The Handoff Was the Control: OpenAI Measured Role Dissolution and Called It Expansion

One sentence in OpenAI’s July 2026 usage analysis deserves the attention of everyone who signs off on regulated work: “some activities that once required a handoff can now be done by the person who first encounters the need.”

OpenAI files that under expansion. In any function carrying a compliance obligation, the handoff it describes was the control.

What was measured

The study covers more than 800,000 messages from US ChatGPT users. OpenAI’s headline finding: “16.8% of work-related messages and 43.5% of occupation-specific messages are about tasks associated with another occupation.”

Those are two different denominators and they get conflated constantly. The 16.8% is the share of all work-related messages. The 43.5% is the share of a narrower slice: messages that carry an occupational signature at all. According to the chart accompanying the piece, 61.5% of work messages are classified as generic (writing, summarizing, scheduling, activities shared too broadly to indicate anything) and 38.5% are non-generic. Within that non-generic portion, 56.5% sit inside or near the user’s own occupation and 43.5% sit outside it.

So when someone is doing work that identifiably belongs to a profession, it is close to a coin flip whether that profession is their own.

The per-function breakdown is where it stops being an abstraction. Share of occupation-specific messages falling outside the user’s own occupation: customer experience 77%, designers 75%, HR 69%, legal 56%, marketers 53%.

A handoff is not paperwork

Review, sign-off and segregation of duties survive in most organizations as documents: a policy PDF, a checklist, a workflow step in a ticketing tool. The document is the residue. The control is the physical fact that work stopped, changed hands, and continued only after a second qualified person touched it.

That moment does four things at once. It puts the work in front of someone who holds the relevant qualification. It creates a second read by a party who did not author the thing. It produces a record with a name and a timestamp, which is what an auditor actually samples. And it separates the person who wants an outcome from the person authorized to approve it.

Delete the crossing and all four go with it. Not one of them survives in the document, because the document only ever described the crossing.

This is the same failure shape we traced when permissions stopped being a system of record: the artifact that proves a control existed disappears before anyone notices the control did.

There is a sanctioned version of this, and we wrote it up in June: 37% of UX organizations cut staff and one polymath plus AI absorbed an entire pipeline. That kind of checkpoint loss is auditable. Someone decided to collapse the roles, the org chart records the decision, and you can name every review step that went away with it. The OpenAI data describes the unsanctioned version. No decision was made, no role was consolidated, no reporting line moved. People reach outside their occupation ad hoc, at 43.5%, and the crossing leaves nothing behind. A consolidation you can audit. Drift is invisible until an auditor asks who reviewed something and there is no answer.

The rates are worst where the controls are mandatory

HR at 69% and legal at 56% are not incidental entries in a table. These are the two functions where the identity of the person performing the task is itself a legal fact.

An employment decision documented by someone with no HR qualification is an employment decision with no defensible record of how it was reached. A contract term drafted outside the legal function, then never routed through it, is an unreviewed obligation that the company is nonetheless bound by. Discrimination exposure, works council obligations, data subject rights, privilege: all of these attach to who did the work, not to how good the output looks.

At 69%, most identifiable HR work in these logs is being done by people whose job title is something else. We wrote earlier this year about HR as a governance surface on the strength of adoption numbers. This is a sharper reading of the same function. Adoption tells you the tool is present. Crossover tells you the role boundary is already gone.

OpenAI adds two details that widen the exposure. “Financial calculation and technology troubleshooting each appear among the three most common outside tasks in all seven other occupation groups in the analysis.” Every group in the sample is doing finance work and IT work. And marketing travels furthest of anything measured: “creating marketing materials appears across five other groups and is especially prominent among design users,” with marketing showing the highest outward share in the sample at 24.3%.

Financial calculation performed outside finance, at scale, with no reviewer, is the precondition for a restatement.

Engineering is the useful contrast

The heatmap diagonal, the share of each group’s messages that concern its own occupation, reads: customer experience 11%, design 12%, engineering 53%, finance 23%, HR 10%, legal 31%, marketing 36%, sales 12%.

Engineering stays in its lane at four to five times the rate of HR or customer experience. The function with the most mature review culture, where pull requests, code owners and required approvals are ambient rather than aspirational, is also the function whose boundaries hold. Correlation, on first-party data, not causation. It is still the single most instructive number in the table, and it points at where the missing machinery has to come from.

The benign reading, and what separates it from the bad one

OpenAI offers the charitable interpretation directly: “AI may be especially useful as a generalist tool where specialist resources are scarce.” That is real. A five-person company with no lawyer is better off with a drafted contract than with no contract, and the size data supports the scarcity story: outside-occupation task share falls from 18.9% for workspaces with 2 to 5 seats to 16.3% for those over 100 seats, though OpenAI notes the pattern breaks among the heaviest users.

The scarcity reading and the controls failure are the same behavior under two different conditions. What separates them is whether anything downstream knows the crossing happened. Marketing copy drafted by a designer and then reviewed by marketing is delegation working. The same copy shipped without that step is an unowned claim. Nothing in the usage data distinguishes the two, and nothing in a typical company’s tooling does either, which is the absence of governance tooling outside engineering showing up as measured behavior rather than as a prediction.

What this data is, precisely

OpenAI is measuring OpenAI’s own logs and framing role dissolution as expansion. Treat the framing accordingly.

The piece is published under the organization name alone. No individual researcher is credited anywhere on the page, and OpenAI files it under company announcements rather than research. It is not peer reviewed and it is not externally replicable, because nobody outside OpenAI can see the message corpus. Sample composition beyond “more than 800,000 messages from U.S. ChatGPT users” is not stated on the page: no participant count, no date range, no sampling frame. The methodology lives in a linked PDF. The 61.5 / 38.5 / 56.5 / 43.5 split comes from chart alt text rather than prose.

None of that makes the numbers wrong. It makes them a directional signal from an interested party, which is how they should be cited. OpenAI’s own claim for them is modest and probably correct: “usage patterns may provide an early signal of occupational change that conventional labor-market statistics will capture only later.”

Do this now

Pick your two most regulated functions, almost certainly HR and legal, and run a one-hour reconstruction on last month’s output.

Take ten artifacts each: an offer letter, a performance write-up, a vendor clause, a policy change. For each one, answer two questions from records that already exist. Who drafted it? Who with the relevant qualification reviewed it before it took effect?

Where you cannot answer the second question from a record, you have found an undocumented crossing. Count them. That count is your current exposure, and it is measurable today without buying anything.

Then fix the cheapest thing: make the crossing produce a record again. A required reviewer field on the template, an approval step that names a qualification rather than a person, a log line that captures who authored versus who approved. This is what role-scoped agent tooling has to be built on top of, and it is worth having in place before agents rather than after.

The handoff is not coming back. It was slow, and the speed it cost is exactly why people route around it. What has to come back is the artifact the handoff used to leave behind.

This analysis synthesizes How AI is expanding what people do at work (OpenAI, July 2026).

Victorino Group helps organizations rebuild review and sign-off records in functions where AI has already dissolved the role boundary. Let’s talk.