43 Percent of AI Use Is Somebody Else’s Job

OpenAI read more than 800,000 work messages from American ChatGPT users at their desks. Strip out the generic requests, and 43.5 percent of what people ask AI to do belongs to a job that is not theirs.

The marketer is writing SQL. The designer is redlining a contract clause. The HR generalist is building the dashboard she used to file a ticket for. None of them got a promotion or a new title out of it. They stopped waiting on the person who owns that work, and 800,000 messages caught them at it.

Who crosses the line most

Some roles now spend most of the day outside their own lane.

The traffic also runs harder one direction than the other. Engineering tasks travel farthest: engineering work shows up in 7.4 percent of messages written by people who are not engineers. Design travels least, 1.7 percent.

Engineering has a right answer you can test. Your query returns rows or it throws an error, and you know which within four seconds. Design hands you no receipt like that, so people borrow it less. Work crosses job lines when it can prove itself to a beginner.

Why the five person company moves faster

Small workspaces cross over at 18.9 percent. Companies past 100 seats sit at 16.3 percent. That is not a rounding error, and the cause is not culture.

In a five person company, nobody is coming. You need the landing page copy, there is no copy team, so you write it. In a five hundred person company there is a team, a queue, a form, and a person whose performance review depends on owning that form. Same tool sits on both desks. What differs is who gets to touch the work.

The part the report does not say out loud

Task crossover is not the same thing as competence. A marketer who ships a query she cannot read has not become an analyst. She has become a person holding a result she cannot evaluate.

That is the risk buried in this data, and it outranks the usual complaint about accuracy. Hallucination is a solvable annoyance once you know to check. Confident output in a domain where you have no taste is different, because you do not know what checking would even look like.

You are already crossing. All that is left to settle is how to do it without pretending you have the training.

Which lane to cross into first

The 7.4 versus 1.7 split hands you a selection rule, and you will get more out of it than out of the headline number.

Engineering travels because engineering fails loudly. Bad SQL returns an error or an obviously wrong row count. Bad design returns something that looks fine to you and makes a customer close the tab six weeks later without telling you why. One of those teaches you. The other lets you stay wrong for a year.

So rank the tasks you are tempted to take on by how fast they tell you that you blew it. A spreadsheet formula tells you in seconds. A contract clause tells you in a year, in court. A pricing page tells you in a quarter, and only if somebody is measuring. Cross into the fast feedback work first, and treat the slow feedback work as something you learn to brief rather than something you learn to do.

It also explains the top of the list. Customer experience at 77 percent, HR at 69 percent: those people sit where every other department’s work lands as a problem due today. They were crossing lines before any of this. AI removed the last excuse, which was tooling.

The Crossover Ladder

Four rungs, each mapped to a SeedStacking phase. All four together cost less time than the meeting you would have booked instead.

Seed: name the border, 10 minutes

Write down one task you keep handing to somebody else. Be specific enough that it has a deliverable. Not “I wish I understood data.” Try “I hand off every request for a cohort retention number.”

Sprout: borrow the vocabulary, 20 minutes

Do not ask AI to do the task yet. Ask it what a competent person in that role would ask before starting. You are after the questions, not the answer. A good analyst asks which cohort, over what window, counting what as active. If you cannot answer those, you were never ready to receive the number anyway.

Grow: build a throwaway, 30 minutes

Produce the thing badly, on purpose, knowing you will delete it. Then hand your own output back and ask what a senior person in that field would flag. The gap between your version and that critique is your actual skill gap, measured rather than guessed.

Harvest: show it to the owner, 20 minutes

Take it to the person whose job it really is and ask what you got wrong. This is the rung people skip, and skipping it is what turns crossover into damage. It also does something the report cannot measure: the analyst now knows you respect the craft, which is how you get help the next time instead of a queue position.

Run all four and you have spent about 80 minutes. You will not be an analyst. You will be someone who can hold an analyst’s output and tell whether it is any good, which is what the 43.5 percent asks of you.

What to do this week

Pick the single task you hand off most often and run it up the ladder once. One task, one week. Not a transformation program, not a tool rollout.

And if you manage people, read the 18.9 versus 16.3 number again. Your team is crossing over less than a five person startup, and it is not because they are less capable. It is because somewhere in your process there is a form standing between a person and work they could already do.