THE NUMBER
360,000.
That's how many lawyer-hours JPMorgan was spending every year reviewing 12,000 commercial credit agreements - open each one, pull the key terms, flag anything odd. A document-reading agent now handles all of it. That's what this newsletter is about.
---
3 THINGS HAPPENING RIGHT NOW
The people using agents most are handing over the wheel
Anthropic released data from millions of Claude sessions showing how user behavior changes over time. People who've run 750 or more sessions now let the agent make decisions without intervening more than 40% of the time - double the rate for new users. The longest individual tasks - those in the top 0.1% by duration - now run over 45 minutes without interruption, up from under 25 minutes in October 2025 - trust that accrues from what agents have actually proven they can do.
H&M's AI recruiter cut hiring time by two-thirds
Maki, an AI recruiting startup, built an agent called Maria for H&M that interviews candidates around the clock in multiple languages. Maria runs skills assessments, holds 5-to-15-minute conversations, and filters applicants before a human recruiter gets involved. 80% of the recruiting process now runs without a recruiter present, and time-to-hire dropped by two-thirds. (Source: TechCrunch, January 2025)
Duolingo’s first 100 language courses took years to build. AI agents created 148 more in one.
Duolingo launched 148 new language courses in a single year - all generated by AI agents, not human contractors. The company’s first 100 courses took years to develop. (Source: TechCrunch, April 2025)
---
THE DEEP DIVE
The Contract No One Wanted to Read
JPMorgan's legal team had a problem that didn't feel like a technology problem. Every year, 12,000 commercial credit agreements came in. Each one required someone to open it, read through the standard terms, record the key figures, and flag any clause that strayed from the norm. Mostly they didn't stray. It was the same job 12,000 times.
The bank built an agent called COiN - Contract Intelligence. It reads each agreement, extracts the terms that matter, and surfaces the exceptions. What used to consume 360,000 lawyer-hours a year now runs in the background.
The real math: 360,000 hours returned to work that actually needs legal judgment. At roughly $150 an hour - a reasonable estimate for in-house legal staff costs at a large bank - that's $54 million a year in legal time redirected. The agent handles the pattern. The people handle the calls that don't fit the pattern.
The same job exists in every back office that isn't JPMorgan's. A bookkeeper reconciling the same bank statement format every month. If the task is repetitive and rule-based, an agent belongs at the front of the pile.
---
ONE THING TO TRY THIS WEEK
COiN processes 12,000 contracts without a lawyer in the room. Your version is probably sitting in your downloads folder right now: bank statements you meant to reconcile, invoices to check, intake forms you haven't gotten to.
Here is how to hand them all off at once.
Create a folder called to-process in any directory where you have Claude Code open. Drop in 3-5 documents of the same type - same format each time.
Then paste this, with one change:
```
Process every file in ./to-process/ - do not stop to ask me questions.
For each file: [replace this bracket with one sentence describing
what you want - "extract the total amount and due date" for invoices,
"find the closing balance and any flagged items" for statements].
Write all results to summary.md using each filename as a header.
Write DONE at the bottom when finished.
```
Minimize the window. Do something else. When you come back: DONE. Open summary.md.
Every file processed, results organized. You were not in the room.
Next batch: drop new documents in to-process, run the same command. Nothing to maintain. The work runs while you are elsewhere.
Stuck? Reply to this email. I'll help.
---
WHAT'S COMING
Next issue: the morning standup your team holds every day - same questions, same format, same people. An agent asks them overnight, collects the answers, and delivers a summary before anyone's first meeting. Here's how it works and what three teams stopped doing once they set it up.
Manu