Back to ClaimsIQ
Data discipline · ClaimsIQ

What we capture, what we don't, and why.

ClaimsIQ is a claims intelligence platform. It exists to help handlers run cleaner claims, give managers a clearer view of the work, and produce a defensible audit trail. None of that requires storing the customer's identity.

The principle

Capture the picture, not the people. The platform sees the cause, the severity, the cover position, the reserve, the vulnerability flag, the cost. It doesn't see the name on the front of the policy.

This isn't a marketing posture. It's an architectural commitment. The PII layer runs across every input on every tool, redacting names, phone numbers, emails, full addresses, account numbers, NHS numbers, NI numbers, dates of birth, vehicle registrations, and other identifiers before they reach the canonical store. Where a handler types "Mrs Patel called from 07700 900123 about her flat at 14 Acacia Road, M14 1AE", the platform stores "[NAME] called from [PHONE] about [ADDRESS], [POSTCODE_TRIMMED:M14]". The tools work just as well on that record. The customer remains a person to the handler, not a data point in a system.

What we never capture

TypeExamplesStatus
Names (with title)Mrs Patel, Mr O'Brien, Dr WilliamsRedacted to [NAME]
Names (first + surname)James Smith, Sarah WilliamsRedacted to [NAME]
Phone numbersUK and internationalRedacted to [PHONE]
Email addressesAny formatRedacted to [EMAIL]
Street addresses14 Acacia Road, 23 High StreetRedacted to [ADDRESS]
NHS numbers485 777 3456Redacted to [NHS_NUMBER]
National Insurance numbersAB 12 34 56 CRedacted to [NI_NUMBER]
Bank detailsSort code, account number, IBANRedacted
Card numbers16-digit primary account numberRedacted to [CARD_NUMBER]
Dates of birthAny format following DOB / bornRedacted to [DOB]
Driving licence numbersStandard UK formatRedacted
Vehicle registrationsAB12 CDERedacted to [VEHICLE_REG]

What we capture, treated carefully

FieldWhat we keepWhy
PostcodeOutward code only by default (M14, not M14 1AE)Regional cost benchmarking and accommodation searches need geography. Full postcodes only stored where a tool's function genuinely requires it (CostIQ regional rates, AAIQ accommodation search) — and they remain server-scoped, never shared across teams.
Cause of lossStandard categories (EoW, Fire, Storm, Subsidence, etc.)Drives every downstream tool. Not an identifier.
SeverityBanded (Minor, Moderate, Significant, Severe)Drives reserve and routing logic.
Property typeBanded (Detached, Semi-detached, Flat, Commercial, etc.)Drives cost benchmarking and accommodation logic.
Reserve estimate / claim costNumeric valuesThe financial picture of the claim.
Vulnerability flagCategory only (cognitive / financial / digital / sensory / capacity / language)Required for FCA Consumer Duty compliance. Specific medical conditions are not captured.
Cover positionSections engaged, exclusions in play, decline rationaleThe senior's read on the claim. Captured against wording references, not against a person's name.
DatesLoss date, FNOL date, milestone datesThe timeline of the work.
Tool outputsVerdicts, scores, findingsWhat the AI flagged, what the handler decided.

Free-text discipline

Handler notes are where personal data leaks most easily — a handler under time pressure types "Mrs Patel said her son James has just left for university". The platform handles this in two layers:

Live, while typing. An amber strip below the field flags any detected identifier. Handler can manually redact at the click of a button.

On commit (blur or save). If any identifier remains, it's automatically redacted in place. The textarea visibly updates so the handler can see what was kept.

Pasted notifications — broker emails, portal submissions, eFNOL exports — get the same treatment before reaching the AI extractor or any storage layer. A green confirmation banner shows what was redacted.

Special category data

Health information (a vulnerability flag pointing at cognitive impairment, or a customer mention of mental health affecting their handling of the claim) falls under Article 9 of UK GDPR. ClaimsIQ captures this as a category — "cognitive" rather than "Alzheimer's diagnosed 2023" — never as a specific medical detail. The handler should record any specific clinical detail in their primary claim system if needed; ClaimsIQ holds only the category needed for proper claim handling under Consumer Duty.

Lawful basis for processing health-category data on insurance claims: Article 9(2)(f) — necessary for the establishment, exercise or defence of legal claims.

Where data lives

ClaimsIQ uses Supabase (PostgreSQL on AWS Ireland) for the canonical store, plus the handler's browser localStorage as a working cache. All claim data is scoped to the handler's sub-team via row-level security; cross-team visibility requires explicit role-based access (manager, lead, analyst).

AI processing for tools that use it (FNOLIQ written extraction, CostIQ benchmarking, hub AI Review) routes through Anthropic's API. The redacted text is what reaches the model — direct identifiers are stripped before any external call.

If you find something we missed

The PII patterns are conservative — false positives (a legitimate piece of text mistaken for an identifier) are preferable to false negatives. If you spot a case where personal data leaked through anyway, the handler-side migration tool (ClaimsIQPIIMigrate.run({dryRun:true}) in the browser console) sweeps your existing claim records and reports anything that needs cleanup. For platform-level concerns, the route is via your sub-team admin — they have visibility into the audit log and can escalate.

Three principles, in summary

01
Capture the picture, not the people. What the claim is about, not who it's about.
02
Redact at every input, not just at output. Personal data shouldn't enter the platform in the first place.
03
Discipline by architecture, not by policy. The pattern enforces itself — no handler training required for it to hold.