A week in the network: an M&A attorney's review queue
Forty redlines, two second reviews and one long argument about indemnities. What a week of expert work actually looks like, from one of the attorneys in our network.
Written by Lena Sørensen · Research Lead, Evaluations
- Updated Nov 13, 2026
- 3 min read

People often ask what "expert work" in AI actually involves. The honest answer is that it looks a lot like ordinary professional work, done carefully and written down. Here is one week, from an M&A attorney who has been in the network for eight months. We've changed a few details to keep clients anonymous.
The setup
Our attorneys work in short, focused sessions from a queue. Each task arrives with a brief, the documents, a rubric and a deadline. Most tasks take between forty minutes and two hours. Nobody is asked to do more than they have said they want to do.
- Hours this week: 14, spread across four evenings and a Saturday morning.
- Tasks completed: 40 redlines and 2 second reviews.
- Rate: $235 an hour, paid the following Friday.
Monday and Tuesday: share purchase agreements
Most of the week went on share purchase agreements, the long contracts that govern buying a company. The model under test had produced redlines of each one, and the task was to grade them: what did it catch, what did it miss, and what did it get wrong?
The pattern became clear quickly. The model was good at the obvious: inconsistent defined terms, missing signature blocks, dates that didn't match. It was much weaker on the provisions that actually move money.
Wednesday: the indemnity argument
On Wednesday a second reviewer disagreed with one of her scores. The question was whether a cap on indemnity that applied to "losses" also applied to legal fees. She thought it did; the reviewer thought the contract was ambiguous and the model should have flagged it.
We argued about it for twenty minutes, then rewrote the rubric line. That's the job, honestly. Not being right, but being clear.
The rubric line was split in two, the whole batch was re-graded against the new version, and agreement on that line went from 62% to 91%.
Thursday and Saturday: writing it down
The last part of the week was rationales. Every score needs a short written explanation, in plain language, of what the redline should have said and why. These rationales are what make the data useful for training, not just for measurement.
- State what the provision does in one sentence.
- Say what the model's redline did or didn't do.
- Say what a competent associate would have written instead.
Why she does it
We asked why someone billing at a law firm would spend evenings grading a model's contracts. Her answer was practical: the work is flexible, it's paid weekly, and it's interesting in a way that much of her day job isn't. "I'm teaching the thing that will eventually do the boring half of my job," she said. "I'd rather it learned from someone careful."
If you're a practising professional and this sounds like work you'd enjoy, join the network. Applications start with a fifteen-minute work sample from your own field.



