cx-data-minimisation-review
Use to find personal data support collects and stores that nothing actually uses, and to stop collecting it. Trigger for "do we collect too much data", "which fields do we actually use", data minimisa
SKILL.mdskills/compliance/cx-data-minimisation-review/SKILL.md
GitHubData minimisation in support
Support forms and ticket schemas grow. Every incident adds a field, every integration adds a property, every team adds a custom attribute — and nothing is ever removed, because removing a field feels riskier than keeping it.
The result is a system holding personal data with no current purpose. That is both a compliance problem and, more concretely, the single cheapest way to reduce the impact of a future breach: data you do not hold cannot be exposed.
This is the least glamorous analysis in the compliance set and one of the highest return-on-effort, because the findings are usually uncontroversial once someone actually looks.
Inventory, then test for use
List every field where personal data can land:
- Ticket and conversation fields, standard and custom
- Contact and customer record fields
- Intake and contact-form fields, per form and per channel
- Fields populated by integrations
- Fields populated by automation
- Free-text fields, which can contain anything
- Attachments, and what customers are asked to attach
Then, for each, establish whether anything uses it:
- Is it populated? A field empty in 95% of records is either unused or inconsistently used, and both are findings.
- Is it read? Query it in reports, dashboards, automations, routing rules, exports and integrations. A field written by everyone and read by nothing is the target.
- Is it read by a human? A field displayed on a screen nobody looks at is not used.
- When was it last used? A report that has not run for a year does not justify a field.
- Who asked for it, and is that need current? Fields frequently outlive the incident or the person that prompted them.
The pattern to look for is a field that is required at intake, populated for every customer, and read by nothing. Those are pure liability.
The categories that recur
- Identity documents collected "in case" — the highest-risk holding and frequently unnecessary once verification is done. Ask whether the document needs retaining or whether the verification outcome is sufficient. Almost always the latter.
- Full card numbers or bank details in free-text or attachments, where a reference or last four would do. Also a payment-security matter.
- Date of birth, collected for verification and retained indefinitely.
- Full addresses, where a postcode or a country would serve the purpose.
- Phone numbers on channels that never call.
- Health information collected once for a specific accommodation and retained afterwards.
- Optional demographic fields nobody analyses.
- Free-text "additional information" boxes, which reliably accumulate whatever the customer felt like sharing and are almost never read.
- Duplicated fields — the same data in three places because three integrations each created their own.
Collection, not just retention
Retention review asks how long you keep it. Minimisation asks whether to collect it at all, and that is the more effective intervention because it removes the data from every downstream system at once.
For each intake field:
- Is it required? A required field forces every customer to provide it, including the large majority whose case does not need it. Conditional collection — asked only when the case needs it — is usually the right answer and is rarely how forms are built.
- Is it asked before it is needed? Collecting verification data before establishing that verification is necessary.
- Is it asked because a small fraction of cases need it? Ask that fraction.
- Does the customer know why? A field with no explanation gets junk answers as well as being a transparency problem, so this finding usually improves data quality too.
Free text is the hard part
You cannot minimise what customers volunteer, and support conversations are free text by nature. What you can do:
- Do not prompt for more than you need. A field labelled "please describe your medical circumstances" will produce special-category data at volume. Rewording the prompt is the cheapest control available anywhere in this area.
- Do not require detail the case does not need, since customers will over-share to be helpful.
- Know what accumulates, by sampling free-text fields for the categories of sensitive data present. This informs handling and retention even where collection cannot be reduced.
- Handle attachments deliberately. Asking a customer to "send a screenshot" reliably produces screenshots containing far more than the issue.
Removing a field safely
The reason nothing gets removed is that removal feels risky. Make it a bounded change:
- Confirm nothing reads it — reports, automations, routing rules, integrations, exports, and any warehouse model. Check for reads you did not build.
- Stop collecting it first, and leave the historical data in place. This captures most of the benefit for new records with almost no risk.
- Then decide about the history, which is a retention question with its own obligations — and possibly a reason to keep it.
- Check nothing broke, over a period long enough for monthly processes to run.
- Deleting a field's data is usually irreversible, and it may destroy something an obligation requires. Treat it as a reviewed change with a plan, not a cleanup — and get a documented basis before deleting rather than after.
Guardrails
- Do not delete fields or data. This produces a proposal. Deletion is irreversible and may remove records something requires.
- Do not remove a field that a retention obligation or an open matter depends on. Check before proposing, and route the question rather than assuming.
- Identity documents and payment data findings carry higher priority and involve security as well as privacy.
- Special-category data found in free text is a handling and retention finding as well as a minimisation one. Do not quote it; report categories and counts.
- Whether a purpose justifies a field is a legal determination where the answer is contested. Provide the usage evidence.
- Do not extract the data to prove it exists. Field names, population rates and counts are sufficient.
Present results to the user
- The field inventory, with population rate and evidence of whether anything reads it.
- Collected-but-unread fields, ranked by sensitivity then by population — the primary finding.
- Required-at-intake fields that most cases do not need, with a conditional-collection proposal.
- High-risk holdings — identity documents, payment data, date of birth, health — with whether the outcome could replace the artefact.
- Free-text and attachment findings, including prompts that invite sensitive data, with the rewording proposal.
- Duplicated fields across systems.
- A removal plan per field: stop-collecting first, read-dependency check, verification period, and the separate historical-data decision.
- What needs a legal determination or a retention check before anything is removed.