The future of work · data, not hype

Automation vs. augmentation: what people really use AI for

The headline fear is that AI does your job instead of you. The data on how people actually use these tools points somewhere else entirely — toward AI doing your job with you. That difference isn't wishful thinking. It's measurable, and it changes what you should be good at.

"Will AI replace me?" is the question everyone asks. But it smuggles in an assumption — that AI is a substitute, a thing that steps into your role and pushes you out. There's a second possibility the word "replace" hides: that AI is a complement, a tool that makes the person using it faster and better.

Economists have names for these two futures. Automation is AI doing a task instead of a human. Augmentation is a human and AI doing it together. They lead to very different worlds — and we no longer have to guess which one we're in. We can look.

What the usage data actually shows

Anthropic did something unusually concrete: it analysed millions of anonymized conversations with its Claude models and sorted them into how people were interacting — not what they were talking about, but the shape of the collaboration. It found five patterns, which fall into two camps:

• Automation — directive (hand off a task, take the result) and feedback loop (the AI works, self-corrects, delivers).
• Augmentation — learning (ask for explanations), task iteration (build something together, back and forth), and validation (get your own work checked).

The dominant mode, in other words, isn't "do it for me." It's "help me do it": brainstorm with me, explain this, check my draft, let's iterate. Even on the actual frontier of the technology, most people reach for AI as a collaborator, not a replacement.

There's an important wrinkle. When AI is embedded inside business software through an API — quietly powering a feature rather than being chatted with — automation dominates, at roughly 77% of that traffic. So the honest summary is: when humans choose how to use AI directly, they mostly augment; when companies wire it into a pipeline, they mostly automate. The mix of the two is a choice, not a destiny.

What augmentation does to output

If AI were just a novelty, none of this would matter. But when researchers measure augmented work against unaided work under controlled conditions, the gaps are large and consistent.

These aren't AI working alone. They're people working with AI, and the combined output beats the person by margins that would be extraordinary in any normal productivity story. Which is exactly what "augmentation" is supposed to look like: the tool doesn't replace the worker — it raises the ceiling on what the worker can produce.

The part that surprises people: it levels

Here's where the data breaks the intuition hardest. If AI mostly helps skilled experts, it widens the gap between the strong and the weak. It does the opposite.

In the largest real-world study of its kind, MIT and Stanford researchers watched 5,179 customer-support agents get access to an AI assistant across three million conversations.

The AI worked by capturing the tacit know-how of the best agents and putting it in front of everyone else. Newcomers climbed the experience curve faster; the people who were already excellent gained little, because they already knew what the model was suggesting. Augmentation, in this study, was a great equalizer — it lifted the floor, not the ceiling.

The clearest pattern in the productivity research isn't machines replacing people. It's AI compressing the gap between novices and experts — which reshapes what "experience" is even worth.

The catch: the frontier is jagged

None of this means AI is a free lunch, and the same BCG study is where the story turns sharp. The researchers gave consultants a second kind of task — one that looked similar but sat just outside what the model could reliably do. On those tasks, the people with AI did 19 percentage points worse than the people without it.

They coined a phrase for this: the jagged technological frontier. AI is brilliant at some tasks and confidently wrong at others, and the two can look almost identical from the outside. The consultants who lost ground weren't lazy. They trusted a fluent, plausible answer that happened to be wrong — because it takes real judgment to tell the difference.

So what's the human actually doing?

Put the findings together and a job description emerges — not for a job that AI erases, but for the human half of the augmented pair. In every study above, the person is still doing the parts that don't automate: deciding what's worth doing, directing the tool toward it, and judging whether what comes back is any good.

Those are not typing skills. They're the ability to frame a problem, to think critically about a confident answer, to know your field well enough to catch a subtle error, and to own the decision at the end. As routine execution gets cheaper, these become the scarce, valuable part of almost every knowledge job.

Augmentation doesn't remove the human from the loop. It moves the human up the loop — from doing the task to directing and judging it.

The bottom line

The replacement story isn't wrong because AI is weak. It's incomplete because it counts only one of the two things AI can do. The measured reality, right now, is that most people use these tools to amplify their own work; that the amplification is large; that it helps the less experienced most; and that it pays off only for people with the judgment to steer it and catch its mistakes.

Which points at a genuinely useful question — not "will a machine replace me?" but "am I good at the things that make augmentation work?" Directing, questioning, judging, catching the confident error. Those are human capabilities, and unlike the technology, they're ones you can measure and build.