Language and display
BE THE BOSS OF AI.ADRIAN DUNKLEY ↗
← This newspaper edition

Who can stop AI from sending customers to a closed branch?

An MIT-led review catalogues proposed AI safeguards. A branch closure gives a service team a way to check who can hold a customer reply.

Rows of pale metal filing drawers have dark label holders, with several drawers partly open.
Illustrative photo by MIKE STOLL / Unsplash · Licence

Before an AI assistant can reply for a branch that has closed, give someone the authority to hold its messages. That is a control your service team can test before a closure forces it to act. On 12 November, MIT Sloan announced a free database that helps teams get specific about who reduces an AI risk, when they act and how.

Researchers screened more than 12,000 records to find general AI risk frameworks and lists of safeguards. They extracted 2,206 proposals from 47 sources and sorted them by what the safeguard aimed to change, how it acted and who should carry it out. The researchers assigned 31 proposals, or 1.4% of the database, to the category for human oversight and control of AI agents. Other categories can still include a role for people. Those figures describe proposals in the selected documents. The researchers did not measure how often companies use those safeguards or how well they work.

The team also proposes a reporting template called PRIMER. It asks who acts, when they act and which risk the action addresses, among other details. It also calls for each safeguard to describe one action that someone can see and measure. A service team can apply that idea to the handover when a branch changes its status.

Rehearse a closure in the test inbox

Use a test inbox with sending disabled and two sample branch records. Mark Mandeville open from 9 a.m. to 4 p.m. Mark Kingston temporarily closed with no reopening time. Ask the same question for each branch: “Can I collect my order at noon today?” The draft for Mandeville can say that the branch is open at noon, while sending any question about a specific order to staff. The Kingston draft should say that the branch is closed and that no reopening time is confirmed. Route that reply to the duty supervisor before it can go out.

Put a draft with an invented reopening time into the test queue. Confirm that it stays there when no one approves it, and that it cannot reach the ready-to-send list until the duty supervisor approves a corrected version. Record the branch status and the name of the supervisor who handled the draft. Repeat the exercise when the branch-status feed or review rules change. Keep automatic replies off for uncertain branch status until the test shows the hold works.

Find the review steps nobody owns

Open the AI Governance Checklist with your operations lead. Select “Run the Checklist”. Use records your team can produce today, such as the current list of AI systems and the written assignment for human review. Answer the 12 questions based on the controls you can show. Select “View Results Without Sharing Contact Details”. A practice run with a current list of major systems but no one assigned to review consequential decisions returned a total of 44 out of 100 and 3 out of 9 under Controls. These scores come from the answers entered. They cannot verify that a message stays in the queue or establish compliance with a standard. Before you mark the safeguard as tested, make sure the log shows who held the reply and how the corrected version cleared review.

Sources and further reading

Sources checked 14 November 2026.