AI’s effect on revenue or market value. ArXivSignals But economically, this may actually be the more interesting problem.
What If AI Makes the Productivity Gap Bigger?
There’s a comforting story about general-purpose AI: Everyone gets access to approximately the same intelligence.
Therefore technology democratizes capability. Small companies can suddenly do what giant corporations do. And there is certainly some truth to that.
A ten-person startup now has access to research, coding and analytical capabilities that would have required considerably more people a few years ago. But there is another possibility.
What if the organizations best positioned to exploit AI are precisely the ones that already have: better processes, better data, better employees, better internal software, better management, more R&D, and more capital to redesign themselves?
Then equal access to models doesn’t equal equal outcomes. It does the opposite. The best organizations integrate the technology faster. They discover more use cases. Their employees develop better habits. They build internal tooling. They connect models to proprietary context. They create reusable workflows. They measure results. They automate more. Then they use the resulting productivity gains to invest even more.
Now you have a compounding loop. OpenAI’s current enterprise data already hints at precisely this divergence: its frontier firms have moved from 2.6× the output tokens per active user of typical firms in January to 8.3× by June. OpenAI
Same technological revolution. Very different trajectories.
We Are Still at the Splashing-Around Stage
My biggest takeaway from the paper is actually reassuringly simple.
Nobody has completely figured this out yet.
Companies are still discovering where AI belongs. Usage is spreading across functions. Junior employees use it differently from executives. Large companies adopt faster but struggle to diffuse intensive usage. Industries have their own specializations but share a surprisingly large common core. And existing customers keep dramatically increasing consumption as capabilities improve.
This looks exactly like an early general-purpose technology. Lots of experimentation. Lots of mediocre implementations. A few extraordinary ones. And enormous differences in organizational learning speed.
Which is why purchasing licenses is the least interesting part. The actual work begins afterward. Processes have to be redesigned. Employees need to learn what to delegate and what not to delegate. Successful individual workflows need to become organizational workflows. Models need access to company context and tools. Permissions and verification need to exist. And eventually companies need to measure whether any of this is producing better outcomes rather than merely producing more tokens.
That’s considerably harder than adding “AI-first” to the About page.
The Small Companies Should Probably Run Faster
There’s an obvious optimistic interpretation of AI for startups: intelligence is getting cheaper. Wonderful. But large companies aren’t asleep.
The organizations adopting earliest are already larger, richer, more R&D-intensive and often more productive. And they’re learning too.
So the advantage of being small isn’t that giant companies won’t use AI. They absolutely will. The advantage is that a small company can redesign itself faster. No twelve-month procurement cycle. No 40-person transformation committee. No legacy department whose primary objective is ensuring that the legacy department continues to exist. That speed may matter enormously. Because if this research is pointing in the right direction, AI isn’t automatically leveling the corporate playing field. It may actually make organizational quality more valuable.
Everyone can buy access to intelligence. Not everyone knows what to do with it. And that difference may become one of the most important competitive advantages of the next decade.