Half your AI usage is documentation.

What people are actually doing.

When the researchers classified those 17.4 million messages, the picture was less futuristic than the marketing suggests. Over half of active users were doing documentation and technical writing. Nearly half were doing technical digital work. Then a long tail: communication, research summaries, sales and marketing, legal, finance, planning.

In other words, the dominant enterprise use of AI is high-volume, routine, well-specified production work. Write the doc. Summarize the thing. Draft the update. It is the least exotic work in the building, and it is where the tokens go.

And it’s growing fast.

Aggregate output tokens grew sevenfold in nine months. More striking: among firms that had already adopted at the start of the window, usage still deepened roughly fourfold. That’s not new customers arriving. That’s the same teams using it more.

We flagged this pattern before as Jevons — efficiency doesn’t shrink the bill, it raises usage. This is that claim, measured. Nobody is going to optimize their way to a flat AI invoice.

Put the two facts together.

Your biggest AI workload is routine production writing. Your usage of it is compounding several times a year. And unless somebody made a deliberate decision, all of it is running on whatever model happened to be the default in the box.

This is what makes routing more than a clever trick. Route by task, and the task you’re routing is the one your team runs all day, every day, at compounding volume. Our model guide puts routine drafting and extraction squarely in the fast-and-cheap tier. The gap between that tier and a frontier default, multiplied across half your messages, is the bill.

The honest catch (three of them).

• Messages aren’t value. The authors are explicit that they measure classified message content, not “downstream work products, productivity effects, or changes in organizational routines.” Half your messages being documentation doesn’t mean half your value is. Volume tells you where to look, not what to cut.
• Not all documentation is routine. A throwaway status update and a customer-facing technical spec are the same category to a classifier and very different jobs to a business. Downgrade the first freely; test before you touch the second.
• It’s one vendor’s window. ChatGPT Enterprise only — not your API calls, internal tools, or personal accounts. Directionally useful, not a census.

One more finding worth your attention.

The researchers also found that adoption is only the beginning — getting access to the tool predicted very little on its own. What separated firms was the complementary work around it: the workflows, the routines, the co-invention. Usage was wildly uneven too, with early-career workers sending far more messages than average.

That’s the same thing we argued in the floor and the ceiling: buying the technology is not the move. Redesigning the work around it is. Now there’s a dataset behind it.