Home Services Work About Blog FAQ Contact
← Back to blog

Claude's new usage policy lands November 12 — here's who actually has to change anything.

Anthropic published a usage policy update on October 8 that takes effect November 12. Most of it clarifies existing rules. A few parts are new requirements, and they land hardest on agents that touch health, legal, financial, or hiring decisions — or anything with a motor attached.

If you're running a Claude-based agent for your business, you probably skipped the usage-policy email. Fair. Most policy updates are lawyers restating things that were already true. This one mostly is, too — but it adds two real requirements, and if your agent touches what Anthropic now calls a "High-risk Use Case," you need to know about them before November 12.

A qualified person must meaningfully review a High-risk AI Recommendation, with the authority to change or modify it, before it reaches anyone it affects.

That line is the center of the update. It's not brand new — Anthropic has required human oversight for high-stakes decisions for a while. What's new is how specific the policy now gets about which decisions count, and a second requirement that didn't exist before: telling the affected person that AI was involved at all.

What actually counts as high-risk

The policy names four areas most small businesses will recognize. Healthcare covers "diagnosing a condition" and "recommending, starting, stopping, or adjusting a treatment or medication," plus anything that sets "triage or waitlist priority." Legal covers "advising an individual on their legal rights, obligations, or liability." Finance covers a "personalized recommendation to buy, sell, hold, or allocate among specific investment products," along with individual tax advice. Employment covers "screening, ranking, advancing, or rejecting candidates" — and pay and termination decisions, too.

Notice what's not on that list: drafting a reply, summarizing a document, scoring a lead, writing a social post. Those stay exactly as unregulated as they were last week. The line Anthropic drew isn't "did AI touch this," it's "does the output change what happens to a specific person's health, money, legal standing, or job."

If you've built a lead-qualification agent, a support triage bot, or a content pipeline, this update changes nothing for you. If you've built — or a client has asked you to build — anything that screens job applicants, triages patients, or gives personalized financial guidance, it changes two things: a qualified human has to be able to catch and override the output before it reaches anyone, and that person has to be told AI produced it.

The part that surprised me: hardware

Buried further down is a requirement for anything where Claude controls equipment that takes physical action. The policy requires "a qualified individual who is able to observe the equipment's operation and stop it at any time," and it requires that safe operating limits be "enforced by the equipment or a controller independent of model output" — meaning the model itself can't be the only thing standing between normal operation and something going wrong.

We don't see many clients wiring Claude directly into machinery. But it's worth a second look if you've got anything AI-driven touching irrigation controllers, warehouse equipment, or a drone flight path — the model recommending an action isn't enough anymore; something outside the model has to be able to pull the plug.

What to actually do before November 12

Three checks, in order of how long each one takes.

  1. Walk your agent's outputs against the four high-risk categories. Most agents don't touch any of them. If yours does — even one step in a longer pipeline — that step needs a named human reviewer with real authority to change the result, not just a rubber-stamp approval click.
  2. Add the disclosure, if you haven't already. A one-line note — "this recommendation was generated with AI assistance and reviewed by [name]" — satisfies the letter of it. It should already be in your output template if you built the approval step correctly in the first place.
  3. Check who owns the kill switch on anything physical. If nothing you've built controls hardware, skip this. If something does, confirm the stop mechanism doesn't depend on the model agreeing to stop.

Most of what you'll find doing this walkthrough is that you're already compliant, because the underlying advice hasn't changed — we wrote about where to put approval gates back in August, and the same irreversibility logic applies here. The new part is narrower than it sounds: name the reviewer, say the words "AI was used," and don't let a model be the only thing holding a physical system's safety limits.

The rest of the update — new rules on fake accounts and election-related deception, a tightened weapons section, limits on nonconsensual surveillance — matters if you're building anything in those categories. For a typical small-business AI build, it's not the part to lose sleep over.

If you want a second set of eyes on whether an agent you're running (or planning) crosses into one of these categories, that's a quick conversation, not a redesign — reach out and we'll look at it together.

Sources

— Cole

Not sure if your agent crosses into "high-risk"?

Send us what it does. We'll tell you plainly whether anything needs to change before November 12.

Book a Discovery Call →