What you will be able to do
- Check a document's classification and your organization's AI policy before sharing it with Claude
- Cut a file down to the fields a task actually needs, using consistent placeholders in a working copy
- Tell redaction from anonymization, and choose which one a task calls for
- Decide what has to leave a file before it is uploaded for code-executed analysis, including content you can't see on screen
Key concept
Data minimization before upload — Start from the result you need, work out which data that result depends on, and remove everything else from a working copy before Claude sees it. Claude can't expose what it was never given, whatever the plan or settings.
1.Classification and policy come first
Check a document's classification before you share it with Claude, because the classification decides what you are allowed to do with it. Donor lists, contracts that contain personal data, salary bands and pricing sheets covered by a non-disclosure agreement all carry rules about who may see them. Those rules still apply when the reader is an AI tool. Anthropic's consumer guidance names the categories to be most careful with: financial identifiers such as SSNs and card numbers, health records, passwords and login credentials, and confidential business or personal documents.
The tool and plan are part of the same decision. A free account you use to brainstorm event themes does not give the same protection as a paid account with strict data retention. Commercial plans such as Claude for Work are not used for model training by default. Under those plans the customer organization controls its users' data, and Anthropic processes that data on the organization's behalf. Your organization's policy decides which of its accounts may receive which classes of data. Whether a task suits AI at all is a separate question, covered under use-case appropriateness.
Sometimes the policy says nothing about the material you hold. An AI policy might not mention vendor-confidential pricing, for example. That silence does not give you permission. Ask the people who own the policy before you upload. The vendor's NDA still binds you either way, and an answer from administrators takes far less time than undoing a disclosure.
Checkpoint 1 of 5· Check yourself
A program officer wants Claude's help finding giving patterns in a spreadsheet of donor names, gift amounts and contact details. They have a free personal Claude account and their organization's Claude for Work account. What should they do first?
The classification and the organization's policy decide whether the data may be shared and which account may receive it. A plan with stronger protection is part of that decision, but it doesn't replace it, and a policy that says nothing about the material is not permission.
“Match your tool to your task—tools with more protection allow for safer sharing of sensitive data.”Source: academy.claude.com
2.Removing what the task does not need
Most analysis needs far less data than the source file holds. If you want to find patterns in donor giving, Claude needs amounts, dates and segments. It does not need names, email addresses or phone numbers. You decide what to remove by asking three questions of every field: does it contain PII, is it essential to the analysis, and could it be removed or anonymized without losing analytical value.
Use a consistent placeholder for each removed value, not a blank. If every reference to one client becomes "Client A" and every reference to another becomes "Client B", Claude can still follow who did what across a contract or a set of survey responses. Mixing placeholders, or deleting values outright, breaks that link and makes the summary less reliable. Make these edits in a working copy so the original record stays intact.
Removing names may not be enough. Ask what the worst case would be if the data were exposed. Suppose a team has two survey respondents and their free-text comments are reported by team. Even without names, managers can tell who wrote what. Before you run the analysis, combine that group with another or drop it, so the result can't point to a person.
Redaction and anonymization are different operations, and a task may call for either. Redaction takes a value out of the working copy, such as a name, an account number or a clause, and leaves nothing or a placeholder behind. Anonymization goes further: the aim is that no combination of the remaining fields points to a person, which is why the two-respondent team needs more than a deleted name column. Redact when the analysis doesn't need the field at all. Anonymize when it needs the records but not the identities, by swapping names for consistent labels and merging groups too small to hide anyone.
These steps have a fixed order, because each step depends on the one before it:
- Look up the document's classification and the policy rules that apply to it.
- Identify the fields the task actually needs, such as clause text and dates for a contract summary.
- In a working copy, replace or remove the fields you don't need, using consistent placeholders.
- Search the finished working copy for leftovers: names, hidden sheets and columns, or text layers under visual markup.
- Upload only the verified working copy.
Checkpoint 2 of 5· Check yourself
A coordinator plans to ask Claude to compare course completion rates across program cohorts. Their participant spreadsheet has names, email addresses, cohort, enrollment date and completion status. What should the working copy they upload contain?
Start from the result you need. A cohort comparison depends on cohort, dates and completion status, not on who the participants are, so the identifying fields come out of the working copy. A consistent placeholder such as 'Participant A' keeps track of each person without using real names, and a stronger plan doesn't replace removing data.
“you likely don't need names, contact details, or other PII. Work backwards from your actual goal to determine what data is truly necessary”Source: academy.claude.com
Checkpoint 3 of 5· Exam question
A project coordinator must prepare a client contract containing personal data so that Claude can summarize it. The summary only needs the clause text and dates. Which sequence should the coordinator follow?
Correct answer: A — Look up classification and policy rules → Identify fields the summary needs → Replace unneeded fields in a working copy → Search for leftover names and hidden content → Upload the file to Claude
- A. Classification and policy rules come first because they set the standard, then the task defines which fields are needed, then the unneeded fields are replaced in a copy. The search for leftover names and hidden content runs on the edited copy, so the file is checked in its final form before the upload.
- B. The search for leftover names and hidden sheets happens before the replacement. The replacement edits can leave new traces or miss hidden layers, so the file that gets uploaded was never checked in its final form.
- C. Searching for leftover content happens before the needed fields are identified. The search then has no clear standard for what must go, and the later replacement edits are never checked, so personal data can reach the upload unnoticed.
- D. The replacement is done before the document's classification and policy rules are looked up. The rules set what may be shared, so the coordinator may remove the wrong fields or keep ones the policy forbids, and the later search only confirms the wrong edits.
Sources2
3.Uploads for code-executed analysis
Uploading a file for analysis is different from pasting text into a chat. With code execution and file creation, Claude works in a sandboxed computing environment. It writes Python scripts that open and process the CSV, TSV, spreadsheet or other data file you gave it. Files in a project are also available in that environment. A script works on the file's contents, not on the view you had on screen.
The upload is the disclosure, whatever Claude does with the file afterwards. Hidden sheets, hidden columns, comments, and text under a black box drawn with an annotation tool are all still in the file, and a script can read every one of them. Renaming the file, adding a password to one sheet or hiding a column does not remove the data. Removing it means deleting it from the working copy. The leftover check in step 4 matters most here: before you upload, unhide everything and confirm that only the needed fields remain.
Checkpoint 4 of 5· Check yourself
Before a workbook goes to Claude for code-executed analysis of quarterly refunds, its client name and email columns sit hidden on the main sheet. Which preparation keeps those details from being shared?
Claude's Python scripts read the file that was stored, not the view you had on screen, and the upload is the disclosure whatever Claude does with the file afterwards. Hiding columns, adding a password or renaming the file leaves the data in place. Only deleting it from a working copy removes it.
“What would you remove or modify before sharing with AI?”Source: academy.claude.com
Checkpoint 5 of 5· Exam question
An operations analyst uploads an Excel workbook so Claude can run code to compute monthly refund totals. The 'Refunds' sheet is clean, but the workbook also holds a hidden sheet of customer card-holder names and a hidden column of phone numbers. Policy bars personal data from Claude. What went wrong?
Correct answer: B — Code execution can read the whole workbook, including hidden sheets and columns, so those needed removing from a working copy before upload.
- A. Incorrect. Renaming the extension does not convert the file or limit it to one sheet, and it can break the analysis.
- B. Correct. Hiding is a display setting, not removal. The code can open every sheet and column, so unneeded personal data must be deleted before upload.
- C. Incorrect. Hidden sheets are part of the file and are readable by code, so the personal data was shared on upload.
- D. Incorrect. A password would be a different matter, and relying on hiding or passwords is not data minimization. The clean fix is to remove the data not needed.
Sources5
Sources
(5)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“we also recommend being thoughtful about sharing highly sensitive personal details such as”
↩︎ Classification and policy come first - 2.https://academy.claude.com/courses/ai-fluency-for-nonprofits/understanding-privacy-and-dataOfficial docs
“Match your tool to your task—tools with more protection allow for safer sharing of sensitive data”
↩︎ Classification and policy come first“For pattern analysis, you likely don't need names, contact details, or other PII.”
↩︎ Removing what the task does not need“Which information could be removed or anonymized without losing analytical value?”
↩︎ Removing what the task does not need“What's the worst-case scenario if this data were exposed?”
↩︎ Removing what the task does not need“You can often remove identifying information entirely”
↩︎ Removing what the task does not need“Work backwards from your actual goal to determine what data is truly necessary”
↩︎ Key concept“Match your tool to your task—tools with more protection allow for safer sharing of sensitive data.”
↩︎ Checkpoint“you likely don't need names, contact details, or other PII. Work backwards from your actual goal to determine what data is truly necessary”
↩︎ Checkpoint“What would you remove or modify before sharing with AI?”
↩︎ Checkpoint - 3.
“By default, we will not use your inputs or outputs from our commercial products”
↩︎ Classification and policy come first - 4.https://privacy.claude.com/en/articles/9267385-does-anthropic-act-as-a-data-processor-or-controllerOfficial docs
“You should consult with your organization's administrators regarding Claude usage policies.”
↩︎ Classification and policy come first“You should consult with your organization's administrators regarding Claude usage policies.”
↩︎ Exam trap 2 - 5.
“Code execution and file creation gives Claude a sandboxed computing environment.”
↩︎ Uploads for code-executed analysis“Processing CSV, TSV, and other data files”
↩︎ Uploads for code-executed analysis“Files in your projects are now accessible through Claude's computing environment while remaining in context.”
↩︎ Uploads for code-executed analysis“Creating Python scripts for data analysis”
↩︎ Exam trap 1