Task statement 6.2 · Governance, risk and responsible use

6.2 · Data Sensitivity, Regulation and Privacy

In Wrenfield Housing, the course example, the income team, part of operations, wants Claude to find patterns in rent arrears. Their spreadsheet export holds tenants' names, addresses, rent account numbers and some vulnerability notes, so it cannot go to Claude as it is: AI policy rule 1 keeps tenants' personal data out of prompts unless it is anonymised first.

Why it's tested

Three choices look reasonable here, and each one fails. Uploading the file as it is feels safe because the analysis is internal, but the personal data still leaves the system where it is protected and lands in a chat. Adding "don't keep this data" to the prompt feels like a safeguard, but a sentence to Claude does not change what the service stores or for how long. Skipping the analysis protects the tenants, but it gives up useful work that anonymising would have allowed. The right move keeps the task and removes the people from the data first.

KnowledgePersonal data, sensitive data, and the law

Personal data is any information about a person who can be identified, either directly or by putting details together. An identifier is a detail that points to one person: a name, a home address, an account or reference number, a phone number or an email address. Some kinds of personal data are especially sensitive, such as health, disability or other vulnerability details, because exposing them can cause more harm. Data-protection law, the rules many countries set for handling personal data, generally expects you to use only the data a task needs and to take more care with sensitive kinds. Which rules apply depends on where you work and what the data is. Your organisation's data-protection lead is the person who can tell you, so any answer that treats a legal question as yours to settle alone is wrong.

Anonymising means changing data so that no one in it can be identified. Redacting is the simplest form: removing a detail completely. Work through the file column by column before it reaches a prompt or an upload:

  1. List the columns and write down the question the analysis must answer.
  2. Drop every column the question does not need, such as names and free-text notes.
  3. Code each identifier you need later to act on a result: replace it with a made-up code, and keep the key that links codes to real records outside Claude.
  4. Generalise detail the question needs only roughly: an address becomes an area, a date of birth becomes an age band.
  5. Check a sample of rows: could anyone name a person from one row?

Interactive · what reaches Claude as each step is applied

Five groups of columns from the rent arrears export on the left, each with an arrow labelled by the step applied to it, leading to what reaches Claude on the right. Names and vulnerability notes are dropped. Rent account numbers become codes, and the key stays outside Claude. Addresses become the estate only. Estate, rent, balance, weeks behind and payment method are kept. A column still holding personal data is dashed and labelled personal. EXPORT COLUMN WHAT REACHES CLAUDE Tenant name one per row DROP AS IS REMOVED PERSONAL every tenant named Address house number, street, estate GENERALISE AS IS ESTATE ONLY PERSONAL for example Millbrook points to one home Rent account number needed to act on results CODE AS IS CODE ONLY PERSONAL A0001 to A1480 the real account Vulnerability note free text, some rows DROP AS IS REMOVED PERSONAL, SENSITIVE health and support needs Arrears columns rent, balance, weeks, method KEEP KEPT what the question needs Key: code to account STAYS OUTSIDE CLAUDE · INCOME TEAM Dashed and PERSONAL: a person can be identified.

Step through the arrears export, from the file as it comes out of the rent system to the version that may reach Claude.

As exported: four of the five column groups hold personal data, and the vulnerability notes add health and support details. Uploading this breaks AI policy rule 1.

Drop: names and vulnerability notes go, because the arrears question needs neither. Addresses and account numbers still point to real people.

Code: each account number becomes a code such as A0417. The key stays with the income team, so results can be acted on later. Addresses still identify homes.

Generalise: each address becomes its estate. No row now names a person, and the arrears columns the question needs are all still there.

Dashed and labelled PERSONAL: that column still identifies a tenant. Solid: it no longer does. The key box sits outside Claude on purpose.

A coded file is safer, but it is not automatically anonymous. Whoever holds the key can turn every code back into a person, which is why the key never goes into a chat or a Project. Rare combinations can also point to one person, such as the only household on a small estate paying by a certain method. When the data includes sensitive kinds, or you are unsure whether coding is enough, ask your data-protection lead before you upload.

KnowledgeWhat happens to data you send, and why instructions are not controls

Once you send a message or upload a file, what happens to it is set by your plan's terms and your settings, not by what you ask Claude. Claude cannot promise to forget a chat, because storing it is the service's job, not the model's. By default, a chat you delete leaves your history at once and Anthropic's back-end storage within 30 days. Chats flagged for a policy breach are kept longer. Commercial plans (Team and Enterprise, which Anthropic calls Claude for Work) and consumer plans (Free, Pro and Max) differ on training, so check which plan you are on. Both were checked on 2026-10-15, and the terms can change:

Team, Enterprise

Commercial plans

By default, Anthropic does not use your inputs or outputs to train its models. Exceptions are feedback you send, such as thumbs up or down, and data you choose to share. Chats stay in the product until you delete them; Enterprise can also set custom retention.

Still true: your organisation's policy decides what may be uploaded.

Free, Pro, Max

Consumer plans

The Model Improvement setting in Privacy Settings decides whether your chats are used to improve Claude. With it on, data may be kept in de-identified form for up to five years. Incognito chats are not used, even with the setting on.

Still true: a setting is not permission to upload personal data.

Neither column makes personal data safe to paste. A training setting answers one question: does Anthropic learn from your chats? Data-protection law and your policy ask another: should this data leave its system at all? That is why Wrenfield keeps the shared tenant-enquiries inbox unconnected (5.2), and why the first-try acknowledgement in 4.4 broke rule 1 the moment a tenant's name was pasted in.

SkillWorked example: the arrears export, anonymised before upload

The income officer's export held 1,480 rows, one per tenancy in arrears. The first plan was to upload it whole and add a line to the prompt. The officer checked it against the five steps before sending anything:

# before: the export as it leaves the rent system
Name | Address | Rent account | Estate | Weekly rent | Balance | Weeks behind | Payment method | Vulnerability note
Prompt: "Find patterns in these arrears. Don't keep this data afterwards."
→ 1,480 named tenants, some with health notes, in a chat; the last line changes nothing
# after: dropped, coded, generalised
Code  | Estate    | Weekly rent | Balance   | Weeks behind | Payment method
A0417 | Millbrook | £112.40     | £1,286.00 | 11           | cash
Key (code → rent account): the income team's own folder, never uploaded
→ the same arrears question, and no row that names a tenant

The cause was treating a prompt line as a control. The fix followed the steps. Names and vulnerability notes were dropped, because the question is about patterns, not people. Account numbers became codes A0001 to A1480, and addresses became the estate. Claude found that accounts eight or more weeks behind were mostly paying by cash. It returned the codes behind that pattern, and the officer looked them up in the key, outside Claude. What to do for each household is then a person's decision, as 4.1 set out.

Ask one question: could anyone identify a person from what you are about to paste or upload? Yes → drop, code or generalise it first, and keep any key outside Claude; ask your data-protection lead about sensitive kinds. No → go ahead on your organisation's approved plan.
Takeaway
  • Anonymise before use: keep the task, take the people out of the data.
  • Plan terms and settings decide what happens to your data; a sentence to Claude does not.
  • Health and vulnerability details need the most care, and the data-protection lead settles what the law allows.

check yourself

0 / 3 answered

A charity's fundraising officer wants Claude to find giving patterns in a donor list: names, email addresses, donor numbers, gift amounts and dates, and some notes about bereavements. The charity's policy allows donors' personal data in AI tools only once it has been anonymised. The officer must contact some donors once the patterns are known.What is the most appropriate first step?

Select one.

An HR officer at a 40-person firm wants Claude to spot patterns in a year of sickness absence: name, job title, department, dates and reason, such as "back injury" or "anxiety". They plan to replace names with codes kept outside Claude. Two departments have only two staff each, and each job title there is held by one person.Which change most directly protects the staff while keeping the analysis possible?

Select one.

A training manager wants Claude to summarise common skill gaps across 60 job applications for next year's training plan. A colleague says: "We're on a Team plan, so pasting the CVs as they are is fine." The CVs include home addresses and some health disclosures, and company policy requires anonymising applicants' data first.Which TWO responses best correct the colleague's reasoning?

Select two.

sources

  1. Anthropic Privacy Center, "Is my data used for model training?" (commercial products): by default, inputs and outputs from commercial products such as Claude for Work are not used to train models; exceptions for feedback and data you choose to share. Commercial plans. Checked 2026-10-15.
  2. Anthropic Privacy Center, "How long do you store my organization's data?": chats kept in the product; deleted chats removed from back-end storage within 30 days; feedback kept for 5 years. Commercial plans. Checked 2026-10-15.
  3. Anthropic Privacy Center, "Can you delete data that I sent via Claude for Work?": Enterprise plan customers can set custom retention for their organisation's data. Team and Enterprise. Checked 2026-10-15.
  4. Anthropic Privacy Center, "Is my data used for model training?" (consumer products): the Model Improvement setting in Privacy Settings; Incognito chats are not used to improve Claude. Free, Pro and Max. Checked 2026-10-15.
  5. Anthropic Privacy Center, "How long do you store my data?": with Model Improvement on, de-identified data kept for up to 5 years; deleted chats removed from back-end storage within 30 days. Free, Pro and Max. Checked 2026-10-15.
  6. Claude Help Center, "Claude for Work (Team and Enterprise Plans)": Claude for Work is the name for the Team and Enterprise plans. Checked 2026-10-15.
  7. Wrenfield's arrears export, its columns, the codes and the cash-payment pattern are this course's example.