Claude Certification Guide CCAO-F

Curriculum › Governance, Risk, and Responsible Use · 15% of the exam

Data sensitivity, privacy and organizational policy

What you’ll be able to do

  • Decide what can be shared with Claude based on your organization’s data classification
  • Prepare a file so the analysis works without exposing protected information
  • Choose the approved route — tool, account, connector — and escalate when policy is unclear

What you need to know

Classify before you share

Most organizations sort information into classes — something like public, internal, confidential and restricted — and set rules for what each class may be used for and where it may go. Personal data about customers, employees or students; health or financial information; contract terms; unreleased financial results; and anything covered by a confidentiality agreement usually sits in the stricter classes. Privacy laws such as GDPR and sector-specific rules add obligations on top. You don't need to be a lawyer, but you do need to ask one question before every upload or paste: what class is this, and does policy allow it in this tool?

Redact or anonymize before upload — instructions to the model are not a control

Once data has been shared, typing "don't remember this" or "delete this after you're done" doesn't undo the sharing or satisfy a policy rule. Policy controls act on the data before it leaves your hands, or on the configuration your organization has approved — not on a request to the model. So when a task involves sensitive data:

  • Remove what the task doesn't need. To find themes in complaint emails, Claude needs the complaint text, not names, email addresses or order numbers.
  • Replace identifiers with neutral labels — "Employee A," "Client 3," "Region North."
  • Aggregate or summarize — share counts or ranges rather than individual records where that answers the question.
  • Watch for indirect identifiers. A job title plus a small office plus a start date can point to one person even with the name removed.

The aim is to let the work go ahead safely. Abandoning a useful analysis is rarely necessary when a cleaned version of the data answers the same question.

Use the approved tools and settings

Your organization may have an approved Claude workspace with settings chosen by administrators — which features and connectors are enabled, how data is handled, who can see shared Projects. Work data belongs there, not in a personal account or an unapproved AI tool, even if the personal one is quicker to open. Equally, connectors (for example, Google Drive or Gmail) let Claude read from those sources once you authorize access; connect only the sources your organization allows, and remember that anything you connect is something Claude can read. Shared Projects deserve the same care: knowledge files uploaded to a team Project are visible to its members.

Feature availability and data-handling settings differ by plan and admin configuration, so check your organization's guidance rather than assuming.

Outputs and shared spaces carry data too

Data sensitivity doesn't stop at upload. A summary Claude writes from confidential material is itself confidential, and so is an Artifact built from it. Before you paste an output into a team channel, share a Project with colleagues, or publish an Artifact, ask whether everyone who will see it is allowed to see the information it contains. Similarly, if memory is enabled, think about what you're asking Claude to carry from one conversation to the next, and review what it remembers where your settings allow. Keeping sensitive work in a clearly scoped Project, shared only with the people who need it, is usually safer than mixing it into general chats.

When the policy doesn’t give you an answer

Policies can't anticipate every case. If you're unsure whether a dataset is allowed, whether anonymization is good enough, or whether a new connector is approved, ask your privacy, security, compliance or IT team before proceeding. If you realize you've already shared something you shouldn't have, report it through your organization's process promptly — don't quietly delete the chat and hope. Prompt reporting lets the people responsible take the right steps.

Key concept

Apply data controls before the data reaches Claude — classify it, strip or anonymize what the task doesn’t need, and use only approved tools — because instructing the model afterwards is not a control.

Expect options that upload restricted data "but tell Claude not to retain it," justify sharing as "internal only," or abandon the task. The best answer almost always prepares the data so the task proceeds within policy, or checks with the responsible team when policy is unclear.

Practice scenario

ScenarioA school's pastoral lead wants to use Claude to spot common themes in 200 parent emails about a new homework policy, so she can brief the head teacher. The emails include parents' names, children's names and classes, and a few mention children's health conditions. The school's policy restricts sharing student personal data with external tools without approval.
Work it through, then open this

Theme analysis needs the content of the concerns, not who wrote them. The appropriate step is to remove parent and child names, classes and health details (or replace them with neutral labels) before uploading, using the school’s approved Claude workspace — and to check with the data-protection lead if she isn’t sure the cleaned version is acceptable. Uploading the emails with an instruction not to retain them doesn’t meet the policy; calling it “internal” doesn’t change who the data is shared with; and dropping the analysis wastes a legitimate, achievable task. She should also watch for indirect identifiers, such as a rare health condition in a small class.

Hands-on exercise — Prepare a dataset for safe analysis

Beginner · 15 min

What you’ll learn

  • Spotting direct and indirect identifiers
  • Stripping data down to what the task needs
  • Recording which policy rule applies

  1. Take an invented sample — for example, ten fictional customer-feedback entries with names, emails, order numbers and comments — and write down the analysis you want (e.g. “top three complaint themes”).

    • Why: Being clear about the question tells you which fields you actually need.
    • You should see: A short list of required fields, usually far shorter than the full record.

  2. Make a cleaned copy yourself: remove direct identifiers, replace any you need to track with labels, and look for indirect ones. Only then paste the cleaned version into Claude for the analysis.

    • Why: The control has to act before the data is shared — this is the habit to build.
    • You should see: An analysis that’s just as useful as it would have been with the identifiers.

  3. Look up (or ask your manager about) your organization’s data classification and approved-tools guidance and note which class your real data would fall into.

    • Why: Exam scenarios and real work both turn on following the organization’s own rules.
    • You should see: A clear answer, or a question to take to your privacy or IT team.

Exam traps

Uploading restricted data and telling Claude not to retain it

A request to the model happens after the data is shared and isn’t a policy control. Redact or anonymize first.

Assuming internal use means policy doesn’t apply

Policies govern where data goes and which tools may process it, regardless of who reads the output.

Using a personal account or unapproved AI tool for work data

Work data belongs in the organization’s approved workspace, with the settings its administrators chose.

Skipping the task entirely when redaction would make it safe

When a cleaned dataset answers the same question, abandoning the task is unnecessary.

Thinking removing names alone always anonymizes the data

Combinations of details — role, location, dates, rare conditions — can still identify a person.

Connecting a work data source without checking it is approved

A connector gives Claude read access to that source; connect only what your organization permits.

Sources

Quick check

A team leader has to write annual performance reviews for eight direct reports. A colleague says AI should never be used for anything involving performance. Which approach best balances value and responsible use?