Anthropic’s January 2026 Economic Index analyzed two million real conversations with Claude and landed on a conclusion that’s awkward for Anthropic itself: across most of the occupations they studied, the net effect of using the model is a skill loss, not a gain. The reason is simple. Claude covers tasks that require 14.4 years of average formal education, roughly an associate’s degree, against 13.2 years for the economy’s average task. The AI isn’t picking up the routine work. It’s picking up the most skilled parts of each job, which happen to be exactly the parts that give a person the practice to get better at their work.

This isn’t an outside warning or a competitor’s critique. It’s the lab behind Claude publishing the number itself. And it’s why any company designing an AI-assisted workflow needs to ask a different question than the usual one. Not “what task can I automate,” but “which part of this task does my team actually need to keep practicing.”

What Anthropic’s index found, in detail

The January report is the fourth edition of Anthropic’s Economic Index, and this time they broke it down by occupation with real examples. Technical writers lose tasks like “analyze developments in a specific field to determine the need for revisions” (18.7 years of education-equivalent) and keep tasks like “draw sketches to illustrate specified materials” (13.6 years). Travel agents lose complex itinerary planning and keep routine ticket purchasing and payment collection. Not every occupation loses. Property managers actually gain skill, because AI takes over routine bookkeeping and frees them for contract negotiation and tenant management, which are higher-level tasks.

The pattern holds across the report: when AI takes the easy part of a job, the person moves up. When it takes the hard part, the person stalls or slides back. And in most of the occupations Anthropic studied, it’s taking the hard part.

The empathy case shows the mechanism in miniature

If you want to see this pattern in one concrete example, AI-assisted communication is the clearest one available. A study published in Nature Human Behaviour compared how people value human empathy against AI-generated empathy and found something counterintuitive: in blind tests, people rate AI responses as equal or higher quality. But once they’re told which response came from where, the preference flips completely. People still prefer to receive empathy from a human, even knowing the AI response scored better on average.

A companion study in PNAS confirms the same effect from a different angle: a response that felt genuine and empathic stops feeling that way the instant the recipient discovers AI wrote it, even though not a single word of the text changed. And a third study from earlier this year found that AI-generated empathic responses score well on average but tend to follow repetitive, templated patterns that an attentive human notices if they’re paying close enough attention.

None of these three studies say AI is bad at writing empathic messages. They say something more specific: fully delegating the task carries a cost, both in how the other side receives it and in how genuine the communication actually ends up being.

The experiment that shows the way out

Here’s the part almost nobody cites when this topic comes up. A Stanford and Northwestern experiment tested something different from the usual setup. Instead of letting AI write the full message, researchers had participants draft their own response first, then use a language model to get specific feedback on how to communicate more empathically. Across nearly 34,000 messages and more than 2,900 conversations, they found this way of using AI actually improved people’s real empathic skill, measured by how they communicated afterward, without the assistant present.

The difference between that experiment and the typical “ask ChatGPT to write the email for me” scenario isn’t the model. It’s where in the workflow human judgment sits. In one case, the person never practices. They receive a finished message, approve it, send it. In the other, the person practices first, gets a specific correction, and rewrites using their own judgment. The skill outcome is the opposite, even though the underlying model is the same one. We’ve written about this exact distinction before: cognitive delegation is not cognitive surrender, and here it shows up again with the data to back it.

What this means for your workflow

This isn’t a reason to distrust AI in business communication. It’s a reason to design the workflow with this distinction in mind, especially in the areas where communication is the most sensitive part of the job: customer support, HR, sales, any interaction where tone matters as much as content.

In AI Maestro discovery engagements, this is literally one of the design decisions we work through with a client before touching any code. For every task being automated with AI, does the human end up as a passive approver of an already-finished message, or as the author who receives feedback and decides? The first option is faster to build. The second is the one that keeps the team capable six months later, when the tool changes, breaks, or the client needs something the model couldn’t actually solve. It’s the same logic behind why adoption is not the same as transformation: automating a task without redesigning the workflow around it rarely produces the result a company expected.

Anthropic’s data doesn’t say you have to choose between speed and skill. It says that if you don’t design that choice on purpose, you’ll lose it by default. And in communication with customers or employees, that’s one of the worst ways to lose it.