Is Claude safe for client data?

Not by default on the business plans. As of August 2026, Team and Enterprise content is excluded from model training by default, which is the single fact most owners are actually asking about. On individual plans, training preferences are a setting to check rather than an assumption to make.

Verify rather than trust this guide: the current terms are at claude.com, and Anthropic publishes its security and compliance documentation publicly. A fact this important should be confirmed at the source and then written into your policy with the date you checked.

The business tiers carry the controls a data-protection review looks for, and the table below maps them to what they actually do for a small office, as of August 2026.

Which tier a given office needs is a cost question as much as a controls question; the arithmetic is in What Claude actually costs a small business.

From the rollout decisions, far more than from the platform. The failure patterns we design engagements against are ordinary and human:

• Personal accounts. Staff quietly paste client material into consumer AI tools where none of the business-tier protections apply, which is exactly what a ban without an approved alternative produces.
• Over-broad access. A connector pointed at the whole drive when the task needed one folder. Every connection should be scoped to what the work requires.
• No review rule. Output goes straight out because nobody wrote down that it must not. The review step is a rule, not a vibe.
• No offboarding. Someone leaves; their access does not. Business tiers make this fixable in one place, but only if someone owns it.
• Nothing in writing. The office "knows" the rules until the week it turns out everyone knew different ones.

Policy-first means the rules are written, signed, and trained before the tool goes to work, not patched in after an incident. In our engagements the AI use policy is a deliverable with a date: plain language, covering approved tools, allowed data, supervision, and disclosure, and it is signed before go-live.

A good policy is short enough to be followed. If it cannot be read in one sitting by the newest hire, it is a compliance document, not a working rule.

What that document contains, section by section, with a sample clause to adapt for each one, is drafted out in An AI use policy for a small firm.

This is the governance piece of our flat-fee engagement, from $3,500, alongside setup and training. The full scope is on the pricing section of our main page.

The usual entries: material under a protective order or court seal, anything a client has restricted in writing, trade secrets you have contracted to keep siloed, and data a regulator has already ruled on when nobody in the office has read the ruling. Every office has a version of this list; the point is that it is written down rather than intuited.

The excluded list is also allowed to shrink. As controls are verified and comfort grows, categories can move from excluded to allowed-with-review. That direction of travel, cautious first, is a feature of the approach and not a cost of it.

By checking, on a date someone owns. Access reviews confirm the right people and only the right people; audit logs, on plans that carry them, get read rather than merely collected; the training-data and retention settings get re-confirmed rather than remembered.

Justo’s engagements build this in as the day-30 check: usage and access reviewed, the audit trail confirmed, and time saved measured against the baseline captured before rollout. A firm with ongoing obligations can keep the same review running monthly through the managed option. For the law-firm version of this whole analysis, privilege included, read Claude for law firms.