Integration · security
Written for the person who has to sign the third-party risk assessment. It includes what we have not certified yet, in the same typeface as everything else, because you will find that out in due diligence anyway and finding it here costs you less time.
The honest version, ahead of the questionnaire.
| Control | Status | Detail |
|---|---|---|
| Australian data residency | By design | ap-southeast-2 only; no cross-border transfer or support access |
| Encryption in transit and at rest | In place | TLS 1.3, AES-256, customer-managed keys available |
| Data minimisation | By design | Enforced by the data model, not by configuration |
| Automated decision transparency | Built | APP 1 disclosure generated per claim ahead of 10 Dec 2026 |
| Protected-attribute firewall | Built, tested | Runtime error, not a filter |
| Audit logging | Built | Append-only; reports replayable to a rule-set fingerprint |
| Penetration test | Not yet performed | Scheduled before first production tenant |
| ISO 27001 / SOC 2 Type II | Not certified | Controls designed against both; no audit engaged yet |
| Production operating history | None | No production tenant to date |
Australian claim data stays in Australia. Not "primarily", not "unless support requires otherwise" - the architecture has no path out of the region.
ap-southeast-2 (Sydney).The strongest privacy control available is not holding the data. The data model is built to need very little.
Necessary to answer the question
Claim identifier, product class, jurisdiction, scheme. Claim events with dates. Documents you send us, their classification and extracted facts. Findings and their anchors.
Stays in your system of record
Customer names and contact details unless they appear in a document you send. Reserves, payments and financials. Claim notes. Any copy of your policy administration data.
Firewalled in code
Postcode, ethnicity, language, interpreter need, disability, mental health, Centrelink receipt, credit score, prior AFCA complaint, legal representation - excluded from integrity assessment by a runtime error, not a filter.
| Layer | Control |
|---|---|
| In transit | TLS 1.3, HSTS with preload. mTLS available. AWS PrivateLink where public internet transit is unacceptable. |
| At rest | AES-256. Documents encrypted per tenant with a distinct data key. |
| Key management | AWS KMS in ap-southeast-2. Customer-managed keys supported - you hold revocation, and revoking the key makes your data unreadable to us. |
| Secrets | No credential in source or configuration files. Rotation without downtime; client secrets rotate independently of tokens. |
| In use | Documents are processed in memory and not written to local disk on processing nodes. |
WHERE tenant_id
clause that one missing filter can defeat.events:write
cannot read a document back out. See scopes.You remain accountable for information security regardless of who operates the control. What follows is what we do and what we can evidence, mapped to the paragraphs your assessment will reference.
| Requirement | How it is met |
|---|---|
| Clear roles and responsibilities (para 13) | Named security owner. RACI provided in the due diligence pack. |
| Capability commensurate with threat (paras 15-17) | Controls designed to ISO 27001 Annex A. No external certification yet - see posture. |
| Classification by criticality (paras 18-20) | Claim documents classified as sensitive by default; no lower tier exists for claim content. |
| Third-party assurance (paras 21-22) | Sub-processor list, zero-retention model terms, and this page. Right to audit offered contractually. |
| Testing (paras 27-32) | Automated control tests in CI. Independent penetration test not yet performed - scheduled before first production tenant. |
| Internal audit (paras 33-34) | Your internal audit may review our control design. We have no internal audit function of our own at this stage. |
| Notification (paras 35-36) | Material incident notified within 12 hours, so you can meet your 72-hour obligation with time to assess. See incident response. |
In force since 1 July 2025. If any part of your claims handling depends on us, we are a material service provider and you need the following on file.
| Obligation | Position |
|---|---|
| Tolerance levels for disruption | Agreed per tenant. Because obligation positions are computed rather than stored, a period of unavailability does not corrupt state - positions recompute correctly on recovery. |
| Critical operations mapping | We support claims handling and regulatory reporting. We are not in the payment path and cannot stop a payment. |
| Service provider register | Register entry drafted for you, including sub-processors, in the due diligence pack. |
| Exit plan | Full export of events, documents, findings and reports in the documented JSON schema, on demand and at exit. No proprietary format and no export fee. |
| Business continuity | Multi-AZ within ap-southeast-2. Not yet exercised under production load, because there is no production load yet. |
| Concentration risk | We run on AWS. If that is a concentration your board has already flagged, it applies here too. |
| Principle | Position |
|---|---|
| APP 1 - open and transparent management | The automated decision-making disclosure is generated per claim: what is decided automatically, what is not, what information is used, how a person is involved. Built ahead of the 10 December 2026 commencement rather than after it. |
| APP 3 - collection | We collect only what you send. The data model has no field for information we do not need. |
| APP 6 - use and disclosure | Claim content is used to answer your questions about your claims. It is not used for training, benchmarking, or any cross-tenant analysis. |
| APP 8 - cross-border disclosure | Not engaged. There is no cross-border disclosure of claim content. |
| APP 11 - security | See encryption and access control. |
| APP 12 - access | Where a customer requests their information, every finding about their claim is already anchored to the pages it came from, so an access request is answerable rather than archaeological. |
You remain the APP entity for claim personal information; we are your service provider and act on your instructions. The data processing agreement reflects that.
APRA's April 2026 letter to boards flagged third-party AI risk specifically. Three distinctions matter for your model risk register.
The obligation engine, the five integrity detectors and the seven coverage checks are deterministic code with no model call. Same inputs, same output, every time. This is the part that produces positions and findings.
Document classification and extraction, and drafting a candidate rule from a prompt.
A drafted rule cannot activate itself - it enters review carrying
verified: false and a person activates it.
There is no endpoint that declines, reduces or recovers. recommendation
is hardcoded to "review". This is an architectural property you can
verify in the source rather than a commitment you have to trust.
| Data | Default retention | Notes |
|---|---|---|
| Claim events | 7 years | Configurable to your record-keeping policy; append-only within the period |
| Documents | Your setting | Or hold none - submit by pre-signed URL and we retain extracted facts and anchors only |
| Findings and reports | 7 years | Retained as compliance evidence; deleting a report deletes evidence you may need |
| Audit logs | 7 years | Not deletable by tenant administrators, by design |
Deletion is verified, not merely marked. On exit you get a full export first, then deletion, then written confirmation identifying what was deleted and when.
24/7 alerting on authentication anomalies, egress volume and access-pattern deviation.
Of a material incident affecting your data - deliberately inside your CPS 234 72-hour window so you have time to assess rather than react.
Jointly. You are the APP entity; we provide the forensic detail you need to make the assessment.
Written, shared with you, including what we got wrong. Provided whether or not you ask for it.
Provided on request, under NDA, before any commercial discussion: