OpenAI found the smallest offices get the most out of AI. That's most trade sub back offices.
OpenAI's new research on 800,000 work messages found workers in 2-to-5-person workspaces use ChatGPT to cross into other people's job duties far more than workers at large companies. That's the exact staffing shape of most trade sub and small GC back offices — and it's a specific argument against your next specialist hire, not a vague one.
OpenAI just published the largest study yet of how people actually use ChatGPT at work, and the headline number cuts against how construction back offices usually think about AI: the smallest offices — not the biggest — get the most cross-role value out of it [1][2]. Among users in workspaces of 2 to 5 seats, 18.9% of work messages involved a task outside that person's own occupation, compared with 16.3% in workspaces over 100 seats [1][2]. Most trade sub and small GC back offices are built exactly like the smaller group: three to six people covering estimating, PM support, accounting, HR, and marketing between them, with no dedicated specialist in most of those lanes.
What is "task crossover," and why does OpenAI think it matters?
OpenAI's Work at the Frontier report analyzed more than 800,000 US business ChatGPT messages and found that 43.5% of occupation-specific requests involved a task historically owned by a different job [2][3]. Examples the report gives: a financial analyst using ChatGPT to write marketing copy, a legal worker running financial calculations, a designer drafting release notes, a customer-experience lead writing a SQL query [3][4]. Across all work messages, not just the ones tied to a specific occupation, the crossover rate was 16.8% [4][5]. OpenAI's own explanation for the size effect: smaller businesses "may see stronger effects because they often have fewer specialist teams and less formal delegation," making AI "especially useful as a generalist tool where specialist resources are scarce" [4] — a description that fits a 40-person mechanical sub's office better than a 4,000-person GC's.
Which tasks travel the farthest?
The report broke crossover down by occupation. Customer-experience workers had the highest rate at 77% of their occupation-specific prompts, then designers (75%) and HR staff (69%). Legal came in at 56%, marketing at 53%, sales and finance both at 40%, and engineering lowest at 28% [3]. Marketing and technology-troubleshooting tasks were among the three most common "borrowed" tasks across nearly every other occupation studied [3] — meaning the two functions a small trade sub is least likely to have a dedicated hire for (marketing, IT) are the two that show up most often getting done by someone else with AI's help.
| Function | Crossover rate (share of that role's occupation-specific prompts that are "someone else's" task) |
|---|---|
| Customer experience | 77% |
| Design | 75% |
| HR | 69% |
| Legal | 56% |
| Marketing | 53% |
| Sales | 40% |
| Finance | 40% |
| Engineering | 28% |
Should this change who a trade sub hires next?
Only for a narrow, specific decision, and OpenAI is upfront about the limit: the report measures message patterns, not output quality, and doesn't establish whether AI use reflects a new responsibility or one the person already had informally [4]. It's not evidence that a controller doing occasional HR paperwork is doing it well, and it's not a productivity study. What it does support is a concrete test before the next single-function hire — a part-time marketing person, a first in-house HR hire, a bookkeeper for a task the office manager already handles: check whether the crossover is already happening informally with an existing employee and a ChatGPT or Claude account, the way it already is at 18.9% of small workspaces studied. If it's not landing well, that's the actual signal to hire a specialist — not the absence of the tool.
The honest limit here
This isn't a construction-specific study — OpenAI's occupation categories are broad white-collar roles, not "estimator" or "submittal coordinator," and the 800,000-message sample skews toward however OpenAI's business customer base is distributed, not a construction-representative panel. Read it as a directional signal about small-office AI use generally, one that happens to match the staffing shape of most trade sub back offices, not as data collected on contractors. And crossover usage is not the same as competent output — a PM using AI to draft an offer letter still needs someone who knows employment law to check it before it goes out.
We covered OpenAI's small-business AI push in July and flagged that none of its named integration partners touch construction software directly — this report is the data behind why that push still lands in the back office anyway.
Friday one chart. Every week, one piece of data that should change a decision on your project. Subscribe at constructionaibrief.com.
- What did OpenAI's task crossover report actually find?
- OpenAI analyzed more than 800,000 US business ChatGPT messages and found that 43.5% of occupation-specific work requests involved a task historically associated with a different job — for example, a finance worker writing marketing copy or a designer drafting release notes. Across all work messages (not just occupation-specific ones), the crossover rate was 16.8%.
- Does company size change how much task crossover happens?
- Yes. OpenAI found the crossover rate was 18.9% for users in workspaces of 2 to 5 seats, versus 16.3% for users in workspaces of more than 100 seats — meaning the smallest offices, not the largest, get the most cross-role use out of the same tool.
- Which jobs cross into other roles the most, according to the report?
- Customer-experience workers had the highest crossover rate at 77% of their occupation-specific prompts, followed by designers (75%) and HR staff (69%). Legal was 56%, marketing 53%, sales and finance both 40%, and engineering the lowest at 28%. Marketing and technology-troubleshooting tasks showed up most often as the 'outside' task other roles borrowed.
- Should a trade sub skip hiring a marketing or HR specialist because of this data?
- Not on this data alone — OpenAI is explicit that the report measures usage patterns, not whether the work was done well or whether it changed an actual hiring decision. What it supports is a narrower call: for a single, low-volume specialist function (a part-time marketing hire, a first HR hire), test whether an existing employee with ChatGPT or Claude can already cover it before opening a req.
- What should a small contractor's office actually do with this?
- Map which specialist tasks currently get done informally by non-specialists in your office — a PM writing proposal copy, an estimator drafting an offer letter, a controller doing basic HR paperwork — and check whether that person is already leaning on AI for it. If they are, that's your workspace confirming OpenAI's pattern, and it's a reason to formalize the practice with a shared prompt library or account rather than hire around it.