Epoch AI, a research group that tracks how fast AI is improving, gave six leading models real jobs from its own work: drawing charts in its house style, writing short data articles, researching data centers and designing experiments. Its staff graded the results against their own standards. The two best models, Claude Fable 5.1 and GPT-6 Astra, were reliable on the well-defined parts and weak on the parts that need judgment.
The well-defined steps are going
One job got so reliable that Epoch took it out of the test and now has AI do it: restyling a chart to match Epoch's look, using the team's written instructions and ready-made chart parts. Earlier this year, models of this kind could not move an article from one publishing site to another. Newer ones can. Any step in your week that comes with written instructions, a template and an obvious right answer is in the same position. A team that publishes a chart a day gets a small daily job gone for good.
A wrong filter nobody checked
Another model, Kimi K3, was asked to write a short data article on how often physics papers mention AI. It counted mentions across all the papers on the preprint site arXiv, not only physics, never checked that its filter worked, and built the whole article on the result. None of the leading closed models made a factual error in their data articles, but the lesson holds for any of them. A clean summary of sales by region is only as good as the filter behind it, and the first question to ask of any AI analysis is what it counted.
Examples teach what to copy, not why
Epoch gave Fable 5.1 its whole design file, with months of past work, a color palette and guidelines. For a simple idea, Epoch's designer drew a simple chart, and Fable 5.1 produced a crowded diagram. It had borrowed small "A" and "B" badges from a complex example, where they labeled two concepts that appear in several places. Fable 5.1 used them to label single items, so it copied the look without learning why the original needed them.
Every company has rules like this that nobody wrote down: how long a customer email should be, which details a client cares about, what the audience finds interesting. GPT-6 Astra had a link to every article Epoch had published and still picked a topic too narrow for its readers.
There is a second effect. Three of the six models picked the same topic when asked for a data article, though many were open to them. If your competitors use the same models with the same instructions, their drafts will look like yours. What stays different is the taste and the unwritten standards you bring.
What the spreadsheet did to bookkeeping
When spreadsheet software arrived in the late 1970s, it did in seconds what used to take a person a day. Many bookkeepers and clerks who added up numbers were replaced. But the number of accountant jobs went up, because once answers were cheap, people asked accountants to do more with them.
AI is already beating junior accountants at bookkeeping tasks, so the first half of that story is repeating. The second half depends on the skill Epoch's test shows models lacking: deciding which question to ask and whether an answer can be trusted. A finance analyst who spends two days a month on a cost report could get a first draft in minutes. The hours that matter then go to asking whether the draft counted the right accounts, and whether anything in how it was built could have produced that number on its own.
That habit is worth building: look at a finished-looking result and ask what it counted, what it left out, and what in the setup could have created the pattern by itself. Models do not yet do this for their own work, so for now it is the part of the job that stays yours.