RGBvision RGBvision
Back
aiproduct

Anthropic's new Claude usage policy: what changes for businesses

Anthropic banned cruelty toward Claude, but companies should look elsewhere: AI disclosure in chats, human review and region restrictions.

Published: 2026-10-09

On October 8 Anthropic published a new version of its Usage Policy for Claude, effective November 12. Headlines focused on one clause — "you can't insult Claude" — but for companies building products on language models, the other changes matter more. The short answer: if you run a customer-facing bot or agent on Claude, check three things — it tells users they are talking to AI, high-risk recommendations are reviewed by a person, and your company is allowed to use Claude in the first place.

What was actually banned

The section on cruel and psychologically harmful conduct now includes a new item: no "sustained and needless abusive or cruel behavior toward our models." Anthropic says it targets extreme cases where users repeatedly act cruelly toward a model with no discernible purpose. Ordinary frustration, pushback, dark creative themes, and model testing or research are not covered.

Enforcement relies mainly on ending the conversation: in Claude.ai and Claude Code the model can close a chat with a persistently abusive user. Anthropic has not said that being rude to Claude gets accounts banned, even though some news coverage suggested so.

What matters more for companies

The bot has to say it is a bot

The policy states it directly: all consumer-facing chatbots and AI agents that communicate with external users must disclose that users are interacting with AI rather than a human. The disclosure must come at least at the start of each chat session or in the product interface. You don't have to name Anthropic or the specific model.

In practice that is one line in the greeting or a label next to the chat window.

We handle this in our own product, DialIQ: the company writes the AI operator's role itself, including how it introduces itself to customers. Complex conversations, payments and complaints are handed over to a person.

Recommendations reviewed by a human

For recommendations in high-risk areas, the policy requires two things. A qualified person meaningfully reviews the recommendation before it reaches the customer, can change it, and stays responsible for the result. And the person who receives the advice, or who is the subject of a decision, must be clearly told that AI was used to produce it.

The policy lists eleven high-risk areas: legal, medical, finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits and services, and legal status and adjudication. If your agent helps screen candidates, assess borrowers or suggest legal steps to a customer, this applies to you.

Booking an appointment, checking an order status or answering a pricing question does not. The line depends on what the bot actually does: a "booking assistant" that starts recommending treatment is already in high-risk territory.

Fake activity and "invisible" marketing

A new section on deceptive campaigns brings together rules that used to be scattered across the document: no fake accounts or media outlets, no bots presenting themselves as humans, no fake reviews.

One item addresses what is now called AI search optimization: you may not seed the sources that search engines and AI systems draw answers from with content that hides its real origin. Publishing articles under your own name is fine; generating "independent" reviews of your own product is not.

What this means for companies tied to Russia

Russia is not among the regions where Anthropic allows its services to be used. Alongside the policy update, Anthropic clarified how it enforces this. Its services may not be used by people physically located in an unsupported region, by entities incorporated or headquartered there, or by entities majority-owned or controlled, directly or indirectly, by persons or entities in such regions.

So a foreign company with Russian owners does not solve the access question, and for a product that is a risk: access can be cut off, and everything built on the model stops with it. The same logic applies to any company whose ownership ties it to an unsupported region.

That is why we design agents so the model can be swapped: our own instructions and knowledge base, quality tests on real requests, and a separate layer between the product and the model. We covered how to choose a model in our article on LLMs for business in Russia.

What to do before November 12

  1. List every place Claude is used, including integrations built by contractors.
  2. Check AI disclosure in every customer-facing bot: a greeting line or an interface label. The bot can have any name, but users must know it is AI.
  3. Find high-risk scenarios where the agent gives advice or prepares decisions in the listed areas. Add human review before delivery and tell the customer AI was involved.
  4. Review your marketing: reviews, publications, content aimed at AI search — is all of it published under your name? If an agent controls hardware, also check the new requirements for safely stopping equipment.
  5. Assess vendor risk: are you allowed to use Claude under the policy, and how long would switching to another model take? If the answer is "we don't know," start by testing an alternative model.

We audit agents and migrate them to other models as part of AI agent development. For an example of an AI operator with a role, a knowledge base and handover to a person, see our omnichannel chat.

Sources