Anthropic's 2026 Usage Policy: Audit Your Claude Workflows Before November 12

Anthropic updated its Usage Policy on October 8, 2026. The new version takes effect on November 12. If Claude touches customer decisions, public communications, sensitive data, physical equipment, or autonomous actions, the useful response is not panic. It is a documented workflow audit.

This guide gives you a practical way to find the parts of a Claude workflow that need a qualified reviewer, disclosure, tighter permissions, a safe stop, or a different design. It is an operational checklist, not legal advice.

What changed

Anthropic says most changes clarify existing rules rather than introduce entirely new restrictions. The update does, however, make several areas more explicit:

• Deceptive campaigns and artificial activity now have a consolidated prohibition.
• The elections section is narrower and focuses on deception or disruption.
• Weapons restrictions explicitly include operational software and components.
• Surveillance and criminal justice restrictions are stated more precisely.
• High-risk requirements list covered recommendations more directly.
• Autonomous physical actions now have specific human-control and independent-safety requirements.
• Sustained, needless cruelty toward models is prohibited in extreme cases.
• The Supported Regions Policy clarifies restrictions based on location, incorporation, headquarters, ownership, and control.

The full policy applies to people using Anthropic apps, Claude Code, the API, cloud providers, authorized resellers, and products that integrate Claude. It also says builders remain responsible for actions taken by their agents.

The practical question: what can your workflow do?

Do not audit prompts in isolation. Audit the entire path from input to consequence.

A harmless-looking summarization step can become higher risk when its output ranks job candidates. A drafting assistant can become a deceptive system if it operates fake personas. A planning agent can create physical risk when its output controls equipment without an independent safety layer.

Map each workflow as 6 connected parts:

1. Input: What data enters, and whose data is it?
2. Model task: What judgment, transformation, or recommendation does Claude make?
3. Tools and permissions: What systems can the workflow read or change?
4. Output: What does the model produce?
5. Decision or action: Who or what acts on the output?
6. Affected person: Who could experience a material consequence?

Run the 5-part Anthropic usage policy workflow audit

1. Classify the consequence

Start with what happens after Claude responds.

Do not label a workflow low risk because the first step is ordinary. Follow it to the final decision or action.

2. Check universal prohibitions

Compare the workflow against the policy's universal standards. Pay particular attention when it could:

• impersonate a real person or conceal artificial origin
• generate or operate fake accounts or coordinated artificial activity
• collect, enrich, track, or expose personal data without appropriate permission
• recommend criminal justice or law-enforcement decisions
• enable unauthorized access, malware, surveillance, weapons, or harmful physical activity
• generate spam, fraud, deceptive reviews, or guardrail circumvention

If the workflow's intended operation falls inside a prohibition, stop. A reviewer cannot convert a prohibited use into an acceptable one.

3. Identify high-risk recommendations

Anthropic's policy names areas where a model recommendation can materially affect a person: legal, medical, finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits, and legal status or adjudication.

When a workflow produces a covered recommendation, the policy requires a qualified person to review it meaningfully before it is delivered or used. That person must have the relevant training or experience, hold a license where applicable law requires one, be able to change the recommendation, and remain responsible for it. The affected individual must also be told that AI was used.

The policy also defines exclusions, including general information, internal assistance that is not the final recommendation, and execution of a decision already made by a person. It permits limited exceptions to prior human review for wholly favorable insurance, healthcare, and public-benefit decisions where allowed by law; partial or conditional approvals do not qualify, and disclosure or other obligations may still apply. Check the exact policy rather than treating every related task as covered.

Ask 4 questions:

1. Does Claude apply judgment to an individual's circumstances?
2. Could the output affect rights, health, finances, livelihood, housing, education, or essential services?
3. Is a qualified person reviewing the output before it matters?
4. Is AI involvement disclosed clearly to the affected person?

If the first 2 answers are yes, do not rely on a generic approval button. Record who qualifies as the reviewer, what they inspect, what evidence they receive, and how they change or reject the output.

4. Review agent permissions and physical actions

For agents, document each tool, allowed action, data source, permission boundary, and failure path. Use the smallest permission set that still supports the job. Separate reading from writing where possible. Require confirmation before irreversible or high-impact actions.

If model output is acted on by hardware without human approval and can cause equipment to move, apply force, control hazardous energy or materials, act on a body, control a safety system, or run an industrial process, the updated policy requires stronger controls. A qualified operator must be able to observe and stop the equipment, and the equipment must stop or hold a safe state when the operator intervenes or the Claude connection is lost. Safety limits must be enforced independently of model output.

The policy excludes certain low-consequence household devices, non-hazardous enclosed laboratory automation, monitoring without control, and plans or commands reviewed by a qualified person before equipment executes them. Check those boundaries against the exact use case.

For a covered high-risk physical action, missing controls are a stop condition, not a prompt-engineering problem. Do not deploy until the operator, independent controller, disconnect behavior, and physical acceptance tests are documented.

5. Verify access, disclosure, and evidence

Finish the audit with 3 cross-cutting checks:

• Access: Confirm that users, entities, ownership, and operating locations comply with the current Supported Regions Policy.
• Disclosure: Consumer-facing chatbots and agents must disclose that users are interacting with AI. High-risk recommendations require an additional disclosure that AI was used.
• Evidence: Keep a record of the policy version, workflow owner, review date, classification, controls, tests, exceptions, and next review.

Policies and product behavior can change. Link your record to the current source rather than copying a rule into a document and forgetting it.

Copy this workflow audit record

Use one record for each production workflow.

Workflow name:
Owner:
Audit date:
Policy version and effective date:

1. Intended outcome:
2. Inputs and data owners:
3. Claude task:
4. Connected tools and exact permissions:
5. Output:
6. Final decision or action:
7. People who may be affected:

Universal prohibition check:
[ ] No prohibited intended use identified
[ ] Deception and artificial activity reviewed
[ ] Privacy, surveillance, and identity risks reviewed
[ ] Security and physical-harm risks reviewed

High-risk check:
[ ] No covered recommendation
[ ] Covered recommendation: category __________
[ ] Qualified reviewer named and authorized
[ ] Reviewer can change or reject the output
[ ] AI disclosure is present before impact

Agent and action controls:
[ ] Least-privilege access
[ ] Approval before high-impact action
[ ] Action log and owner
[ ] Manual fallback
[ ] Rollback or containment path
[ ] Safe behavior on model or network failure

Regions and access:
[ ] User location checked
[ ] Entity incorporation and headquarters checked
[ ] Ownership or control checked when relevant

Decision:
[ ] Proceed
[ ] Proceed after controls are added
[ ] Escalate to qualified legal, compliance, safety, or domain review
[ ] Stop because the intended use is prohibited

Open gaps:
Control owner:
Due date:
Next review date:

A worked example

Imagine a recruiting team uses Claude to summarize applications and draft interview notes.

The first version also scores candidates and automatically rejects people below a threshold. That final action changes the classification. Employment screening and rejection are listed high-risk recommendations. The team should not treat a manager's occasional spot check as sufficient review.

A safer redesign separates assistance from decision:

1. Claude extracts job-relevant evidence into a fixed schema.
2. The system records source references and missing information.
3. It does not infer protected attributes or use proxy criteria.
4. A qualified recruiter reviews every record before a candidate advances or is rejected.
5. The recruiter can edit the evidence, reject the model output, and records the decision rationale.
6. The candidate-facing process discloses that AI assisted the review where required by the policy and applicable law.

This example is illustrative. It does not determine legal compliance, and local employment or AI laws may impose additional requirements.

What to do before November 12

Prioritize workflows with the largest consequences and the most autonomy.

1. Inventory every production use of Claude, including third-party products and cloud access.
2. Map the full workflow, not only the prompt.
3. Classify prohibited uses, high-risk recommendations, agent actions, and physical actions.
4. Add named reviewers, disclosures, permission limits, logs, manual fallbacks, and safe stops where required.
5. Re-test the workflow with missing data, ambiguous inputs, tool failure, model refusal, and network loss.
6. Record the result and schedule a policy review after material workflow or policy changes.

The strongest audit outcome is not a completed checklist. It is a workflow whose permissions, review points, disclosures, and failure behavior match the consequence it can create.

Put the audit into practice

• Reduce sensitive inputs with 8 things never to paste into AI.
• Map permissions and human approval with the AI agent blueprint.
• Test escalation behavior with the Claude Haiku model-routing test.
• Assign practical assessment and handover responsibilities using enterprise AI engineer training.

Sources and verification

Verified October 10, 2026:

• Anthropic: 2026 Usage Policy update
• Anthropic Usage Policy, effective November 12, 2026
• Anthropic Supported Regions Policy
• Anthropic guidance for using agents under its Usage Policy

This article summarizes Anthropic's published policy for operational planning. It is not legal, medical, financial, employment, or compliance advice. Review the current policy and applicable law for your use case.