A story went around about Anthropic suspending an entire company's Claude accounts on a Monday morning with no warning, with the staff locked out while the bill supposedly kept running. I want to be careful with that version, because the most dramatic details, a 110-person company and the API billing through the lockout, trace back to a single writer and are not independently confirmed. I am not going to repeat them as fact.
What is confirmed is bad enough, and it is the reason every small business should spend an afternoon on this.
The case that actually checks out
In April 2026, Anthropic abruptly suspended roughly 60 employee accounts at Belo, an Argentina fintech, over a vague usage-policy flag. It turned out to be a false positive, and access came back about 15 hours later. The catch that made it sting: the only route to appeal was a Google Form, and reaching the right channel meant logging into the very accounts that had just been banned. Belo's CTO, Pato Molina, put the lesson plainly in public: never put all your eggs in one basket (Tom's Hardware, OECD.AI incident monitor, NewsBytes).
There is also a real clause worth reading. Anthropic's Commercial Terms include a suspension liability waiver, Section I.3.b, stating the company "will have no liability for any damage, liabilities, losses (including any loss of data or profits)" that a customer incurs because of a service suspension (Anthropic Commercial Terms). The same terms say Anthropic will use reasonable efforts to give written notice of a suspension. They do not say anything about whether your fees keep accruing during one, which is itself a reason to read your own contract closely rather than assume.
One thing to be fair about: this is not an Anthropic-specific failing. Every major AI provider can suspend accounts on an automated flag, most route appeals through slow self-service channels, and most carry a similar liability waiver. The risk is not their policy. The risk is how completely your business depends on a single account staying on.
Your one-page AI continuity plan
You do not need a consultant for this. Write these six lines down, do them once, and you have turned a business-stopping event into an annoyance.
- Configure a second provider behind a simple switch. Route your AI work through one place in your setup that you can point at a different model. If one account goes dark, you flip a setting instead of closing for the day. The cheaper-model and metered-API posts below cover the how.
- Keep your admin and billing identity on a separate email. If your day-to-day user seats get banned, you want a different login that can still reach the billing console and the appeal form. A suspension that locks every email you have is the real trap.
- Keep your data and source files local. Do not let the only copy of anything important live inside the vendor. Export conversations and outputs on a schedule, and keep the originals on your own machine.
- Set a spending cap and a billing alert. Since the terms are silent on charges during a suspension, do not leave it to trust. A hard cap and an alert protect you from both a runaway bill and a quiet one.
- Bookmark the appeal path now, and know it needs a login. Find your provider's suspension and appeal page today, save it, and confirm what it requires. Discovering the catch-22 mid-outage is the worst time to learn it.
- Do not run your whole company through one model account. Separate the truly critical workflow from the experimental one. Blast radius is something you control by how you set things up, not something the vendor decides for you.
That is the entire plan. It costs a little setup time and the price of a second provider you may rarely touch, and it converts a Monday-morning lockout into a shrug.
This is the same single-vendor discipline I argue for across a lean tool stack in my book The $20 Dollar Agency (search the title on Amazon Kindle): use the cheap tools, but never let one of them become a single point of failure for the business.
Related reading
- Why Use the Separate Anthropic API, and How It Actually Works: the metered, more controllable way to consume a model.
- Western Tech Is Routing AI Workloads to Chinese Models: the router pattern that makes a second provider a one-line switch.
- Running Your Own AI On-Prem in 2026: the ultimate fallback when you cannot be locked out at all.
- How to Read an AI Subscription Before You Pay for the Big Tier: reading the limits and terms before you commit.
- How a Small Business Runs AI Agents Without a Surprise Bill: the spending caps that belong in this plan.
Fact-check notes and sources
- Belo suspension: roughly 60 employee accounts cut off in April 2026 over a vague usage-policy flag, restored after about 15 hours, with a Google Form as the only appeal route (Tom's Hardware, OECD.AI, NewsBytes, Inshorts).
- Suspension liability waiver: Section I.3.b of Anthropic's Commercial Terms; the terms also commit to reasonable efforts to give written notice, and are silent on fees during a suspension.
- Appeals require a login: per Anthropic's safeguards and appeals help article.
- Not stated as fact here: the widely-shared "110-person company" and the claim that "API keys kept billing during the lockout" appear in a single author's account and were not corroborated by primary reporting, so I have left them out and framed billing-during-suspension as a risk to check in your own contract rather than an established fact.
This post is informational, not legal advice. Contract terms change and vary by plan; read your own agreement. Mentions of Anthropic, Claude, Belo, and other third parties are nominative fair use. No affiliation is implied.
