Anthropic runs its hardest work on AI drafts. Do you?

Anthropic said this week that Claude now writes more than 80% of the code merged into its own systems. The number isn’t the story. Where your first drafts come from is.

Read that stat again. The company building the frontier model already trusts that model to draft the work that keeps the model alive. Code is their hardest knowledge work. And they let the machine start it.

Most people will read this as a headline about coders losing jobs. That’s the lazy take. The real signal is about the blank page.

Here’s the thing nobody measures. In most shops, the slow part isn’t the final answer. It’s the start. The distributor’s team knows how to answer an RFP. They’ve answered a hundred. The pain is the empty document at 8am and the four hours it takes to get to a rough version worth editing. Kill that four hours and the whole shop moves faster.

The blank page is the bottleneck. The expert was never the problem.

That’s the flip the 80% number is pointing at. Anthropic didn’t replace its engineers. It moved them off the blank page and onto the merge. The machine drafts. The human decides what ships. Judgment stayed with the person, and the grind moved to the model, and the output went up because the expensive hours now go to the hard part instead of the cold start.

We saw this inside a Fortune 500 services engagement. The win wasn’t fewer people. It was the first draft landing in minutes, so the expert spent their hours on judgment instead of staring at a cursor. Same team. More work out the door. Higher quality, because they had time to think instead of time to type.

The move

Pick one workflow. One that’s heavy on documents and light on mystery.

A distributor’s RFP responses. An accounting firm’s first-pass memos. A law shop’s initial discovery summaries. Something you produce over and over, where the shape is known and only the details change.

Now measure one thing this quarter. What share of that workflow could an AI draft? Not finish. Draft.

Then run it small. One workflow. One person. One week. The human owns the merge. Nothing ships without them signing it. That’s the whole guardrail, and it’s the same guardrail Anthropic uses when it lets Claude write most of its code and keeps a person on the commit.

You’ll hear the objection before you finish the sentence. “Our team can’t. Our team won’t.” Fine. You’re not asking the team to change. You’re asking one person to try one thing for five days on work they already do. The objection needs a big rollout to be true. There is no big rollout. There’s one memo, drafted by a machine, edited by a human, shipped by Friday.

Start there and the fear shrinks to its actual size.

Track a number, not a feeling. Two numbers, really. How long the first draft used to take. How long it takes now. The gap is your answer. If the machine gets you to a workable draft in ten minutes instead of three hours, you just found where your quarter’s slack was hiding.

The point isn’t that AI is smart. The point is that the person who knows the answer shouldn’t be spending their morning building a scaffold to write it on.

Anthropic already ran this experiment on the hardest work they have. The result was 80%. Your work is easier than training a frontier model. So the question isn’t whether the draft can come from somewhere else.

It’s why yours still doesn’t.