Are you still necessary, when AI does most of the writing?
Your Essential Report from the Frontlines of Applied AI
At Gunma University Hospital in Japan, doctors no longer write discharge summaries from scratch.
This month the hospital started using AI to draft them from the electronic medical records, along with referral letters and medical certificates. The hospital says every draft is checked and corrected by a doctor before it is used.
The head of its systems integration centre told the system’s maker that the drafts sometimes catch details he had missed himself.
So what does a doctor’s job become?
Checking and correcting, instead of writing. And that is where many of our jobs are heading.
It’s clear to us that:
- AI drafts, and you check.
- Without your judgement, AI output is unverified raw material.
- Your name goes on what you sign.
Checking is still faster than writing. At Gunma, the hospital wants the time it saves to go back to patients.
So this week is about being the check. There’s a prompt below from AI+ Librarians. It has AI suggest subject headings and rate how sure it is of each, so you can verify every one. And there’s a case study on Sullivan & Cromwell. Its AI contract review hands every finding back to a lawyer to review, adapt or reject.
Plus the week’s news in applied AI — Uber gives 30,000 staff an AI agent, a US appeals court reassigns a case after a judge’s staff used AI to draft an order full of errors, Bojangles’ AI drive-thru takes a million orders a week, and almost 80% of Linklaters partners now use Legora every month.
AI drafts. The competent person reviews, approves and issues.
On a building site, a method statement or a risk assessment is a legal safety document. We give construction teams one rule for it: AI drafts only.
A qualified competent person checks the draft against the planned work, confirms it and signs it off. Work starts only then.
However well written the draft, it is not valid until then. The duty cannot be handed to AI.
The same rule holds for anything that goes out under your name.
AI drafts. You are the competent person.
- Business operations
Uber gives more than 30,000 employees an AI agent to research, prepare briefs, build presentations and analyse data
After a pilot with 1,000 staff, Uber is rolling out OpenAI’s ChatGPT Work across the company. It connects to Google Drive, Gmail, Google Calendar and Slack, and comes with an internal library of AI skills, approved integrations and training. The companies say they will share results as the rollout goes on.
OpenAI, 22 October 2026
- Courts
A US appeals court reassigns a case after a judge’s staff used AI to draft an order full of errors
The order named people who were not parties to the case, misquoted state law and cited a case that does not exist. In its ruling on 22 October, the Fifth Circuit called the proceedings “a botch job, to say the least”. It gave the case, over a Mississippi education law, to a new judge.
Mississippi Today, 23 October 2026
- Retail
Bojangles’ AI drive-thru takes over a million orders a week across 500 restaurants
The chain says Hi Auto’s voice AI gets 98% of orders right, and it is adding the system to 100 more drive-thrus by mid-2027. Each order is now paired with a picture of the car that placed it, so staff can check it when two lanes merge. The accuracy figure is the company’s own.
Restaurant Dive, 22 October 2026
- Legal
Almost 80% of Linklaters partners now use the legal AI platform Legora every month
Monthly active partners rose from 65% to almost 80%, and 97% of weekly users were working in its AI agent within a month. The firm credits peer teaching: twelve partners were trained first, then taught the others. The figures are the firm’s and Legora’s own.
Legora, 22 October 2026
Paste in the source material only, with nothing confidential. The prompt asks AI for five to eight subject headings, each with how sure it is and why. It also tells AI you will verify each one against the authority file, so a heading it isn’t sure of is marked low. Use the same pattern with any AI suggestion you have a source of truth for.
Suggest 5–8 candidate subject headings for the work described below, drawing on Library of Congress Subject Headings (LCSH) — I will verify each against the authority file before use. For each: - The exact heading as you propose it. - Your confidence that the heading is an authorised form in LCSH (high / medium / low). - A one-line justification from the source material. If you are uncertain whether a proposed heading is an authorised LCSH form, mark it "low" and suggest what the authorised equivalent might be. Source material: [paste]
Prompt 4.2 from AI+ Librarians, word for word.
Build the human check into the AI tool itself.
Sullivan & Cromwell partner John Savva built Agreement Analyzer to review draft agreements, and the firm then developed it with OpenAI.
A lawyer adds the draft, their chosen precedents and a playbook of firm and client positions. The tool reviews the agreement several times from different angles.
It returns a memo and proposed tracked changes “for a lawyer to review, adapt or reject”.
Savva’s test for trusting it is plain: “If an AI tool makes a claim about market practice, you need to be able to see the source of that information.”
Whatever AI drafts for you, ask for the source behind every claim, and keep the final say.
