What Anthropic's two million conversations say about hours saved

Anthropic published research in January 2026 drawn from two million conversations. It puts a figure on the saving that every AI ad promises you. A task a person would do alone in about 3.1 hours took about 15 minutes with AI in that sample. Roughly twelve times faster, on one task, when the output came back usable.

As of 13 September 2026.

What the two million conversations actually measured

The report is Anthropic's Economic Index, published on 15 January 2026. It draws on one million conversations from Claude.ai and one million records from Anthropic's own API traffic, sampled over one week in November 2025. Classifiers estimated two things for each task: how long a person would need alone, and how long the task took with AI. The median alone was 3.1 hours. The median with AI was about 15 minutes. The report is at https://www.anthropic.com/research/anthropic-economic-index-january-2026-report

The same report scored whether the task worked. Classifiers labeled 67 percent of the Claude.ai sessions successful. On the API traffic the figure was near 49 percent. That second number is the one an owner should write down, because it sets the terms of the first one.

What can I use AI for in my business?

The saving lands on written and analytical work you already do every week. It does not land on new work. In the sample, the ten most common tasks on Claude.ai accounted for 24 percent of usage on that service. For an owner, the recognizable ones look like this. A month-end debtor ageing list read down to who to call first and what to say. A supplier comparison pulled into one table with the source URL beside every row. A monthly sales report turned into the top insights, the questions to ask, and the next actions.

Each of those has a right answer you can check in minutes. That is why hours come off them. Work with no checkable output does not move: a pricing call, a hiring call, a decision about a long-standing partner.

Why the 15 minutes will not show up in your month

One session in three was not labeled successful in the Claude.ai sample. The hour you save on a task is only saved if you can tell which third you are holding. Judging that for yourself is the work, and no figure in the report measures it for you.

One other thing in the data cuts the headline down. Coding was roughly a third of the conversations, so the average carries work your business does not do.

So the honest version of the finding is small and useful. On a task you can check, the time falls from hours to minutes. On a task you cannot check, the time moves from doing the work to arguing with an output.

How do I get better with AI from here?

You get better by timing one task, not by adding course hours. Take one recurring task this week, run it both ways, and time both. Then check three claims in the output against a document you already trust. That check, repeated, is what turns a 15 minute draft into 15 minutes you can bank.

How do I know when an output is wrong?

Check the claims you can already verify. Take three facts from the output, a number, a name, a date, and match them against your own records or the source page. If one of the three is wrong, read the rest slowly and check more before you act on any of it.

Which task should I start with this week?

Start with the task you do every week that produces a document somebody else reads. A sales summary, a supplier note, a customer reply. It repeats, so the time back compounds, and you know what right looks like, so you can catch a wrong output on sight.