Key takeaways
- You can measure how much of your team's work AI actually leads, task by task, with two yes-or-no questions drawn from the six-level scale Anthropic used on its own research. Does AI do large chunks of the task? If it does, does AI carry most of the task end to end, so the person only reads the result and decides instead of staying with it to keep the work moving?
- The check runs in six steps: list last week's real tasks, split each person's week evenly across their own tasks, with a separate line for each person's version of a task, have that person and one other answer those two questions about each line separately, settle the disagreements before computing two shares, at "leads" or above and at "collaborates" or above, then freeze the list and its weights and have the same two raters rate each line again next quarter. The rated list then tells you whose job is now the review, which work to test next, and where to hold off.
- Anthropic found Claude leading 26% of its AI research work in August 2026, up from under 1% in February, a figure the company measured itself. A level says who does the work and nothing about its quality. Compare your team with itself, and keep the ratings out of performance reviews.
In this article
Ask a leadership team how much of its work AI does now, and you’ll hear a few kinds of answer. Someone points at the license dashboard, someone says “a lot, honestly,” and someone else says it’s hard to tell. All of them can be sincere, and none of them is a measurement.
The question that gets a real answer is smaller: for each task your team did last week, who led the work, a person or the AI?
Anthropic has asked that about its own research and published the answer. As of August 2026, its Claude models “lead” 26% of Anthropic’s AI research and development work, up from under 1% in February 2026 (Anthropic, “Measurements for understanding the pace of AI development inside frontier labs”).
The method behind that number is the useful part, because a scaled-down version of it doesn’t need a frontier lab.
Collaborates or leads
Anthropic rated its work on a six-level scale proposed by Epoch AI: no AI involvement, minimal AI involvement, AI assists, AI collaborates, AI leads, and AI operates fully autonomously.
The line that matters sits between the fourth and fifth levels. When AI collaborates, it “can do large chunks of work under close human direction.” When it leads, it “can complete most of the task end-to-end from a high-level prompt, while the human supervises.”
Anthropic’s own example makes the line concrete. A data pipeline breaks. At “collaborates,” an engineer brings Claude the logs, stays with it, and decides what to do each time a new problem turns up.
At “leads,” the engineer hands over the failure alert and gets on with something else. Claude works through the logs, finds the cause, writes and tests the fix, handles the surprises, and then tags the engineer, who reads the write-up and decides whether it ships tonight.
Anthropic rated how much AI is actually used in each kind of work, not what it could do, so the test is about how the work is done today. Does the person have to stay tuned in to keep the work moving, or only read the result and decide?
Even at “leads,” a person still supervises every one of those tasks, and Anthropic says the top level is one it hasn’t reached for any of the work it measured. More than 90% of the work sits at “collaborates” or higher.
How they counted
Four choices in the method are worth borrowing.
They listed the work people actually did. For each week of July 2026, Anthropic sampled 20% of staff in each research department, and a Claude research agent went through each sampled person’s week in Slack and internal documents and listed the tasks they worked on. The result was about 15,000 tasks, grouped into 378 specific kinds of work, from “serving incident postmortems” to “eval platform defect diagnosis and fixes.”
They weighted by people’s time, roughly. Each person got one unit of weight per week, split evenly across what they worked on, so work that many people spend time on counts for more. If someone worked on four tasks, each got 0.25, and Anthropic itself calls the scheme “a crude approximation.”
They froze the list. Every month is rated against the same July 2026 basket of work, using only evidence from that month or earlier, so a rising number can’t come from a different mix of work.
And they checked the ratings against people. A separate Claude judge rated each kind of work, and staff who own each area rated it separately, without seeing what the model had found. The model’s rating matched a staff member’s exactly 59% of the time and was within one level 97% of the time, while two staff members matched each other exactly only 35% of the time.
That 35% is the figure to sit with. Two people who know the same area gave it different ratings most of the time, and Anthropic names one borderline case where that happens: “where exactly ‘AI collaborates’ ends and ‘AI leads’ begins.”
Run it on one team
Pick one team and hand this to its lead. The steps follow Anthropic’s method, scaled down for one team, and where they differ from it, the choices are ours.
- List last week’s tasks, person by person, from what each person did. Calendars, tickets and sent files are better evidence than a role description, which hides the small recurring tasks where AI may already do part of the work. Name each one as a kind of work that comes back, “client proposals” rather than “the Acme proposal,” so the same list still fits next quarter.
- Split each person’s week across their own tasks, one line per person and task. Each person gets one unit for the week, split evenly across their tasks, so someone who spent the week on four tasks gives each line a quarter. Two colleagues who both write client proposals get a line each, because they may not use AI the same way.
- Ask two yes-or-no questions about every line, as that person does the work today. Does AI do large chunks of it, and if so, does AI carry most of the task end to end, so the person only reads the result and decides instead of staying with it to keep the work moving? A no to the first puts the line below “collaborates,” a yes then a no makes it “collaborates,” and two yeses make it “leads.” Two people answer separately, the person whose line it is and someone else who knows that work, such as whoever reviews it, because two people who know the same work can still see it differently.
- Read the disagreements before you compute the shares. Go through every line where the two raters answered differently and agree one answer, because a share computed from ratings nobody agrees on measures the raters, not the work. A split on the second question is normal, since Anthropic itself names that boundary as borderline, and where a pair still can’t agree, count the lower answer and mark the line. A line one rater puts at “leads” and the other below “collaborates” is worth a conversation: the raters may be picturing different work, or one of them may not know how the AI is being used.
- Compute two shares: the weight at “leads” or above, and the weight at “collaborates” or above, each as a share of the total. Anthropic reports both, and the gap between them is work where AI already does large chunks but someone still has to stay with it.
- Freeze the list and its weights, and rate the same list again next quarter. The frozen list is what makes the second number comparable with the first, as long as the same two people rate each line again.
Take a proposal your sales team sends. If AI suggests sentences while someone writes, it assists, and if it drafts whole sections while the person steers each one, it collaborates.
If the person hands over the brief and the client file, and the AI produces a complete draft, flags what it couldn’t confirm, and leaves the person to read it and decide whether it goes out, it leads. The task and the tool stay the same while the level changes, and what changes it is who keeps the work moving.
The fine print on a task list
A level tells you who does the work. Whether the work is good, or the team got faster, needs its own numbers, measured use case by use case rather than as adoption overall.
Don’t compare your share with Anthropic’s 26%. The work is different, the raters are different, and Anthropic’s figure is the company measuring itself, with independent outside evaluators still being set up.
On a small team the percentage will jump around, because the list is short and one line changing level moves it. Read the task list, not the decimal.
A frozen list can’t see new work, either. If people move onto tasks that didn’t exist in the base week, the old list won’t show it. Anthropic checked for this separately, found no rise in new kinds of tasks in its own research, and says it plans to rebuild its list periodically.
Keep the ratings out of performance reviews. Once a level affects someone’s review, people will rate to protect themselves, the same trap as ranking employees by how much AI they use. That point is our own reasoning, and Anthropic didn’t measure it.
Who reviews, what to test, where to wait
The rated list doesn’t need a benchmark to be useful. It points to three decisions.
Where a line sits at “leads,” the job on that task is now the review, the last check before the work goes out. Decide who owns that review, and whether they have the time and the knowledge to do it well.
Where a lot of weight sits at “collaborates,” AI already does large chunks of the work. That’s the work to test next: pick one of those tasks and measure its quality and speed now (for proposals, say, how many go out without rework and how long each takes). Then hand it over the way the proposal example does, with AI given the brief and the material and the person only reading the result and deciding, and at the next quarterly rating keep it there only if both numbers are at least as good.
Where step 4 marked a line, or found one rater at “leads” and the other below “collaborates,” hold off on changing that work until the people involved agree how it’s actually done, not just how to rate it.
Anthropic’s number went from under 1% to 26% in about six months, and that movement is readable because every month was rated against the same list of work, in the same way. A team that starts its list this quarter will be able to see its own movement next quarter. One that doesn’t will keep answering the question with a dashboard and a guess.
Questions this article gets
What does it mean that AI 'leads' a task?
On the scale Anthropic used, proposed by Epoch AI, AI leads when it can complete most of a task end-to-end from a high-level prompt while a person supervises. One level down, AI collaborates: it does large chunks of work under close human direction. The practical test is whether the person has to stay with the AI to keep the work moving, or only reads the result and decides.
How did Anthropic arrive at 26%?
It sampled 20% of staff in each research department for each week of July 2026, had a Claude research agent list the tasks each person worked on, and grouped them into kinds of work. Each kind was weighted by splitting each sampled person's week evenly across the tasks they worked on, and a separate Claude judge rated each kind on the six levels. Staff who own each area rated it separately as a check. The figure is Anthropic measuring itself, and the company says it is now setting up independent outside evaluators to monitor numbers like this one.
Should we compare our team's share with Anthropic's 26%?
No. The work is different, the raters are different, and Anthropic's figure is self-measured. The useful comparison is your own team against itself, on the same frozen list of tasks, a quarter apart.
