AI Is Not About Doing the Same Things Faster. New OpenAI Data Shows What It Actually Changes

I first proposed Impact per Hour as a guiding metric in September 2025, and it has become the spine of how we teach AI to leaders.

The idea starts from a complaint about measurement. We have measured productivity in hours and outputs for far too long, and optimizing for outputs quietly loses sight of what those outputs were meant to achieve.

Impact per hour measures the value someone creates in an hour of work, not the hours they log. It grows along two separate dimensions, and this is the part that matters here.

The first dimension is capacity. This means freeing people from the roughly 30 to 40 percent of time typically lost to low-value, repetitive work. Often that means redesigning the workflow so the low-value part disappears entirely, rather than doing the same steps faster.

The second dimension is capability. This means AI enabling people to do things they simply could not do before. Someone who is not a strong writer produces strong writing. Someone who never learned design creates good design. Someone who struggles to build presentations now builds excellent ones with AI PowerPoint generators.

Capacity gives you back time. Capability expands what you are able to do at all.

Increase both together, and impact grows faster than either change delivers alone.

Almost Everyone Optimizes the First Half and Ignores the Second

Here is the honest problem with how impact per hour gets received. Capacity is easy to sell and easy to count. You can put "saved four hours a week" on a slide.

It is also the half that leaks. Time you reclaim and never direct behaves like a budget nobody allocates, which is why we ask leaders to set a North Star for the reclaimed time before it arrives.

Capability is harder still. It asks a leader to accept that their own job description is now negotiable, which is a much bigger ask than adopting a tool.

So which one actually shows up in real usage? Until now, we could only argue from principle and from what we saw in the room. The OpenAI data answers it directly.

The method is worth understanding, because it is what makes the finding credible.

The researchers took work-related messages from users across eight occupation groups: customer experience, design, engineering, finance, human resources, legal, marketing, and sales. They mapped each message to a single work activity in O*NET, the US Department of Labor's database of occupations and tasks.

Then they compared the task in the message against the occupation of the person sending it. Every message landed in one of three buckets.

• Generic, meaning broadly shared work like writing emails or scheduling meetings, which came to 61.5 percent of all messages.
• Within occupation, meaning tasks historically tied to the sender's own job, at 21.8 percent.
• Cross-occupation, meaning tasks historically associated with a different job entirely, at 16.8 percent.

Strip out the generic work, and the headline number appears. Among occupation-specific messages, 43.5 percent involve tasks that historically belonged to another occupation. They call this task crossover.

(A word of caution on that 43.5 percent, since it will get quoted everywhere. It only exists after removing the 61.5 percent of generic writing and scheduling. The honest denominator for "how much AI use crosses job boundaries" is 16.8 percent. Both numbers are real, and they answer different questions. We keep both in our running collection of AI statistics.)

This Is the Capability Dimension, Measured

Look at what people are actually asking for outside their own lane.

Among non-marketing users, the single most common marketing-related task is developing promotional materials, at 25 percent of their marketing-related messages. The report describes what sits inside that category: requests to create ads, social posts, flyers, promotional videos, presentations, product sheets, and campaign graphics.

Among non-finance users, calculating financial data ranks in the top three finance tasks in every single one of the other seven occupations. Among non-engineering users, troubleshooting computer applications and systems does exactly the same thing.

Read that list again next to the capability definition. Someone who never learned design is making graphics. Someone who is not an analyst is running the numbers. Someone who cannot code is troubleshooting the application.

That is not the same work performed faster. That is work the person could not previously do at all, which is precisely the claim we have been making.

Cross-occupation work is now the majority of occupation-specific AI use in five of the eight groups: 77 percent for customer experience, 75 percent for design, 69 percent for human resources, 56 percent for legal, and 53 percent for marketing.

The report finds something more useful than a single average, because crossover runs in two directions and they do not move together.

Designers import heavily. About 35.2 percent of messages from designers involve work usually associated with another occupation, which puts them at the top of a tight cluster that also includes sales at 32.1 percent and both human resources and customer experience at 30.4 percent. Yet design tasks make up only 1.7 percent of messages sent by everyone else.

Engineering is close to the reverse. Only 18.5 percent of engineering messages reach outside the field, but engineering tasks account for 7.4 percent of messages from workers in other occupations.

Marketing does both. Marketers spend 24.3 percent of their messages on other occupations' work, and marketing tasks show up in 8.9 percent of everyone else's messages, the highest share in the sample.

Why does this matter for how you train people? Because it means a designer and an engineer need materially different AI training. One is constantly borrowing capability from other functions. The other is the function everyone else is borrowing from.

A single curriculum serves neither of them well, which is why we design company-wide AI training around what each function actually reaches for rather than one shared syllabus.

The full picture is a matrix of every function against every task source, and it makes the pull toward marketing and engineering hard to miss. 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. Engineering tasks account for 28 percent among design users, and about 20 to 22 percent among customer-experience and finance users.

Smaller Teams Cross More, and That Fits the Theory

There is one more finding worth your attention. Among typical-volume users, defined as the middle 50 percent by message count, cross-occupation use falls from 18.9 percent in workspaces of 2 to 5 seats to 16.3 percent in workspaces of 101 or more.

That is a gap of roughly 2.5 percentage points, or about 13 percent relative to the smallest workspaces.

The researchers offer a plain explanation. A worker in a small business may use AI to draft marketing copy, troubleshoot software, review a contract, or run basic analysis because there is no specialist to delegate to. In a larger organization, the same person hands it to a team.

They are careful to note the limits, and so should we. Among heavy users there is no consistent trend at all, and workspace seat count is not the same as company size.

I want to be precise here, because this research is going to get overclaimed.

The report is descriptive, and it says so repeatedly. The unit of analysis is a message, not an hour, a project, or a job. The researchers do not observe whether the output was used, how good it was, how much time it saved, or whether a specialist reviewed it afterward. The sample is not representative of the US workforce.

They are also explicit that this is not a story about occupations disappearing. AI may make a task easier for an outsider to attempt while specialists remain critical for expert-level judgment and review. Their words are that the division of work is shifting, not that particular jobs are being eliminated.

That distinction is the same one we draw in people-centric AI transformation. There is no fixed ceiling on how much can be automated, but there is a hard line on accountability.

And the report lands on that line itself. Its conclusion notes that if this becomes permanent, workers will need training to evaluate AI-assisted work outside their established expertise, and organizations will need clear processes for review and accountability.

A marketer generating a financial model is a capability gain. A marketer generating a financial model that nobody qualified ever checks is a liability with a nice chart on it.

Your Org Chart Is Drifting From Reality

Buried near the end is a line that deserves far more attention than it will get.

The researchers point out that most studies of AI and work begin with a fixed list of tasks and ask which ones a model can perform. Their evidence suggests the list itself is changing. And if AI changes which workers perform which tasks, then measures built only on existing job descriptions will gradually drift away from how work is actually organized.

Your org chart is becoming a less accurate description of your company every quarter.

This is why we keep insisting that you start with tasks, not tools, and why Workflow Literacy is the skill underneath all the others. Work has to be broken down to the atomic task level before AI can be applied to it well. OpenAI arrived at the same unit of analysis from the data side.

It also explains the capability overhang we keep running into. The capability is already available to almost everyone. The breadth of use is what separates the leaders from the dabblers, and that is a question of AI fluency, not access.

The Bottom Line

1. Capacity is the easy half of impact per hour, and capability is where the leverage is. Time saved on your existing tasks has a ceiling. Doing work you previously could not do at all does not.
2. 43.5 percent of occupation-specific AI use now involves another occupation's work, according to OpenAI's July 2026 analysis of over 800,000 messages. Use 16.8 percent when you want the conservative figure across all work-related use.
3. Marketing and engineering tasks travel furthest. If you are choosing where to build shared capability across your organization, start with the ability to make things and the ability to troubleshoot things.
4. Train by direction, not by average. Some functions borrow constantly and some get borrowed from. Use the GED-RT framework to decide which of those borrowed workflows are worth building for first.
5. Capability without review is a liability. Decide now who is accountable for judging AI-assisted work that falls outside the person's own expertise.

The most important question this research raises is not whether your people are using AI. It is whether you have decided what the expanded version of each role is supposed to look like, or whether you are going to let it happen by accident and find out from a dashboard.

This connects to something I wrote about recently on measuring useful work rather than adoption. Counting who has a license tells you nothing. Counting how far outside their old job description someone can now operate tells you a great deal.

We do not fix a broken world of work by doing yesterday's tasks faster. We fix it by deliberately raising what a person is able to do, and then deciding what that new capacity is for. That is the work we do with leaders in AI Leader Advanced.