Anthropic Usage Policy from 12 November 2026: what HR, lending and GEO teams change

Summary. Anthropic's rewritten Usage Policy, announced on 8 October 2026, takes effect on 12 November 2026. It binds everyone who sends inputs to Claude, including API customers, cloud and reseller customers, and end users of products built on Claude. Against the version in force since 15 September 2025, three changes matter most. First, the high-risk requirements now name 11 areas instead of 7 categories, separating credit from finance and spelling out employment to cover screening, shifts, pay, performance reviews and dismissal. A qualified person who can change the output must review each recommendation before it is used, and the person affected must be told AI was involved; Anthropic says these duties are unchanged and only clarified. Second, a new deceptive-campaigns section bans seeding search engines or AI answer sources with content that hides its origin, which rules out GEO site networks posing as independent sources. Third, reselling or proxying Claude, including through consumer subscriptions, is now named as platform abuse. Automatically published media content is no longer a listed high-risk use. This article is not legal advice.

About this article. For product, HR-tech, lending and marketing teams building on Claude. On 11 November 2026 we compared Anthropic's current Usage Policy with its archived 15 September 2025 version, and read Anthropic's announcement, Supported Regions Policy and agent guidance, the European Commission's AI Act pages and California's ADMT regulation text. Readings of specific workflows are ours. This is general information, not legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What takes effect on 12 November 2026

Anthropic published its 2026 Usage Policy update on 8 October 2026, and on 11 November 2026 its legal centre still listed the version effective 12 November 2026 as current. The previous version took effect on 15 September 2025. The Supported Regions Policy carries the same 12 November date, and India, the United States, the United Kingdom and all 27 EU member states remain on its list of supported countries.

The policy has three parts:

• Universal Usage Standards: prohibitions that apply to every user and use case, now in 15 sections against 14 in the 2025 text.

• High-risk Use Case Requirements: controls for recommendations affecting a person's health, legal rights, finances, livelihood, housing, education or essential services, plus new controls for hardware that takes physical actions.

• Additional Use Case Guidelines: rules for consumer-facing chatbots and agents, products serving minors, agentic use and software listed in Anthropic's directories.

Anthropic says most changes clarify existing rules. Its post lists a new deceptive-activity section, a narrower and renamed elections section, clearer weapons and surveillance language, the rewritten high-risk section with controls for physical actions, a ban on sustained, needless abuse of the models, and clearer Supported Regions enforcement. The regions page now spells out that the restriction covers people in an unsupported region, entities incorporated or headquartered there, and entities majority-owned or controlled from there; the ownership rule dates from 2025.

Enforcement runs through access. If Anthropic suspects a breach, it may warn you, or throttle, limit, suspend or terminate access, and a real-time safeguard that blocks an output is not by itself a breach. The policy also tells users to follow local law where it conflicts with or adds to the high-risk requirements, so treat it as a floor, not a ceiling.

September 2025 versus November 2026: the clauses that moved

Anthropic keeps the previous version in its legal archive. The table compares the archived 15 September 2025 policy with the 12 November 2026 policy on the points that matter here.

Two caveats apply. Anthropic says the high-risk requirements themselves have not changed, so treat the scope row as a wording change to settle with your account team, not a confirmed expansion. And the 2026 exclusions are examples, not a closed list.

HR and hiring teams: employment is now spelled out

The 2026 Employment area covers screening, ranking, advancing or rejecting candidates; hiring; promotion and demotion; setting or changing shifts, hours or pay; and evaluating a worker's performance or behaviour, including discipline, suspension and dismissal. The 2025 text put employability, resume screening, hiring tools and other employment determinations in one line, a catch-all that arguably already reached workforce management. The new list removes the doubt: a Claude-powered shift scheduler or performance-review assistant is plainly covered.

When a Claude output is a recommendation in one of those areas, three things must be true before anyone acts on it:

1. A qualified person reviews it. That means someone with the training or experience to judge the output in that field, licensed where the law requires, who can change the recommendation and stays responsible for the outcome.

1. The individual is told. A candidate ranked or rejected with Claude's help is the subject of a decision and must be clearly told AI was used. You need not name Anthropic, Claude or the model.

1. Local law still applies on top. The EU AI Act, California's ADMT rules and local employment law each add their own duties.

There is no favourable-decision shortcut for hiring. Where the law allows, human review can be skipped only for four wholly favourable outcomes: paying an insurance claim, approving coverage or prior authorisation for care a clinician has ordered, granting eligibility for a healthcare service, and granting or continuing a public benefit. Advancing a candidate is named in the Employment area, so it needs review.

The exclusions help with the work around the decision. The table gives our reading of them for common HR workflows.

Several Universal Usage Standards also bite on HR products. The policy bans emotion recognition except for medical or safety reasons, which rules out confidence or enthusiasm scoring from video interviews built on Claude. It also bans biometric categorisation to infer traits such as race, religion or sexual orientation, trustworthiness or social scoring without notice or consent, and tracking a person's online activity or communications without consent. Its anti-discrimination clause now names caste and disability alongside race, religion, national origin and gender, which matters for Indian employers; the 2025 clause named neither. The surveillance section does not prohibit tracking people have agreed to, but whether an employee's agreement counts as valid consent is a question for local law.

Lending and credit teams: approvals need review too

The Credit area covers approving, denying, or setting the amount, rate or limit of a loan, mortgage or credit line, reducing or closing one, and producing a credit score or creditworthiness assessment used for those decisions. Finance is now narrower: personalised investment recommendations and individual tax advice or returns. In 2025, loan approvals and creditworthiness sat inside one broad Finance category. Housing covers approving, denying or ranking rental and purchase applications, ending a tenancy, and eviction or foreclosure.

Three points stand out for lenders:

• Approvals are not exempt. Credit is not on the favourable-decision list, so an AI-assisted approval needs qualified human review just as a decline does. A limit increase recommended by Claude is a covered recommendation.

• The borrower must be told. An applicant who is the subject of a credit decision based on a Claude recommendation must be told AI was used.

• Scores count. A creditworthiness assessment produced with Claude and used for these decisions is covered even if a separate system makes the final call.

The exclusions draw a practical line. An underwriter who uses Claude to summarise bank statements or flag inconsistencies in an application file is doing internal analysis that is not the final recommendation, which the policy excludes. Anthropic's 9 November 2026 launch post for Claude Fable 5.2 carries a tester's account of the model spotting that salary deposits on a loan application did not match the bank statement balance, which is exactly this kind of work. The line is crossed when the model's output becomes the recommendation that sets approval, rate or limit. Applying a fixed formula where the model makes no judgement, such as computing a debt-to-income ratio, is also excluded.

On prohibitions, the policy keeps bans on predatory lending and abusive debt collection that exploit people, and on products that use manipulative or deceptive techniques to distort decisions. The 2025 text named payday and title loans; the 2026 text folds them into predatory lending. That is a drafting change, not a permission. Fraud and anti-money-laundering screening on a financial account are given as examples of agreed tracking, which the surveillance section allows. Collections voice bots must disclose that they are AI; on a call, that in practice means at the start.

GEO, content and marketing teams: deception now has its own section

The new section, Do Not Engage in Deceptive Campaigns or Artificial Activity, gathers rules Anthropic says were spread across its elections, fraud, privacy and disinformation sections, and covers commercial as well as political work. It prohibits using Claude to:

• run fake personas, accounts, outlets or organisations that mislead people about who is behind content, how widely a view is held, or whether they are dealing with a human, such as sockpuppets, astroturfing, fake reviews and bots posing as people;

• hide or misstate who sponsored content meant to sway public opinion, officials or other decision-makers;

• push content through networks of fake or seemingly independent outlets, sites or accounts to hide a common source;

• manipulate the sources that search engines or AI systems draw answers from, by seeding them with content that misstates its origin, authorship or independence;

• build tools or infrastructure for such campaigns, or sell them as a product or service.

GEO teams should read the fourth bullet twice. The policy's example is a group of sites presented as unrelated sources that back up the same claims. Fake reviews, fake personas and passing AI output off as human were already banned in the 2025 text, in its fraud, misinformation and privacy sections; the search and AI-source bullet has no counterpart anywhere in it. Anthropic's agent guidance adds that agents must not be used for click farming, artificial engagement, or manipulating online polls or traffic metrics.

What stays allowed, on our reading: publishing your own research, data and opinions under your own brand; structured data and technical SEO; earning genuine third-party citations; and paid placements that are clearly labelled. The test is whether content misrepresents its origin, authorship or independence. Our guide to AEO, GEO and SEO covers the legitimate side of that work.

The rewrite also loosens one thing. The 2025 text listed media content generated automatically and published for external consumption as a high-risk use case needing human review and AI disclosure. That category is absent from the 2026 list. Other rules still apply: spam is banned, AI-assisted work needs permission or attribution, and the Disinformation section bars fake personas that convince people they are talking to a human (a rule that sat under privacy in 2025). An outbound sales agent posing as a person breaks that rule and the chatbot disclosure rule. Claude's text watermarking, covered in our piece on Claude watermarking for content teams, is a separate technical matter.

Chatbots, agents and resale

The chatbot disclosure rule is not new, but its form changed. Every consumer-facing chatbot or AI agent that talks directly with external users must disclose that it is AI, at least at the start of each chat session or in the product interface, and the disclosure need not name Anthropic, Claude or the model. In 2025 the minimum was the start of each session.

For agents, the 2026 text makes users responsible for ensuring that agents they build or deploy comply, including actions taken through tools, browsers or connected systems. On our reading, a builder cannot point to the model as the actor.

The Platform Abuse section adds two named prohibitions: reselling or proxying Claude through unauthorised means, including services that route requests through consumer subscriptions or misrepresent the client in use; and using stolen or unauthorised payment methods. Neither appears in the 2025 platform section. On our reading, a shared tool routing a team's or client's requests through one person's consumer subscription is the named pattern, so move such tools to API or business-plan access. Training a model on Claude outputs without authorisation remains prohibited.

A quick decision table for builders

The table condenses the requirements above; logging is our recommendation, not a policy requirement.

How the policy sits next to the EU AI Act and California's ADMT rules

Anthropic's policy is a contract term, not a law, but it overlaps with two regimes covered in our guides to the EU AI Act high-risk rules for hiring and credit and California's ADMT rules for AI hiring and lending. The EU rows draw on the Commission's AI Act pages, the California rows on the state privacy agency's regulation text and guidance.

The AI Omnibus entered into force on 27 July 2026 and moved the Annex III start date to 2 December 2027. The California rules were approved in September 2025, and a business already using ADMT for a significant decision must comply by 1 January 2027.

For Indian teams, the Digital Personal Data Protection Act, 2023 governs the personal data in candidate and borrower files, and the policy says local law prevails where it adds requirements. Designed once, one review log and one disclosure notice can often serve all four regimes.

Implementation checklist

1. Inventory every Claude call and map each to a high-risk area, an exclusion, or neither.

1. Name the qualified reviewer for each covered feature, record their training or licence, and give them a real override in the interface.

1. Log the review: who reviewed, what they changed, and when, before the advice was delivered or the decision applied.

1. Add disclosure to candidate, applicant and borrower notices, and to every chatbot or voice agent session or interface.

1. Remove prohibited features such as emotion inference from interviews and undisclosed monitoring.

1. Audit GEO and content tooling for networks of unattributed sites, fake reviews, sockpuppet accounts and hidden sponsorship.

1. Move off consumer subscriptions for anything a team or client uses programmatically.

1. Write to your Anthropic account team if your B2B design relied on the 2025 consumer-facing wording, and keep the reply.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team helps products built on Claude map features to the high-risk areas, add reviewer workflows with real override and audit logs, and put AI disclosure into candidate, borrower and chatbot journeys. Talk to us at /contact-us/.

Last updated: 11 November 2026.

References

1. 2026 Usage Policy update, 8 October 2026 — Anthropic

1. Usage Policy, effective 12 November 2026 — Anthropic

1. Usage Policy, archived version effective 15 September 2025 — Anthropic

1. Legal Center — Anthropic

1. Usage Policy update, 15 August 2025 — Anthropic

1. Supported Regions Policy, effective 12 November 2026 — Anthropic

1. Usage Policy (PDF), effective 12 November 2026 — Anthropic

1. Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026 — Anthropic

1. Using agents according to our Usage Policy — Claude Help Center

1. AI Omnibus enters into force, 27 July 2026 — European Commission, Shaping Europe's digital future

1. AI Act Annex III: high-risk AI systems — AI Act Service Desk, European Commission

1. Article 26: Obligations of deployers of high-risk AI systems — AI Act Service Desk, European Commission

1. Article 99: Penalties — AI Act Service Desk, European Commission

1. Text of regulations: CCPA updates, cybersecurity audits, risk assessments, ADMT and insurance — California Privacy Protection Agency

1. Frequently asked questions — California Privacy Protection Agency

1. California finalizes regulations to strengthen consumers' privacy, 23 September 2025 — California Privacy Protection Agency