Nearly Half of Job-Specific AI Use Is People Doing Another Job's Work. At Your Dealership, That Job Is Marketing.

This morning OpenAI published research that isn't about models, benchmarks, or agents. It's about who, inside a company, actually does which job, and the answer turns out to be changing faster than any org chart.

The report is Work at the Frontier: How AI is Expanding What People Do at Work, by OpenAI's economic research team. They took a random sample of over 800,000 work-related messages from the individual ChatGPT accounts of U.S. users whose occupations were known, sorted every message into a task using O*NET (the U.S. Department of Labor's database of occupations and work activities), and then asked a simple question: is this person's AI request inside their own job description, or somebody else's?

The headline number: 16.8% of all work-related messages, and 43.5% of occupation-specific messages, concern tasks historically associated with another occupation. OpenAI calls it task crossover. Nearly half the time someone uses AI for job-specific work, they're doing a job that isn't technically theirs.

Then they broke it down by which job. And one category traveled farther than every other one in the sample.

Marketing.

The work that moves, and the work that stays put

The study covers eight occupation groups: customer experience, design, engineering, finance, HR, legal, marketing, and sales. Two directions matter: how much a group borrows from other jobs, and how much its own work gets borrowed by everyone else.

Design borrows heavily and exports almost nothing: 35.2% of designers' messages involve other people's work, while design tasks make up just 1.7% of everyone else's. Engineering is the mirror image: only 18.5% borrowing, but engineering tasks show up in 7.4% of other occupations' messages.

Marketing does both. Marketers spend 24.3% of their messages on other people's jobs, and marketing tasks account for 8.9% of messages among workers in other fields, the highest outward share in the sample. Marketing is simultaneously the most porous department and the most exported one.

Now look at who's doing the exporting-into. Per the report, "marketing tasks account for about 28 to 29 percent of non-generic messages among sales and design users and 26 percent among customer-experience users."

Read that again with a dealership in your head. Among salespeople, more than a quarter of job-specific AI use is marketing work. Among customer-experience workers, the closest thing in this taxonomy to a BDC, it's 26%, and their overall cross-occupation rate is the highest of any group at 77%.

And what specifically are non-marketers making? OpenAI lists it plainly: developing promotional materials is the single most common marketing task among non-marketing occupations, at 25% of their marketing-related messages, and the requests include "ads, social posts, flyers, promotional videos, presentations, product sheets, and graphics for marketing campaigns."

Ads. Flyers. Product sheets. Made by people whose job title says sales.

Why this lands harder on a rooftop than on a Fortune 500

Here's the finding most people will skip, and it's the one that decides whether this is your problem.

Crossover is more common at small companies, not large ones. Among typical-volume users, the cross-occupation share falls from 18.9% at the smallest workspaces to 16.3% at workspaces with 101 or more seats. OpenAI's explanation is almost a description of a car store: "A worker in a small business may use AI intermittently to draft marketing copy, troubleshoot software, review a contract, or perform basic analysis because there is no specialist to delegate to."

Or, as the summary puts it: "AI may be especially useful as a generalist tool where specialist resources are scarce."

Your store probably runs its entire marketing function with one person, or with none, the work spread across a handful of vendors who don't talk to each other. There is no in-house designer to hand it to. There is no analyst. When a salesperson needs a flyer for the used F-150 that's been sitting 70 days, the specialist they'd delegate to does not exist. That's precisely the condition this research says produces the most crossover.

None of this is theoretical, and it isn't only OpenAI's finding. Atlassian's Teamwork Lab surveyed 1,000 U.S. knowledge workers between May 27 and June 9 and reported on July 6 that 92% say their responsibilities expanded beyond their original job description in the past year. The heaviest AI users were nearly twice as likely to take on work from other teams, and twice as likely to handle specialized tasks without looping in an expert.

That last clause is the whole risk in nine words.

What this actually means for your dealership

Strip the labor economics away and here's the operational reality: your store already has more people making marketing than you think, and none of it is going through whatever review process you believe exists.

Four specific consequences.

1. Customer-facing creative is being produced outside your approval chain. A salesperson generating a payment-focused social post or a vehicle flyer is producing advertising with your dealership's name on it. Price, payment, term, rebate eligibility, OEM logos, stock photography: every one of those elements carries rules that came from your compliance policy, your OEM's co-op guidelines, or your state's advertising requirements. The person making the flyer at 8:40 on a Tuesday night is not thinking about any of them. They're thinking about the F-150.

2. The math is moving too. Calculating financial data ranks among the three most common finance-related tasks in every non-finance occupation group in the study (9.8% of all finance-related messages sent by non-finance workers), and Table 2 lists sales as the top group for it, at 14.3%. In most industries that means a rep modeling a discount. In this one it means somebody who is not in F&I doing payment math with a chatbot and then saying that number out loud to a customer.

3. Your vendor scope just got a real test. If a quarter of your salespeople's AI use is already marketing work, some tasks you're paying for by the month have quietly collapsed to near-zero marginal cost inside your own building. That is worth knowing before your next renewal. But OpenAI is careful here, and so should you be: "AI may make a task easier for an outsider to attempt while specialists remain critical for expert-level judgment and review." The production of a flyer collapsed. The judgment about whether that flyer is legal, on-brand, correctly priced, and pointed at the right inventory did not. Confusing the two is how dealers end up paying an agency for work they can't evaluate, or firing the agency and discovering nobody was watching the disclaimers.

4. Nobody is reviewing the output, including the people who made it. OpenAI is blunt in its own limitations: they don't observe "whether the output was used, how good it was, how much time it saved, whether the user could have completed the task without AI, or whether a specialist reviewed it." The study's own conclusion says workers "may need training to evaluate AI-assisted work outside their established expertise, and organizations will need to prepare clear processes for review and accountability." Nearly every dealership I know has a process for who can approve a deal. Almost none has one for who can publish an ad.

The limits, stated plainly

You'll see this study quoted in vendor decks by Friday, so here's what it does not say.

It's first-party data from a company that sells the product: OpenAI benefits commercially from the finding that ChatGPT expands what workers can do. The sample is drawn from users linked to ChatGPT Business accounts and, in OpenAI's words, is "not representative of the entire U.S. workforce"; the estimates explicitly should not be generalized to Enterprise users. The unit of analysis is a message, not an hour, a project, or a job. And the results are, in their own words, "descriptive": they show what people asked AI to do, not what happened next, and not what those workers would have done without it.

Most importantly for us: there is no automotive occupation group in this study. "Sales" here means salespeople generally, not car salespeople specifically. Nobody measured your showroom. What the data gives you is a strong directional signal about how work redistributes in small, specialist-poor organizations, and a dealership is the textbook case of one.

What to do about it

1. Find out who's already doing it: this week, without a policy memo. Ask your sales managers and BDC lead one question: who on your team is using ChatGPT or Gemini to write customer-facing anything? You will be surprised by the number, and you cannot govern what you haven't counted.
2. Set one rule, not a handbook. Anything with a price, a payment, a term, an APR, a rebate, or an OEM logo gets reviewed by whoever owns your advertising compliance before it goes anywhere public. One rule people remember beats a twelve-page policy nobody reads.
3. Give the crossover somewhere to land. A shared folder with approved disclaimer language, current OEM-compliant logos, real photography of your actual inventory, and three or four prompts that already work. The work is happening either way. Your only choice is whether it starts from your assets or from a blank box.
4. Train evaluation, not prompting. Your people can already get output. What they can't reliably do is judge whether the output is right, which is exactly the gap both studies point at. Teach the sales floor what a compliant price claim looks like at your store, and you've converted a liability into capacity.
5. Re-read your vendor scopes with this lens. Separate production from judgment, line by line. Pay for the judgment, the strategy, the review, and the things that require access you don't have. Reconsider the retainer line items that are pure production your team is already doing at 9 p.m. anyway, and if you're not sure which is which, start with what you're actually getting for your content spend.
6. Watch the photos specifically. AI-generated and AI-edited vehicle imagery is the fastest-moving piece of this, and it now carries disclosure exposure of its own. A salesperson touching up a stock photo is the most likely place your store gets embarrassed first.

The bottom line

OpenAI set out to study the shape of work and produced, incidentally, the clearest picture yet of what's happening on dealership sales floors: the marketing department is now everybody, and nobody signed off on that.

This is not a reason to lock it down. Crossover in a small business is mostly a good thing: people solving problems at the point they encounter them instead of waiting a week for a vendor ticket. The dealers who get hurt won't be the ones whose staff started making marketing. They'll be the ones who never noticed, never gave those people the right assets, and never built the ten-minute review step that catches a wrong payment before it reaches a customer.

The job descriptions in your DMS are about to be as out of date as the government's are. Update the review process first; the org chart can wait.

If you want an outside read on what your store is actually publishing right now (and what an AI says about your pricing and your fees when a shopper asks), book a walkthrough with us and we'll go through the real answers together.