At Anthropic, Kristen Swanson’s role is to help people become better at using AI.
Often that means they know when to do the work themselves.
“My job is to teach people how to use AI, and I probably spend at least half my time telling people when not to use AI,” says Swanson, who leads research and learning on AI fluency at the company.
The smartest AI users, she said, aren’t necessarily the ones who rush into every task. What sets them apart is their judgment about what to delegate.
That will be one increasingly important skills in the workplace as AI makes its way into more of what humans do. The challenge is not just learning to use the technology. It’s knowing when it can save time, when it could create more workand when a task is better left to one person.
Knowing what not to delegate
In some cases, Swanson says, evaluating what AI produces takes more effort than doing the task yourself. It’s what she calls a “distinction tax.”
If people hand the right things over to AI, the time spent reviewing the results could be worth it, Swanson says. But for tasks that people are already skilled at, evaluating the work of AI can eat up the time they were hoping to save.
The right way assess what AI producessaid Swanson, “it’s the highest order thinking you can do. It’s actually exhausting.”
That’s why it might make sense to ask AI for help with tedious data analysis, she said. But if you write a social media post about something you know about, you can do more work and “work through all this stuff.”
The distinction tax isn’t the only reason to keep some work away from AI. The limitations of technology can also make delegation risky.
“If you ask AI about a really niche research article or researcher, and the company didn’t have a lot of that information in its training, then it could be hallucinating,” Swanson says.
Bosses are also trying to set clearer boundaries around when employees should turn to AI. While many companies are pushing employees to embrace the technology, some CEOs have warned against allowing it replaces employees’ own thinking.
Scott Stevenson, head of Spellbook, which is developing an AI tool for drafting and analyzing contracts, recently told employees that he “D+ version” of their ideas instead of proposals polished by AI, because he wanted to see the thinking behind them and avoid bloated memos created with the help of AI.
Avoiding ‘capability overhang’
Another challenge is that the line between what people should delegate and what they should do themselves may shift as AI improves, says Swanson, who oversees the Claude Academy, which offers educational resources around AI. A task that humans can perform themselves now might be better suited to AI in six months.
That’s why more advanced AI users routinely re-examine what the technology can do, Swanson said.
This is a way to avoid what is sometimes called “capability overhang,” where users’ understanding of what AI can do becomes frozen in time based on previous experiences with it. That could lead people to stick with familiar applications, even as AI models themselves become more capable.
That’s one reason, she says, that Claude Academy asks users to write down some of their most difficult tasks. As new models arrive, they can try those tasks again and see how well AI does.
The latest models may perform better at some tasks, but not at others, Swanson said. Retesting a task can give users “a better idea of what’s possible.”
That means continuing to experiment, she said.
“Being able to say, ‘I tried this, and it didn’t work, and I’m going to try it a different way next time,’ is much more important than, ‘Did I enable this feature? Did I ask this way?'” Swanson said.
Ultimately, becoming fluent with AI means thinking enough about when and how to use it that those decisions become second nature, rather than simply trying to use as many cues, functions, or AI tools as possible.
“More AI is not always better,” Swanson said. “And more AI is not necessarily fluid AI.”
