Claude data retention after EFS: what small firms get

Claude data retention after Anthropic’s September 1, 2026 Enterprise Frontier Safeguards announcement is a bank-grade custody product. It is not the default you get when a 12-person firm opens Claude.

Anthropic named it EFS: zero-data-retention privacy plus misuse detection, with activity data stored in cloud infrastructure the customer controls. It will roll out in phases later this fall. Eligible customers keep ZDR on Fable 5 and Fable 5.1 until EFS is ready. That sentence is for security teams at Wells Fargo, Stripe, and the other Fortune 100 shops that helped design it. It is not a setting on your Pro plan.

What is Claude data retention under Enterprise Frontier Safeguards?#

Anthropic now treats 30-day retention on Fable 5 as the safety default, and EFS as the opt-in way regulated companies keep those logs in their own AWS, Azure, or Google account. The model does not change. Who holds the tape does.

Mythos-class models such as Claude Fable 5.1 can be misused across sessions and accounts, so per-prompt discard is not enough to catch fraud or stolen credentials. Hence the 30-day window. Anthropic also says it does not train on enterprise data without permission. Both can be true: safety monitoring still needs a copy of traffic for a while.

They built EFS with more than 100 customers plus AWS, Google Cloud, and Microsoft Azure. Customer-owned storage, customer-managed keys, and automated review (flags go to the customer’s team, no Anthropic human in the loop) are each opt-in. Anthropic is not charging for EFS. Your cloud bill for object storage is another story.

Does a small firm get Anthropic’s zero-data-retention setup?#

No. EFS is rolling out to enterprises Anthropic already treats as eligible. A small firm using Claude in the browser, or an API key on a laptop, does not get customer-owned object storage and a Wells-style review queue. You get the product’s default retention and whatever the consumer or team terms say.

“Claude is private” and “our logs live in our VPC under our keys” are different sentences. The second one is EFS. The first one is marketing.

Do not wait for the fall rollout before you decide where customer names, call notes, and Slack threads sit. For AI for small business, pick the workflow, name the system of record, and decide who sees exceptions. Claude is a model. It is not your CRM.

The same questions I already use on AI data privacy for small business still apply: where is the data, how long after you cancel, can they train on it, and how do you delete it. EFS answers those for banks. It does not answer them for your Jobber notes.

Where should your agent logs live instead?#

In the system you already trust — HubSpot, a shared sheet, Slack itself — with the model as a worker, not the archive. That is the same custody split EFS is selling to Fortune 100s, minus the S3 science project.

Trigger: a ticket, an approval, or an internal question in Slack. Agent action: draft, lookup, or route. System of record: the CRM, helpdesk, or channel you already audit. Human escalation: the person who can actually say yes on money, legal, or an angry customer.

That is the Slack AI Agent I still deploy for companies in the 10–250 range: $3,000–$6,000 one time, you own the setup, you keep paying Slack and model usage, not a per-seat safeguards SKU to me. Claude can sit under it. Your transcripts do not have to wait in Anthropic’s 30-day bucket as the only copy that matters.

Skip an owned agent if Claude is only drafting emails you already review. Buy EFS-shaped thinking — your cloud, your keys, your reviewers — only when a security team will actually look at those flags.

This blog is published by the same class of Discord agent I sell. Anthropic’s September 1 write-up changes the enterprise paperwork. It does not change who should hold your customer file.