AI Outperforms Human Accountants on Core Tasks | Aden Barton artwork

AI Outperforms Human Accountants on Core Tasks | Aden Barton

MTS

October 19, 2026

Mercor economics researcher Aden Barton discusses benchmark testing for AI capabilities, revealing that models now outperform human accountants on specialized tasks, while 90% of US workforce time remains completely unrepresented in existing benchmarks. Turn ideas into software people love.

0:00

Aden Barton

Models are really, really good today. They're faster, they're cheaper, and they're far more accurate than human accountants on these studies. Human accountants got around 40% on average, and the models today aced those same tasks. In early 2025, humans were outperforming models on these tasks. And then about a year and a half ago, around May, models reached parity with humans. They both scored around 40%.

And then today, models aced those tasks.

0:25

SPEAKER_2

Joined by Aden Barton, who is an eponymous researcher at Mercor. So Aden, welcome to MTS.

0:30

Aden Barton

Thank you so much.

0:31

SPEAKER_2

Thank you for being here. I think one of the big reasons we are excited to talk to you, I know you guys publish a ton of interesting research just about how AI is changing work, different types of benchmark. Can you tell me a little bit more about this research?

0:45

Aden Barton

Yeah, so stepping back, Mercor is kind of the data layer for the AI economy. So we have millions of experts whose job it is, is to kind of help us power benchmarks, enterprise evals and RL environments. And so as the models get better, it becomes much harder to fail them. And so you have to think very hard about what data can power model improvements, how to measure those improvements, and then how to map those to the real economy. And so that's kind of what the research team is all about. First, kind of post-training, thinking about how data can map to model uplift, and then benchmarking how you measure those improvements on tasks, on occupations, and then the economics team, which is focused on mapping those benchmarks to real economic outcomes. And I think what's unique about Mercor's research team is all three of these parts talk to each other, because you really can't understand benchmarks without thinking about how they are grounded in the real economy, for example.

1:37

SPEAKER_2

Yeah, so you have to have all these different teams that are actually working with each other. And then you guys published some interesting results as well. I know you guys shared some cool stuff around accountants and how AI already outperforms them. Any interesting results you can share with us?

1:53

Aden Barton

Yeah.

So the motivation of the study was, we see benchmark scores going up every day. And it's very hard to know what that actually means. What does a 15 percent on Apex mean? And so one way to answer that is to give those tasks to humans, because at least you know, well, maybe Claude outperforms a human or underperforms a human. And so we took tasks from Apex accounting, which is a benchmark we made with Ramp. And we gave those tasks, simplified versions of those tasks to real accountants, as you mentioned. And we saw kind of two big things. First, just in level terms, models are really, really good today. They're faster, they're cheaper, and they're far more accurate than human accountants on these studies. Human accountants got around 40% on average, and the models today aced those same tasks. And then the second headline finding was even more important than where models are today, how quickly they got there. So in early 2025, humans were outperforming models on these tasks. And then about a year and a half ago, around May, models reached parity with humans. Both scored around 40%.

And then today, models aced those tasks. And so just extrapolating that rate of improvement has kind of dramatic implications about productivity. But even if you were to freeze model progress today, we're already seeing that these models would provide huge productivity uplift to accountants.

3:06

SPEAKER_2

Yeah. So what does this mean? Who are the big people who read these research reports? Does it impact what they actually end up doing and stuff like that?

3:13

Aden Barton

It's a great question. So we got a lot of outreach from accountants, from accounting professors asking, what does this all mean?

And so one thing people said when they saw the study was, accountants are done, the job is totally over. And I think it's important to remember the task is not the job, right? Each job is kind of a bundle of tasks, some of which AI may automate, some of which AI may augment. And so I think our takeaway was that on these kind of well-specified, detail-oriented tasks, models are very good and will keep getting better, but that accounting will shift towards more relational, higher level things like working with clients, how to prioritize through time. And these are things that environments and benchmarks still struggle to capture.

3:54

SPEAKER_2

Why do they struggle to capture that?

Read the other 11 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000796993013