Article · AI, and what it’s actually doing
AI gets you most of the way. The last part is still yours
Epoch AI tested the best models on its own staff's real work. They met 65% of the standard. That's a useful target, not a reason to hand over the rest.
Epoch AI, a research group that tracks how fast AI is improving, has tried something most AI tests don’t. Instead of scoring models on exam-style benchmarks, it gave them the jobs its own staff do every week, such as graphic design and planning research projects, and marked the results against the same standard it uses for people. Epoch published the summary.
The best performers were Claude Fable 5.1 and GPT-6 Astra. On average their work met 65% of the checklist a human employee would be judged on. A score of 100% would mean the task could be handed over entirely. Epoch is careful about what 65% means: the models did well on the easier parts and struggled with the harder ones. They could suggest promising research ideas, for example, but not turn them into experiments that would actually work.
Is 65% good enough?
I think 65% is a good target for most businesses to aim at, and I’d treat it as a ceiling, not a starting point for handing over the rest. I would never want to submit something that was 100% done by AI. It isn’t there yet, and everything needs finishing by a person so it’s genuinely ours.
What AI does very well is speed up the early part of the work. Finding research papers, studies and guides, running the analysis, doing a detailed review and pulling it all together can get you 30%, 40%, sometimes 50% of the way, very quickly. Then the hard work starts: the structure, the narrative, the formatting, and the conclusion you actually want to put across. That’s the part a client is paying for, and it’s the part Epoch found the models are worst at.
It’s the same pattern I wrote about in AI paying back in speed before profit. The gain is real, but it’s at the front of the job.
Who checks the rest?
At Embryo we check AI work at every level, but it always goes up to a senior person for final sign-off. That isn’t caution for its own sake. If a model gets the easy two thirds right and the hard third wrong, the mistakes sit exactly where a junior reviewer is least likely to spot them.
That’s also where most AI projects go wrong. A team sees a tool produce something that looks finished, assumes the job is done, and nobody owns the last third. It’s a version of the pilot that never lands, except here the work does ship, just not to the standard anyone agreed to.
If nobody senior is signing off the last 35%, your clients will be the ones who find it.