Data Handling, Privacy and Confidentiality

The first question any organisation asks about Claude is "what happens to our data?" The answer depends on which plan you are on, which settings are chosen, and what you choose to share. The exam tests the structure of that answer rather than specific retention periods, which Anthropic publishes and updates in its privacy centre and terms. Your job is to know where the dividing lines are and how to behave on each side of them.

Consumer plans versus business plans

Consumer plans (Free, Pro, Max) are governed by Anthropic's consumer terms and privacy policy. Individuals are asked to choose whether their conversations may be used to improve Claude's models. That choice is made in privacy settings and can be changed; it also affects how long conversations are retained, with a longer retention period applying when a user permits training use and a shorter one when they do not. Deleting a conversation removes it from your history and triggers deletion from Anthropic's systems within a published window. Conversations flagged for safety review may be handled differently.

Business plans (Team, Enterprise, and the API used by developers) are governed by commercial terms. Under these, Anthropic does not use customer content to train models by default, the organisation controls retention settings and features, and Enterprise customers can negotiate additional commitments including custom retention and data residency options. Administrators, not individual users, own these decisions.

The practical consequence is simple: company information belongs on a plan governed by commercial terms that your organisation has approved. Using a personal account for work data may breach your employer's policy regardless of your own privacy settings.

Settings worth knowing

• Training preference (consumer plans): the switch that controls whether your conversations may be used to improve models.
• Conversation deletion: per conversation or all at once, from settings.
• Memory controls: view, edit, delete or disable Memory; use incognito chats for conversations that should leave no trace in history or Memory.
• Connector management: see and revoke what Claude can access in other systems.
• Export: you can request a copy of your data.
• Administrator controls (business plans): retention policy, feature enablement, connector allow-lists, audit logs on Enterprise.

Locations and names shift between releases, so know the categories and look them up.

Classifying before you share

Most organisations classify information into tiers such as public, internal, confidential and restricted. Map those tiers to Claude use before you need to decide under pressure. A typical mapping, to be adapted to your own policy:

Tier	Example	Claude use
Public	Published reports, marketing copy	Any approved plan
Internal	Meeting notes, process documents	Business plan only, no public sharing
Confidential	Customer lists, financials, contracts	Business plan with admin-approved settings; minimise what you include
Restricted	Health, payroll, credentials, regulated data	Only where explicitly approved; often not at all

Personal data deserves specific care. Names, contact details, health, financial and employment information about identifiable people are regulated in most jurisdictions. Before including personal data, ask whether the task needs it. Often you can anonymise ("Employee A", "the customer") or summarise, achieving the same result with far less exposure.

Practical rules

1. Minimise. Share the smallest extract that lets Claude do the job. A clause, not the contract; a column, not the database.
2. Anonymise when identity does not matter to the task.
3. Never paste credentials, API keys, passwords or card numbers into a chat, for any reason.
4. Check the plan before uploading company data: is this account governed by commercial terms your organisation approved?
5. Mind the sharing features. Shared links and published Artifacts are public; a confidential chat shared "just to a colleague" via a link is exposed to anyone who obtains the link.
6. Watch connectors. A connector exposes everything your account can see in that system; connect only approved tools and keep write actions confirmed.
7. Clean up. Delete conversations that contained sensitive material once the work is done, and keep Memory tidy.
8. Ask, do not assume. If the data policy is unclear, ask the owner of the data or your information security team.

Scenario: the customer churn analysis

An analyst wants Claude to help understand why customers cancelled last quarter. The export contains names, emails, company names, contract values and free-text cancellation reasons. Applying the rules: the task needs the reasons and values, not identities. She removes names and emails, replaces company names with industry and size bands, and confirms she is working in the company's Team workspace. She uploads the reduced file to a Project whose membership is the analytics team only. Afterwards she deletes the conversation from her history and stores the finished analysis in the approved reporting location. The insight was identical; the exposure was a fraction.

Exam focus

• Consumer plans give individuals a training-use choice that also affects retention; business plans apply commercial terms with no training by default and administrator control.
• Company data belongs on an organisation-approved plan governed by commercial terms; a personal account is not a substitute.
• Minimise and anonymise: share the smallest extract the task needs, and strip identities where they are irrelevant.
• Shared links and published Artifacts are public; connectors expose whatever your account can access; credentials never go into a chat.
• Know the setting categories (training preference, deletion, Memory, connectors, export, admin controls) and look up current details rather than memorising them.