OpenAI put out a number a few weeks ago that’s been making the rounds: their researchers are getting 3x a normal workday out of AI agents. I dug into where that number comes from. It’s not a one-to-one story for accounting and finance, even though most of the AI hype assumes it is.
Researchers are logging 3.14 agent-workdays for every 8-hour human shift. The typical person runs four agents at once, kicks them off, and comes back in the morning to review what got done. That’s where the “3x” comes from. It’s not that the AI got three times smarter. It’s a person supervising three shifts of machine work while only being awake for one.
It’s not free. The median researcher was spending around $700 a day on compute by August. At the 90th percentile, the heaviest users were burning through $7,000 a day. That annualizes to about $2.5 million a year, per person. It’s all operating expense too. A welding robot is a capital purchase. A factory runs it around the clock to earn back what it cost. An AI agent doesn’t work like that. You’re billed for exactly what you use, and nothing stops the number from climbing once the agents are running.
There’s also a catch OpenAI was upfront about. More than half of the complex tasks agents completed still needed a human to fix something. So you’re running three shifts of machine time, but you’re only banking about two shifts of usable work once you account for the rework. Still a real gain. Just not the clean 3x the headline implies.
Why this works for code and not everything else
Code is verifiable in a way most of what we do isn’t. A compiler checks whether the code will even run. A linter checks that it’s structured the way it’s supposed to be. Automated tests check that the pieces do what they’re supposed to do. All of that happens in seconds, with nobody awake to check it by hand. That’s the real engine behind “the AI worked all night.” It’s automatic, cheap grading, nothing more.
Accounting doesn’t have that same layer sitting underneath most of what we do. There’s no compiler for a materiality judgment. There’s no test suite for whether a going-concern call was the right one. Everyone says “AI won’t replace accountants” without saying why. This is why.
Two questions, not one
The way I’d think about applying this in your own practice: ask whether the task has a deterministic check, a rule that tells you the output is right or wrong. Then ask whether you actually have enough volume to make running agents around the clock worth it.
Multi-entity reconciliations, continuous three-way matching, tax calculations across time zones: those have both. A rule exists, and the volume is high enough that there’s basically always a queue of work sitting there. That’s a real candidate for letting agents run overnight.
A hundred reconciling items a day at a smaller business has the rule but not the volume. An agent clears that queue in a few minutes each morning. Running it all night doesn’t get you more output. It just runs up a bill against a queue that was empty most of the time.
And then there’s the harder case: high volume with no clean rule. I’ve worked with clients who got their AI-driven expense categorization up to 98% accurate, across companies with thousands of SKUs. Still had to spot-check it. Still had to review it. Ninety-eight percent sounds like a compiler-level result. It isn’t. There’s no deterministic check confirming any single categorization is correct. It’s a confidence score. High volume there means you sample more carefully. It doesn’t mean you stop watching.
One more thing worth asking before you get excited about any of this: a lot of accounting workflows already have automation running underneath them. If your three-way match is already hitting 85% straight-through today, the real question isn’t whether AI can help. It’s whether getting from 85% to 95% is worth what it costs to run, and how often it needs to run.
The takeaway
- The “3x productivity” number is a runtime number. The real delivered gain, once rework is counted, was closer to 2x.
- Agent spend behaves like a variable cost, not a one-time purchase. It climbs with usage, and nothing built in tells it to stop.
- Ask two questions before automating anything: is there a real check for correctness, and is there enough volume to justify running around the clock?
- High volume with no reliable check is the case to watch closest. Don’t hand it off.
- We’re going to see real benefits from AI in accounting. They’re just not going to look like what developers are getting, because most of what we do doesn’t run on the same kind of verification code does.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: The Compiler Advantage
Leave a Reply