Salespeople Spend More AI Time on Marketing Work Than on Sales Work. OpenAI Has the Numbers.

July 28, 2026. OpenAI published new economic research yesterday that should change how you staff a revenue team. Its Economic Research group analysed more than 800,000 work related messages from US ChatGPT users and found that a large share of the work people do with AI belongs, on paper, to somebody else's job. The report is called Work at the Frontier, and the single most useful number in it, for anyone selling anything, is this: when a salesperson uses AI for something specific to a role, that task most often looks like marketing, not sales.

What the research actually found

1. Across all work related messages, 16.8% involve tasks associated with a different occupation than the user's own. Among occupation specific messages, that share rises to 43.5%.
2. OpenAI first strips out generic work such as writing, summarising and scheduling, because those appear across every job and prove nothing. Generic use accounts for 61.5% of the sample. Of the remaining 38.5%, 43.5% falls outside the user's occupation.
3. The pattern is strongest in customer facing and support roles. Once generic work is excluded, outside occupation tasks account for 77% of occupation specific messages from customer experience workers, 75% from designers, 69% from human resources, 56% from legal and 53% from marketers.
4. Marketing work travels further than any other kind. Marketing tasks account for 8.9% of messages among workers in other fields, the highest outward share in the sample, while marketers themselves spend 24.3% of their messages on tasks associated with other occupations.
5. Engineering is the other big exporter. Only 18.5% of engineering messages involve outside tasks, but engineering tasks show up in 7.4% of messages from everyone else. Financial calculation and technology troubleshooting each rank among the three most common outside tasks in all seven other occupation groups.
6. Design is the reverse case. Designers pull in a lot, with 35.2% of their messages involving other people's work, but design tasks appear in only 1.7% of messages elsewhere.

Why the sales row is the one to read twice

OpenAI published a crossover heatmap that asks, for each non generic message, which occupation the task most closely resembles. For workers whose own job is sales, the answer breaks down like this: marketing 29%, engineering 18%, finance 15%, customer experience 11%, sales 12%, legal 6%, design 5% and human resources 4%.

Read that again. A salesperson's occupation specific AI use resembles marketing work more than twice as often as it resembles sales work. Not because sellers stopped selling, but because the parts of selling that AI is best at, writing the sequence, researching the account, building the one page teardown, segmenting the list, rewriting the subject line, are all tasks a marketing team used to own and used to queue.

The finance and engineering columns tell the same story from a different angle. Sales users are running their own pricing math and their own tooling fixes rather than filing a ticket. That is 33% of their occupation specific AI use going to two functions most small revenue teams do not have on staff at all.

Small teams cross the most lines

The report also breaks crossover down by company size, and the direction is the one you would expect but rarely see measured. Among average users, the outside occupation task share falls from 18.9% in workspaces with 2 to 5 seats to 16.3% in workspaces with more than 100 seats. In a big company the specialist exists and the handoff is available. In a small one, the person who hits the problem is the person who solves it.

OpenAI's own framing is that AI may be especially useful as a generalist tool where specialist resources are scarce. For a five person agency or a founder led sales motion, that is not an insight, it is a description of Tuesday. What is new is that there is now platform level evidence for it, which means the staffing argument you have been making on instinct has a citation.

What it means for operators

The obvious read is that you need fewer people. That is the wrong read, and it is the expensive one. What the data actually shows is that role boundaries have moved, not that headcount has become optional. Three consequences follow.

First, quality control moved with the work. When a rep writes their own campaign copy, nobody with marketing judgement reviews it before it hits two thousand inboxes. That is a deliverability and brand problem before it is a productivity win, and it is exactly the failure mode we see in audits of cold email programmes that scaled faster than their review process.

Second, the shared assets matter more than the individual prompts. If five reps are each independently reinventing the ICP definition, the offer, and the objection handling, you do not have leverage, you have five inconsistent versions of your company. The fix is boring: one source of truth for positioning, one approved message library, one enrichment and scoring standard feeding the whole lead generation motion.

Third, the tasks that crossed over are the ones worth automating properly. Research, enrichment, list building, first draft personalisation and CRM hygiene are now being done manually, in a chat window, by your most expensive customer facing people. That is the definition of a workflow that should be an automation instead, so the rep spends the reclaimed hour on the call rather than on the spreadsheet.

A five step plan for this quarter

1. Ask each seller to list the last ten things they used AI for at work. You will get a task inventory that no job description contains.
2. Sort that list into three buckets: should be automated, should be a shared asset, should stay a human judgement call. Most lists come back roughly a third in each.
3. Automate the first bucket properly, with a system that logs what it did, rather than leaving it in individual chat histories nobody can audit.
4. Put one review gate on anything that goes to a prospect at volume. One person, one pass, before send. It costs minutes and protects the sending domain.
5. Measure one number before and after, ideally meetings booked per rep hour. Task crossover is only a win if it converts into time back on the phone.

What the data does not say

Two honest caveats. This is OpenAI measuring usage of OpenAI's own product, so it captures what people do with ChatGPT, not what they do with every tool, and users of a leading AI assistant are not a random sample of workers. And crossover is a measure of activity, not of quality: the report shows that a salesperson is doing marketing shaped work, not that they are doing it well. OpenAI is explicit that usage patterns are an early signal of occupational change that conventional labour statistics will pick up later, which is the right level of claim.

Even with those limits, the direction is hard to argue with. The job descriptions in your ATS describe an org chart that your own team stopped following months ago. The businesses that win the next year are the ones that redraw the chart on purpose rather than discovering it in an exit interview.