Here’s the number from Anthropic’s robot report that I keep coming back to: about 80% of job tasks, by working time, are exposed to either robots or LLMs.
The obvious question is what’s in the other 20%. That’s the part of work neither kind of machine can do today. If you want to know what your job looks like in ten years, it’s probably more of that.
Two kinds of exposure that barely overlap
Anthropic’s report, What work can robots do? , compares two measures. One is LLM exposure: tasks where an LLM could cut the time needed in half. The other adds robots on top.
- About half of work is exposed to LLMs alone.
- Adding robots raises that to 81%.
- Transportation and moving goes from under 15% exposed to LLMs to about 90% with robots.
- Office and admin support, mostly computer work plus light physical tasks, goes to nearly 100%.
The report puts it simply: “Robots do work where LLMs cannot.” The two technologies hit different jobs. LLMs hit the desk. Robots hit the dock, the road and the line.
I spend my days in AI infrastructure. Most of the conversation in my world is about the first half of that picture. This report is a good reminder that the second half exists, and that it lands on different people.
What’s left
Anthropic grouped the tasks that neither robots nor LLMs can do today and described them. They “tend to be hands-on and face-to-face, and are sometimes regulated.”
Their example: “Administer medications to patients and monitor patients for reactions or side effects.”
That one task has all three. Your hands are on a person. You’re reading the person, not a screen. And there are rules about who’s allowed to do it.
The occupation groups with the lowest exposure fit the same pattern:
- Personal care and service, about 40% exposed
- Installation and repair
- Healthcare support
- Community and social service
And here’s the line I’d underline: no occupation group overall is less than 40% exposed. So the last 20% isn’t a list of safe jobs. It’s a list of safe tasks inside jobs.
What blocks the last 20 percent
The report breaks down what would stop robots from doing physical tasks, even if they could. Four kinds of barriers:
- Capabilities. About 70% of physical tasks. Mostly manipulation, which I wrote about in the hard part is the hands .
- Cost. For nearly all tasks, costs would need to fall. Robots are cost-competitive on just 0.3% today.
- Regulation. Today’s rules wouldn’t allow robots on 14% of physical tasks, mostly in healthcare, protective service and education.
- Human preferences. About a quarter of tasks. Some because people wouldn’t trust a robot, like “Dress children and change diapers.” Some because people want a person, like greeting restaurant guests.
That last category is the one engineers tend to skip. It isn’t a technical problem. Nobody is going to fix it with a better gripper.
The jobs argument, both sides
This is where the conversation usually gets loud. I’d rather stay with what people actually said and measured.
On the worried side, Anthropic’s own backtest is real. Over 50 years, jobs more exposed to robots saw bigger declines in wages and employment. Packers and packagers, the biggest occupation where robots already win on cost, have seen employment fall 22% since 2015, and BLS projects that job will shed the 11th most jobs of any occupation by 2035.
On the other side, at the SCSP AI+ Robotics summit in Washington on October 14, Kai Williams reported that the president of the Association for Advancing Automation, Jeff Burnstein, argued robots add jobs: “We have charts that show whenever robot sales have gone up in America, unemployment has gone down.” Speakers pointed to a projection of up to 1.9 million unfilled manufacturing jobs by 2033. Williams also included a fair counterpoint from economist Susan Helper, who compared that shortage to wanting a new Mercedes for under $10,000 and asked whether that’s a car shortage or a willingness-to-pay shortage.
And there’s the work robots create. FANUC America’s CEO Mike Cicco said at the same event: “You can go to high school, you can get a two-year community college degree and you can go program a robot and be making like $75,000 a year.” Anthropic’s cost estimates back up the idea that people stay in the loop. They often include human supervision, exception handling and repair.
I don’t think either side gets to declare victory from this data. What I’d take from it is that the change is uneven, it’s slow at the system level, and it lands hardest on specific jobs. Those are the people who deserve the attention, not the headline.
What the last 20 percent looks like
Here’s my read, as someone who has spent years in jobs that are mostly talking to people about machines.
The last 20% is the part of work where you’re accountable to another person in the room. The nurse watching for a reaction. The mechanic who has to explain why the repair costs what it costs. The engineer in front of a customer whose system is down.
Anthropic ends the report with a question I like: whether work itself becomes “more social and interpersonal as robots and AI advance.”
I think it does. The machines get the parts of the job that can be specified. People get the parts that have to be trusted.
My barber agrees.