Across 74 countries, the ten most common work activities people bring to ChatGPT made up 56.0% of work messages in early 2025. By the third quarter of 2026, that share had dropped to 51.1%.
That number comes from a new dataset published on the OpenAI blog. It looks like a minor statistical shift. I think it is one of the more useful signals for anyone responsible for how their company adopts AI.
What OpenAI released
The dataset maps ChatGPT work usage to 165 work activity categories across nearly 100 countries, quarterly, from the first quarter of 2025 through the third quarter of 2026. The categories come from O*NET, the standard framework that groups occupation tasks into broader activities, so researchers can compare AI use with employment data.
The activities range from preparing instructional materials and developing marketing content to programming and analysing business data. Countries also specialised over time, with different markets concentrating on different kinds of work.
Two limits matter. The data covers work-related messages on individual ChatGPT plans, and it excludes enterprise accounts and Codex. And about 43% of the employment-weighted task mix in U.S. occupations falls into categories like moving materials or operating vehicles, which barely appear in ChatGPT use at all.
The top ten is no longer the strategy
Most corporate AI programmes I have seen start the same way. A workshop, a list of high-value use cases, a pilot for each. Summarise meetings, draft emails, write first drafts of content, analyse a spreadsheet.
That list made sense when usage was concentrated. The OpenAI data says the concentration is fading. People are taking ChatGPT into narrower, more specific parts of their jobs, the ones no central committee would think to put on a slide.
In marketing teams, I see this constantly. The approved use case is content drafts. The real use is a planner restructuring a media brief, an analyst cleaning a messy export, a coordinator rewriting a supplier contract summary. None of these were in the plan, and together they add up to more value than the official pilots.
The quiet problem in the fine print
Here is the part I would underline for any leadership team. This data is about work use on personal plans. Enterprise accounts are excluded.
So the long tail of work activity OpenAI is describing is, by definition, happening largely outside company-managed accounts. Employees are bringing real work tasks to personal subscriptions. Some of those tasks involve client data, financials or internal plans.
I do not think the answer is to ban it. Bans push the behaviour further out of sight. But it does mean the actual AI footprint of most companies is wider than their official programme, and the gap grows every quarter as usage spreads into more specialised tasks.
I wrote earlier about how people say they distrust AI but use it daily. This dataset is the work version of that gap. The official story and the operational reality are diverging.
What the long tail means for value
There is a second-order effect that cuts the other way. Generic tasks are where AI gains are easiest to see and easiest to copy. Every competitor can summarise meetings with the same tools.
Specialised tasks are where the gains are harder to see and much harder to copy. When a senior analyst builds a personal workflow for a specific forecasting problem, that is know-how. It compounds inside the team and does not show up in a vendor case study.
The risk is that this value lives in individual habits on personal accounts. When that person leaves, the workflow leaves too. Companies that capture it, document it and move it into managed systems keep the advantage. Companies that ignore it lose it one resignation at a time.
The other risk is quality. Specialised work is also where AI errors are hardest to catch. I covered the hidden rework cost of AI marketing efficiency, and that cost is higher when the reviewer is the only person who knows the task.
What I would do with this data
First, run an honest inventory. Ask teams, without judgement, which work tasks they bring to AI and on which accounts. Anonymous surveys work better than audits for this. You want the long tail, not the approved list.
Second, move the valuable workflows into managed accounts. If a task involves client or financial data, it belongs on an enterprise plan with proper data controls. Make that move easy, or people will not make it.
Third, turn the best individual workflows into shared assets. A documented prompt, a template, a small internal tool. That is how a personal trick becomes a team capability.
Finally, update the AI strategy every quarter based on what people actually do, not what the original workshop predicted. The OpenAI data shows usage patterns shifting measurably within eighteen months. A strategy written once is out of date before the second budget cycle.
Your employees already decided where AI is useful. The strategy just has to catch up with them.