Anthropic will let AI approve a claim without a human, but not deny one

The change

Anthropic published a new version of its Usage Policy on October 8, 2026, and it takes effect on November 12. The policy applies to anyone who submits inputs to Anthropic's products: people using the Claude apps, developers and businesses using the API, and "the end users of products or services integrating Claude." The previous version, effective September 15, 2025, is still online in Anthropic's archive, so the two texts can be compared line by line.

The policy has three parts: Universal Usage Standards, which list what is banned for everyone, High-risk Use Case Requirements, and Additional Use Case Guidelines. This article is mostly about the second part, the rules for using Claude in decisions about people's health, money, work, housing, and legal rights.

In its announcement of the update, Anthropic writes: "Most of the updates in the latest version are intended to clarify existing rules." On high-risk uses, it says the two core requirements, a qualified human in the loop and telling the affected person that AI was used, "haven’t changed with the updated version of our Usage Policy." That is true of the two requirements. The text around them changed, and one of the changes decides when the human review is needed at all.

What changed

• Approvals without a reviewer. Where the law allows, a qualified person no longer has to review an AI recommendation before it is acted on when the outcome is wholly favorable, in four listed cases such as paying an insurance claim.

• Eleven high-risk areas instead of seven. The list now names credit, housing, healthcare access, public benefits and services, and legal status and adjudication as separate areas, and media and journalism are no longer on it.

• Not only consumer products. The old section applied to "specific consumer-facing use cases." The new one applies "to anyone who uses our products or services to make a High-risk AI Recommendation."

• Disclosure to the person being decided about. Someone who is the subject of a decision based on a high-risk recommendation must be "clearly told that AI was used to produce it." The old rule covered model outputs "presented directly to individuals or consumers."

• Who is responsible. The old text said "You or your organization are responsible." The new text says "The qualified person remains responsible for the accuracy and appropriateness of what is delivered or decided."

• Physical actions. New requirements apply when Claude's outputs drive hardware without human approval, from vehicles and robots to devices that act on the human body.

• Chatbot disclosure. Consumer-facing chatbots and AI agents must still disclose that they are AI, and the disclosure can now be "in the product interface" instead of only at the beginning of each chat session.

• Agents. The policy now says explicitly that users are "responsible for ensuring that agents they build or deploy, including actions those agents take through tools, browsers, or connected systems, comply with this Usage Policy." The old version said only that agentic use cases "must still comply with the Usage Policy."

• Abuse toward the models. A new line bans "sustained and needless abusive or cruel behavior toward our models." Anthropic's announcement adds that "Claude’s ability to end these interactions will remain the primary enforcement mechanism."

What Anthropic's announcement leaves out

Anthropic's post goes through the update section by section and covers the high-risk part under the heading "Clearer requirements for high-risk use cases." It does not mention the exception for favorable decisions. It does not say that the requirements now reach uses that are not consumer-facing, and it does not mention that the sentence on responsibility now names the qualified person. All three are in the policy text, but you can tell they are new only by reading the new version next to the old one.

Yes without a reviewer, no with one

The new "Favorable Decisions" section reads: "Where permitted by law, human-in-the-loop review is not required before acting on a High-risk AI Recommendation wholly favorable to the individual in the following circumstances:"

• "Paying an insurance claim."

• "Approving coverage or prior authorization for care a clinician has ordered."

• "Granting eligibility for a healthcare service or program."

• "Granting or continuing a public benefit or service."

The opposite outcomes get no such exception. Denying a claim, refusing coverage, or ending a benefit is still a high-risk recommendation that a qualified person has to review before it is acted on. The section also closes the obvious gap: "A partial approval, reduced amount, or approval with conditions is not considered “wholly favorable.”" A claim paid at a reduced amount still needs a person.

The exception covers only the review and says nothing that lifts the policy's disclosure rule. The section adds: "Additional legal obligations such as disclosure may still apply."

The September 2025 version had no exception of this kind. Its rule was that for advice, recommendations, or "subjective decision-making directly affecting individuals or consumers, a qualified professional in that field must review the content or decision prior to dissemination or finalization," and insurance "claims processing, or coverage decisions" were on its list.

The high-risk list grew from seven areas to eleven

The old list had seven entries: legal; healthcare; insurance; finance; employment and housing; academic testing, accreditation and admissions; and media or professional journalistic content. The new one has eleven, and each comes with a specific list of the decisions it covers. Some examples:

• Employment. "screening, ranking, advancing, or rejecting candidates; hiring; promoting or demoting; setting or changing shifts, hours, or pay; evaluating a worker’s performance or behavior, including through disciplining, suspending, or terminating a worker." The old entry spoke of "resume screening, hiring tools, or other employment determinations."

• Credit. Approving, denying, or setting the amount, rate, or limit of a loan, mortgage, or credit line, or producing a credit score used for those decisions. Loan approvals and creditworthiness used to sit under finance.

• Public benefits and services. "determining eligibility for, or the amount of, government benefits, assistance, or essential services, including suspension and termination." The old list did not mention benefits.

• Legal status and adjudication. "deciding or recommending the outcome of a civil, administrative, or arbitral matter; determining immigration, asylum, or citizenship status."

• Healthcare access. "setting triage or waitlist priority; determining eligibility for, denying, or discontinuing a treatment, service, or care program."

Media and journalism are not on the new list. Under the old version, using Claude to "automatically generate content and publish it for external consumption" was a high-risk use. The policy's rules against disinformation and deceptive content still apply, but publishing AI-generated content is no longer a high-risk use on its own.

The rewrite also lists what is not covered. General information "without applying it to the individual's circumstances," wellness content, "Internal drafting, research, summarization, or analysis that is not the ultimate High-risk AI Recommendation delivered," and "Business operations that do not advise or decide about a specific individual" fall outside the requirements. A new note says the requirements "do not replace local legal requirements and restrictions."

Internal tools count, and so does the person being decided about

The old section introduced itself as applying to "specific consumer-facing use cases that pose an elevated risk of harm." The new one says: "These requirements apply to anyone who uses our products or services to make a High-risk AI Recommendation or control equipment capable of High-risk Physical Actions." The consumer-facing condition is gone. A company that uses Claude to rank its own job applicants, set its employees' shifts, or recommend a performance rating falls under the employment entry, even if no customer ever sees the tool.

The disclosure rule moved in the same direction. The old rule applied "If model outputs are presented directly to individuals or consumers." The new one reads: "An individual who receives advice or who is the subject of a decision based on a High-risk AI Recommendation must be clearly told that AI was used to produce it." A job applicant who never sees the ranking that screened them out is still the subject of that decision. The disclosure does not have to name the model: "Users do not need to identify Anthropic, Claude, or the model used."

The reviewer is named as the one responsible

The human in the loop is still required, and the requirement is more specific. A qualified person must "meaningfully review" the recommendation "with the authority to change or modify it." A qualified person is "someone with the training or experience to evaluate the output in that field, and who holds a license where applicable law requires one for the underlying activity."

The last sentence of the requirement has a new subject. The old version said: "You or your organization are responsible for the accuracy and appropriateness of that information." The new one says: "The qualified person remains responsible for the accuracy and appropriateness of what is delivered or decided." The introduction to the section repeats the point: Anthropic requires "that a qualified person remain accountable for what is delivered or decided."

New rules for Claude connected to hardware

The physical actions part has no counterpart in the old version. It applies when Claude's outputs "are acted upon by hardware without human approval" and the equipment can move through shared space, apply force that could cause injury, control hazardous energy or materials, act on the human body, control safety systems, or run industrial processes. In those cases, "there must be a qualified individual who is able to observe the equipment’s operation and stop it at any time," and the equipment must stop or hold a safe state "when that individual intervenes or when connection to our services is lost." Limits such as speed, force, temperature, and dose must be enforced by the equipment or a controller "independent of model output."

What did not change

As Anthropic says, the two core requirements for high-risk uses are still review by a qualified person and telling the affected person that AI was involved. Using Claude to make criminal justice decisions is still prohibited, and the ban now covers making or suggesting decisions in any law enforcement or criminal justice process, where the old version banned only law enforcement uses that violate civil liberties. Consumer-facing chatbots and AI agents must still disclose that they are AI.

Why this matters

Products built on Claude can make or shape decisions about people: applicant tracking, HR and workforce tools, lending and insurance platforms, patient intake, benefits casework. When the Usage Policy changes, the rules for every one of those products change on the same date, whether or not the company behind it read the update. The policy also reaches further than companies that contract with Anthropic directly, and the list of vendors using Claude keeps growing: at least 30 vendors added Anthropic as a subprocessor between September 1 and October 8.

Three changes have practical weight for a company that builds on Claude. The favorable decisions exception changes where a person sits in an automated workflow: in the four listed cases, approvals can go through without prior review, while denials, partial approvals, and reductions have to stop at a qualified person. The end of the consumer-facing condition brings internal tools into scope, including the ones a company uses on its own staff. And the disclosure rule now reaches people who never see the output. Changes like these are a large part of the hidden risk of vendor legal changes, because nothing in the product itself looks different on November 12.

The agents sentence fits a pattern across vendors this fall. HubSpot now makes customers answer for what its AI agents do, OpenAI's terms for dots make you responsible for what they do, and Stripe's new terms make AI agent actions legally binding on you.

Potential impact

For a company that uses Claude in decisions about people, the update raises six practical questions:

• Which of your workflows use Claude to recommend a decision about a specific person's job, money, health, housing, or benefits?

• In those workflows, which outcomes are wholly favorable, and which denials, partial approvals, or reductions need a qualified person before they take effect?

• Does the reviewer have the training, the license where one is required, and the authority to change the outcome?

• Are the people a decision is about told that AI was involved, including the ones who never see the output?

• Do internal tools, such as hiring, scheduling, or performance tools, fall under the employment entry now that the consumer-facing condition is gone?

• If Claude drives any hardware, can a person stop it at any time, and does it stop safely when the connection drops?

How Venpo detected it

Venpo monitors Anthropic's legal documents, including the Usage Policy, as part of continuous vendor risk monitoring. It recorded the new version on October 8, the day Anthropic published it, five weeks before the effective date. Every quote in this article was checked against the live policy and the archived September 2025 version. The same monitoring recorded an earlier Anthropic change, under which Anthropic can override zero-data-retention rules for Covered Models.

Business outcome

Teams that track Anthropic had the new text five weeks before it takes effect, including the parts the announcement does not mention. That is enough time to map which decisions need a reviewer, which can go through, and where a disclosure is missing. A team that read only the announcement learned that the two core high-risk requirements "haven’t changed" and nothing about the exception for favorable decisions.

Key takeaway

From November 12, Anthropic's Usage Policy lets AI approve an insurance claim, coverage, healthcare eligibility, or a public benefit without prior human review where the law allows, while denials and partial approvals still need a qualified person, who is now named as the one responsible. The requirements also reach internal tools and the people a decision is about, and you can tell those changes are new only when you monitor vendor terms of service line by line.