AI is starting to sign in, change files and speed up its own progress, so companies need to decide who approves what it does.
The AI Transformation Brief
AI is moving from answering questions to acting inside the places where work happens. Sierra bought Anon so its AI can sign in to websites on a customer's behalf. Sierra Dropbox now lets AI move, rename and edit files after a person approves the plan. Dropbox An outside reviewer found Anthropic improving its AI about twice as fast as a year ago. METR The faster the tools change, the more each company needs its own record of what AI did and who said yes.
Sierra announced on November 5 that it acquired Anon, a startup whose technology lets AI sign in to websites, handle security codes, fill out forms and work through portals built for people, even where no direct software connection exists. Sierra Anon raised more than $13 million and served hundreds of customers. Sierra says the goal is finishing longer jobs, such as originating a loan or getting prior authorization for a medical procedure. No price was disclosed. AI Weekly
Once AI can sign in wherever a person can, every login becomes a door it can open. The open questions are basic: whose permission the AI uses, how a website can tell, and what record a signed-in session leaves. Sierra's post answers none of them. Sanctity AI Before letting any outside AI sign in for your staff or customers, require separate sign-ins for the AI, limits on what it can do, and a record of every session.
Sierra answers for how its AI signs in. Your company answers for every account it opens. If the AI uses an employee's password, no one can later tell whether a person or the AI acted.
Dropbox announced on November 5 that its AI can now find, summarize, edit, move and rename files from a single request, and the new version has reached more than 70,000 teams. Dropbox It also agreed to buy Tigris Data for $350 million to sell storage to companies building AI software, its largest purchase to date. TechCrunch Customers can switch between AI providers during a session and work on their files in ChatGPT and Claude. Dropbox
The safeguards are the story. Dropbox shows a plan for approval before making changes, stays within existing permissions, lets customers undo changes and keeps a record of what the AI did. Dropbox That is now the minimum any company should expect from software that acts on its records. Ask every supplier for the same four protections in writing, and test the undo before trusting it with client files.
METR, an independent research group, spent 11 weeks reviewing Anthropic's internal data and interviewing staff, and published its findings on November 5. It concluded Anthropic is likely improving its AI on well-defined tasks at about twice the pace of March 2024 to October 2025, with the true figure possibly between 1.3 and three times. Most of the speedup likely comes from Anthropic using its own AI to do research. METR
The bigger issue is what outsiders cannot see. Anthropic redacted a majority of the evidence METR found most informative, so the public has to rely on the reviewer's judgment. METR For buyers, the lesson is speed: an AI tool approved six months ago may now behave differently. Retest your most important AI tasks every quarter, and ask each supplier what outside review it allows and what it withholds.
Palmer warns against handing AI tools your banking password. A typed instruction such as only review my spending does not create a limit the bank enforces. He points to connections where you sign in with the bank and grant read-only access. Shelly Palmer The same test applies to Sierra's sign-in AI: the system itself must enforce the limits.
AI is starting to act inside your accounts, files and customer portals.
This quarter, name an owner for every place AI can sign in or change records, and require a person to approve high-risk actions.
Suppliers now compete on plans you approve, undo buttons and records of every action.
Companies that give AI its own sign-ins will move faster; those relying on shared passwords will stall.
Map every task where AI signs in or edits files.
Give the AI its own sign-in, set what it may change, and review its action record weekly, starting with client and finance files.
Ask management which AI suppliers allow outside review, what they withhold, and how often approved AI tools are retested.
Tools that change every quarter need oversight that checks every quarter.
The shift this edition is AI acting where people work: signing in, editing files and improving itself faster. The assumption it breaks is that approving a tool once settles the question. The decision it forces is to give AI its own sign-ins, require approval before changes and retest important tools every quarter.
If AI acted inside your bank account tomorrow, could you prove whether it was the AI or your employee?
Make AI accountable. Measure the results. Redesign the business.