The AI Usage Policy Changes on 12 November: Three Requirements for Companies Building AI into Operations

Anthropic, the company behind Claude, has notified users that its Usage Policy will be updated on 12 November 2026. The changes were published on 8 October. The headlines have focused on the new ban on abusive behaviour toward the models, but for companies building AI into their operations the important parts are elsewhere.

We run 45 AI features in production in our own project management tool, projectAI, and build AI into our clients' business systems. Read from that position, the update contains three requirements to check at the design stage: human review and disclosure for high-risk uses, a stop mechanism when AI is connected to physical equipment, and a supported-regions test that looks at who controls a company, not only where its people sit.

This article summarises the update briefly and then explains those three requirements alongside how we design for them. As Anthropic itself writes, most of the changes put existing practice into words. If your design already has the policy's shape, nothing needs to change after the update.

The update at a glance

According to the published announcement there are seven changes.

• A new section on deceptive activity. Prohibitions that were spread across several sections, such as amplification through fake accounts and messages that hide who is behind them, are consolidated into one.
• A narrower elections section. It now focuses on deceiving voters and disrupting elections, and the blanket ban on personalised campaign targeting is removed.
• Clearer weapons prohibitions. Software that makes weapons work and the arming of drones are named explicitly.
• Rewritten surveillance and law-enforcement rules. Tracking people without consent, and using Claude to decide whom to investigate, arrest or charge, are prohibited; fraud monitoring, journalism and legal research are listed as permitted.
• Clearer high-risk requirements. The policy now says which recommendations are covered and adds requirements for physical equipment.
• A prohibition on abusive behaviour toward the models. It covers sustained, purposeless cruelty, not ordinary frustration or model testing.
• A clarified Supported Regions Policy, stating that the test includes control of the entity.

The three that bear directly on business systems are high-risk uses, physical equipment and supported regions.

Requirement 1: in high-risk uses, a person reviews and the person affected is told

Where AI is used for decisions touching health, legal rights, finances, livelihood or essential services, in fields such as law, medicine, finance, credit, insurance, housing, employment and education, the policy asks for two things: a qualified person who can review and change the recommendation, and disclosure to the affected person that AI was used. The requirement itself is not new; the update spells out which recommendations it covers.

Our implementations are built on this premise. In projectAI's 45 features, AI drafts, sorts, translates, searches and applies changes, while people decide policy. Steps with large consequences, such as approving a specification, merging a change or releasing to production, always wait for a person's approval.

An example from education: in Nennai Nyushi Navi, a university admissions service, AI scores personal statements and essays. It returns scores by criterion and comments, but never rewrites the text, and the total score is recalculated by program rather than left to the AI. A decision that affects applicants is not finalised on AI output alone.

Disclosure also applies to consumer-facing chat. The AI chat we built into a furniture retailer's online store answers pre-purchase questions and hands over to the inquiry form or phone when it cannot resolve a question. Not hiding that AI is answering, and keeping a path to a person, is the baseline.

Requirement 2: if you connect AI to physical equipment, a person can stop it

New in this update are requirements for connecting Claude to equipment that takes autonomous physical actions and could injure people, such as vehicles, machinery and medical devices. A qualified operator must be able to observe the equipment and stop it, and the equipment must hold a safe state if its connection to Claude is lost.

In projects that connect AI agents to core systems in construction or manufacturing, this becomes a design premise. In the ordering and core systems we built with a construction company, the AI assistant and automatic sorting of site photos are built so that when AI is unavailable the system falls back to normal processing and work never stops. When equipment is involved, the same idea goes one step further: decide in advance what state the equipment stops in when AI stops.

In our process, the stage that decides where AI is used also writes "how a person stops it" and "the state it stops in" into the specification. Deciding this first is cheaper than adding it later.

Requirement 3: supported regions are judged by control, not only location

The Supported Regions Policy defines the countries and regions where Claude may be used. Japan is a supported region. What the update makes clear is that the test looks beyond where a user is located to who controls the entity. Besides entities incorporated or headquartered in an unsupported region, an entity majority-owned or controlled by persons or entities in such a region may not use Claude, even if its staff are in Japan.

A Japanese subsidiary of a foreign parent should check where the parent is based. If the rule applies, features that would use Claude have to be delivered another way.

We run a configuration in projectAI in which each AI feature can be switched between cloud AI and a local LLM we operate ourselves. A switch applies across the system in about 30 seconds without stopping it. Where a feature cannot use cloud AI because of a usage policy or an internal rule, the same feature runs on the local LLM. The configuration and cost approach are described under Local LLM build and implementation support.

What the update explicitly permits

The update spells out permitted uses as well as prohibitions. The surveillance section lists consent-based tracking such as fraud monitoring, content moderation, journalism and legal research as permitted. In the elections section, removing the blanket targeting ban moves legitimate activity such as policy research and civic education clearly to the permitted side.

The uses most common in business systems, such as classifying inquiries, reading documents, summarising minutes, translation and searching internal data, are not high-risk uses. Unless a decision about a person is involved, the update adds no requirements.

What we check when building in AI

When we build AI into a system, we confirm the following at the start of design. The list does not change with this update.

• Classification of the use. Is AI drafting, sorting and classifying, or making a decision about a person?
• High-risk test. If health, legal rights, finances, livelihood or education are involved, the specification includes a point where a person reviews and can change the outcome.
• Disclosure. Consumer-facing chat and assistants show that they are AI and keep a path to a person.
• Stop mechanism. What happens to the work or the equipment when AI stops: fallback to normal processing, or a safe stopped state.
• Where processing runs. Cloud AI or local LLM, chosen per feature to fit supported regions and internal rules.

The process that includes these checks is described under AI implementation support.

FAQ

Does this policy apply to Japanese companies?

Yes. It applies to every Claude user, and Japan is a supported region. A company with a foreign parent should check the control test based on where the parent is based.

Is disclosure required for AI used only internally?

The disclosure requirement is owed to the individual affected. If AI is used in decisions about employees' evaluation or hiring, internal use can fall within it. Uses that involve no decision about a person, such as summarising documents or translation, carry no extra requirement.

What should we do about AI features we have already built?

Classify whether each is a high-risk use; if so, confirm that a person can review and change the outcome and that disclosure is in place. If a feature is connected to equipment, confirm the stop mechanism. Our implementations include human approval and fallback to normal processing from the start, so nothing changes after the update.

If Claude cannot be used, can the same feature run another way?

Yes. We run a configuration in which each AI feature can be switched between cloud AI and a local LLM, so only the affected features move to the local LLM.

Do other AI providers' policies say the same?

They differ. This article covers Anthropic's policy. If you use several AI providers, check each one's treatment of high-risk uses and supported regions.

Summary

With the 12 November update, companies building AI into operations have three things to check: human review and disclosure in high-risk uses, a person who can stop physical equipment, and a supported-regions test that includes control.

What the policy asks for comes down to designs where AI does not decide and can be stopped. Build those two in from the start and the design does not change when the policy does. If you are unsure how to build AI in, ask a project manager directly through our free estimate consultation.

Related articles and pages

• AI implementation support: from PoC to production and in-house capability
• Local LLM build and implementation support: generative AI without sending confidential data outside
• Case study: projectAI in company-wide use, 45 AI features and where human approval sits
• Japanese companies' competitive edge will be decided by local LLMs

Sources: Anthropic, "2026 Usage Policy update" (8 October 2026), "Usage Policy", "Supported Regions Policy". All checked on 14 October 2026.