Sort your Claude workflows before the new usage policy starts

On October 8, 2026, Anthropic published a new version of its Usage Policy. The new version takes effect on November 12. Any business that uses Claude, in the apps or through the API, works under it.

Most of the update clarifies existing rules. One part matters to almost every small business: the section on high-risk use cases. Anthropic says those requirements have not changed. It rewrote the section so that it lists which recommendations are covered and which are not. That list makes a self-check possible.

What the policy requires

The Usage Policy names high-risk areas. They include legal, medical, finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits, and legal status. A recommendation about a specific person in one of those areas is a “High-risk AI Recommendation”.

we require that a qualified person remain accountable for what is delivered or decided

For each High-risk AI Recommendation, the policy sets two controls:

• A qualified person reviews the recommendation before it is delivered or used. That person has the authority to change it. A qualified person has the training or experience to judge the output, and a license where the law requires one.
• The person who receives the advice, or who is the subject of the decision, is clearly told that AI was used. The notice does not need to name Anthropic, Claude or the model.

The policy also lists uses that the requirements do not cover. These matter most to a small business:

• Internal drafting, research, summarization or analysis that is not the final recommendation delivered.
• Carrying out a decision a person has already made, or applying a fixed rule or formula where the model uses no judgment about the person.
• Business operations that do not advise or decide about a specific person.
• General information, such as explaining what a law says, without applying it to the person’s situation.

Two more rules apply whatever the risk level. A consumer-facing chatbot must tell users they are dealing with AI, at least at the start of each chat or in the product interface. And the business is responsible for what its agents do through tools and connected systems. The policy also says local law comes first where it adds requirements.

This article explains Anthropic’s policy text. It is not legal advice. Housing, employment and credit laws in your state or country can add duties of their own.

The task

A 25-person property management firm uses Claude in five workflows. Sort each one against the new policy and say what must change by November 12.

1. Leasing: Claude reads each rental application and drafts a ranked shortlist with an approve or decline note for the leasing manager.
2. Website chat: a Claude-based assistant answers prospective tenants’ questions about viewings, amenities and pet rules.
3. Maintenance rota: Claude drafts next week’s technician shifts from availability forms and open job tickets. The operations lead publishes it.
4. Late rent: Claude writes reminder letters. The lease sets a fixed late fee after five days. Claude fills in the amount and the date.
5. Owner reports: Claude summarizes monthly maintenance spend by building for the property owners.

Emeri’s worked answer

1. Leasing is high-risk. The policy’s housing area covers approving, denying or ranking a rental application. The draft shortlist is not “internal drafting”, because the manager acts on it. Two changes follow. The leasing manager reviews each ranking and note before any applicant hears back, and records any change. The application form and the decision letter tell applicants that AI helped assess their application.
2. Website chat is not a High-risk AI Recommendation as described. It is a consumer-facing chatbot, so it must say it is AI at the start of each chat or in the chat window. Add one more check. The assistant must not tell anyone whether they are likely to qualify for a unit. That answer would move it into the housing area.
3. The maintenance rota is high-risk. This is the one most teams miss. The employment area covers setting or changing shifts, hours or pay. The operations lead already publishes the rota. That counts only if the lead reviews it with real authority to change it, not a quick click. The technicians must also be told that AI drafted their shifts. A line on the rota itself is enough.
4. Late rent is not high-risk as it runs today. The lease formula sets the fee, and Claude applies it without judgment. Move the fee calculation into the rent spreadsheet so the model only writes the letter. Draw a firm line in writing: Claude does not decide to waive a fee, end a tenancy or start an eviction. Ending a tenancy and starting an eviction are in the housing area.
5. Owner reports are not high-risk. They are business operations about buildings, not about a specific person. No change is needed.
6. Write the result down in one table: workflow, high-risk or not, the policy area, the reviewer, the notice, and the date of the check. Name an owner for each row. Recheck the table when a workflow changes.

Result: two of the five workflows are high-risk and need a named reviewer and a notice. One needs a chatbot disclosure. One needs a written limit. One needs no change.

What would score lower

• Marking the maintenance rota as safe because it is “only scheduling”. The employment area names shifts and hours.
• Treating a manager’s approval click as review. The policy asks for meaningful review by a person with the authority to change the recommendation.
• Assuming a favorable outcome needs no review. The policy’s exception for wholly favorable decisions covers insurance claims, care approvals, healthcare eligibility and public benefits. It does not cover housing or employment.
• Asking Claude whether the firm complies, instead of reading the policy text. The tool is not an independent check of its own use.
• Treating the policy as the only rule. Fair housing and employment laws may add their own duties.

Emeri’s certification grading runs on Claude models.