Most people who use an AI tool at work have never read its usage policy. That is understandable, because they are long and written for lawyers. This month’s update from Anthropic is worth a few minutes of a manager’s time anyway.
Anthropic publishes the rules for how Claude may be used, and on 8 October it announced a new version. It takes effect on 12 November 2026. The company says most of the changes “clarify existing rules”, but one section speaks directly to anyone whose team uses AI to inform decisions about people.
What the policy says
The section covers what Anthropic calls high-risk uses. These are cases where Claude’s output can affect someone’s health, legal rights, finances, livelihood or access to essential services.
For those cases the policy asks for a qualified person in the loop. That person must have the authority to review and, if necessary, change Claude’s recommendations. The policy also asks for the individual affected to be told that AI was used.
Two things follow from the wording. The person has to be qualified, which means someone who understands the decision, not whoever was nearest. And the person has to have authority, which means they are allowed to say no, and the process lets them.
A separate part of the update deals with systems that connect Claude to physical equipment. That is a different situation and is not the subject here.
Whose rules these are
This is Anthropic setting terms for its own product. It is a supplier telling its customers what it expects, and it applies to people using Claude, not to everyone using AI.
Other suppliers have their own policies, and they differ. It is also not law. No regulator is enforcing this section, and Anthropic’s agreement with a customer is a commercial matter.
It matters anyway, for two reasons. First, it is a major supplier writing down a floor for consequential decisions, and these floors tend to spread. Second, it names the thing that is easy to skip: a person with real authority to disagree.
What counts as a decision about someone’s livelihood
The wording is wide. A manager reading it should think of hiring, promotion, pay, redundancy and disciplinary work. They should also think of credit, insurance, benefits, tenancy and anything where the output decides whether someone gets something they need.
Most teams do not use AI to make those decisions. Plenty use it near them. Someone pastes a stack of applications into a chatbot and asks for a shortlist of ten. Someone asks it to summarise a grievance file before a meeting.
In neither case does the AI decide. In both, it shapes what the human sees first, and what the human sees first tends to decide the outcome. This is where the idea of a qualified person with authority earns its place.
The failure the policy is guarding against
The risk is not that the AI is wrong. Everyone expects that and plenty of people check for it. The risk is that the person who is supposed to check has no practical way to disagree.
Consider a recruiter given a ranked list of forty candidates, a deadline of Friday, and a note that the tool has been approved. Overruling the ranking means explaining why to a manager, in writing, with no time. So the list is used as given.
On paper, a person reviewed it. In practice, nobody could change it. The policy’s phrase “authority to change” is aimed at exactly this gap between a review that exists and a review that works.
What to do before 12 November
List the consequential uses. Go through the ways your team uses AI and mark anything that touches hiring, pay, performance, customers’ money or access to a service. Include the informal ones. Most teams will find two or three.
Name the person. For each item, write down who can review and change the result. Check they understand the subject and have the time. If the honest answer is “whoever is on shift”, that is the finding.
Make overruling normal. Record when the person changed the AI’s answer, and say in the team that changing it is the job, not an exception. If nobody ever overrules, either the tool is perfect or the review is a formality, and the second is likelier.
Decide what you tell people. The policy expects the affected individual to be told AI was used. Settle your wording now, in a sentence a candidate or customer would understand, before someone asks.
Read the policy yourself. This post covers one section. Read the full text on Anthropic’s site, and check what your other suppliers say, since their terms may differ.
What a good version looks like
A good version is plain. The manager of a recruitment team tells everyone that the AI shortlist is a starting point. A named senior recruiter reads every rejected application in the top and bottom quarter, and may promote anyone. Candidates are told in the job advert that AI helps with first-pass sorting.
None of that needs new software or a committee. It needs one person with the time and the standing to disagree, and a sentence that says so. The policy asks for the same thing, in more formal words.
The short version
If AI helps shape a decision about someone’s job, money or access to something they need, a person should be able to change the result, and the affected individual should know a machine was involved.
Anthropic is not the only party that will ask for this, and it will not be the last. Having the person and the process in place is a small piece of work, and it is far easier to do before anyone asks than after.
Source: Anthropic, 2026 usage policy update, published 8 October 2026, effective 12 November 2026. Quoted passages are from that page.