AI Governance Policy

1. Purpose of AI

Housing Match uses AI to assist — not replace — the people who make housing and support decisions. AI is applied for three purposes only:

• Compatibility matching — scoring how well a participant's needs, preferences and funding align with a property and its existing household (de-identified data).
• Document extraction — reading uploaded documents to pre-fill participant profile fields, reducing manual data entry (consent-gated; the raw document is processed by an overseas AI provider).
• Administrative tasks — generating plain-language match explanations, summaries and reports for authorised users.

The AI Assistant chatbot answers general platform questions from the published Knowledge Base only and does not access participant documents or profiles — it is a separate feature from document extraction.

AI is never used to determine eligibility for disability supports, diagnose conditions, or decide where a participant will live.

2. Responsible AI Principles

• Human decision-making — AI supports informed decisions; it does not make them.
• Transparency — users are told when AI is used, what it uses, and what it does not do.
• Privacy by design — the minimum data necessary is processed, and de-identified wherever possible.
• Accountability — every AI invocation is logged with an honest de-identification flag.
• Fairness — outputs are reviewed for bias and treated as recommendations, not determinations.
• Participant choice & control — participants can withhold or withdraw consent for AI processing.

3. Approved AI Uses

The following AI uses are approved. Any use outside this list requires governance review and explicit approval.

• Compatibility assessment scoring between participants and properties.
• Housemate compatibility analysis within an existing household.
• Structured data extraction from participant-uploaded documents into profile fields.
• Plain-language generation of match explanations and executive/clinical reports.
• Business development outreach drafting and correspondence review.

4. Human Oversight

Every AI output that influences an outcome is reviewed by an authorised human before any action is taken:

• Providers review and approve or decline AI-generated compatibility assessments and applications.
• Support coordinators and allied health professionals verify AI-extracted profile data against source documents.
• Participants and their support networks can see and challenge AI-generated data on their profile.
• Placements are confirmed by a person, never automatically by AI.

5. AI Limitations

• AI may not capture every clinical nuance in a document; extracted data must be verified by a qualified professional.
• Compatibility scores are recommendations, not guarantees of a successful placement.
• AI outputs depend on the quality and completeness of the underlying data.
• AI can reflect biases present in training data; outputs are reviewed and not treated as definitive.

6. No Clinical Decision-Making

Housing Match is not a healthcare provider, clinical assessment tool, diagnostic tool or medical device. AI does not provide clinical, medical or healthcare advice and does not determine clinical support needs, diagnoses, or treatment. Clinical determinations remain the responsibility of qualified health professionals. Housing Match presents information to support those professionals — it does not substitute for their judgement.

7. Automated Matching, Scoring & AI-Assisted Decision Support

Housing Match uses automated matching, scoring and AI-assisted decision-support tools. Housing Match does not use these systems to independently make final housing placement, provider acceptance or support eligibility decisions. Material housing decisions remain subject to human consideration and participant/provider decision-making.

To be transparent about the role of automation, Housing Match distinguishes between four levels of automated processing:

• Automated processing — computer programs that process data (e.g. document extraction, record creation, status tracking). These do not make or materially contribute to decisions about individuals.
• Automated matching and scoring — deterministic algorithms that score and rank housing options based on stated needs, preferences and funding (e.g. the Compatibility Scoring Engine, Hard Exclusion Logic). These substantially and directly relate to housing decisions but do not make the final decision.
• AI-assisted decision support — AI systems that generate compatibility assessments, match explanations and reports from de-identified data. These provide recommendations that inform human decision-makers.
• Solely automated final decision-making — a decision made by a computer program alone, without meaningful human input, that produces legal or similarly significant effects. Housing Match makes no solely automated decisions about housing, support, or eligibility.

All decisions regarding housing, support, referrals and placements are made by authorised people — participants, support coordinators, allied health professionals, providers and administrators. Where AI or automated systems prepare information that informs a decision, the decision-maker remains responsible for the outcome and can override or disregard the automated output.

A full Automated Decision-Making Register is maintained as part of our compliance evidence framework, documenting each automated function, its purpose, the personal information used, and the human oversight in place.

8. Consent

Consent is granular, informed and withdrawable. Separate consent is captured for:

• Sharing a participant's profile with providers.
• Including a participant in provider matching results.
• Overseas AI processing of uploaded documents (distinct from general sharing consent).

Consent is recorded in an immutable, event-based ledger. Participants can withdraw consent at any time; withdrawal is logged and access is revoked accordingly. Proxy consent (guardian, nominee, administrator) is supported and recorded with its authority basis. Consent is enforced at the database level through Row-Level Security — when consent_to_share or matching_consent is true, authorised providers can access the profile for matching; when false, the profile is invisible at the query level. This is a database security boundary, not a UI check.

9. Data Handling

• Only information relevant to housing, support and compatibility is collected and processed.
• NDIS numbers, Medicare numbers and other government identifiers are not required for matching.
• Each provider sees only their own properties, residents, applications and financials, resolved through providerDataScope and platform record-level security. Providers cannot access another organisation's participant data.
• Participant profiles are protected by server-enforced Row-Level Security — access is restricted at the database level to admins, the participant themselves, their managing professional, the profile creator, or participants who have consented to sharing. A provider can only see a participant who has explicitly consented (consent_to_share or matching_consent). This is a database-level boundary, not a UI convention. Validated August 2026.
• Clinical documents are stored privately and accessed only through authorised, time-limited signed links.
• Uploaded participant documents are stored in Australia (AWS S3, Sydney).

10. De-identification & Pseudonymisation

Housing Match applies de-identification proportionate to the AI task:

• Compatibility assessments — de-identified data is sent to the AI provider: name, NDIS number, email, phone and date of birth are stripped before processing.
• Document extraction — the raw uploaded document is processed by an overseas AI provider. This is not de-identified at the document level, and is therefore consent-gated and transparently logged.
• Audit entries record an honest deidentified flag for every AI invocation.

11. Bias Management

• Compatibility scoring uses transparent, reviewable criteria — not opaque "best match" determinations.
• AI outputs are framed as compatibility options, not rankings of "best" participants or providers.
• Match explanations are generated so decision-makers can understand and challenge the basis of a score.
• Feedback from users is captured and reviewed to identify potential bias in outputs.

12. Audit Logging

Every AI invocation creates an immutable audit-log entry recording the actor, purpose, target, whether the data was de-identified, and the timestamp. Document access, consent changes, profile updates and admin actions are likewise logged. The audit trail is the evidentiary basis for accountability and cannot be altered or deleted by application code.

13. Monitoring

• AI invocations are monitored for volume, purpose and de-identification status.
• Unusual access patterns or unexpected AI usage are reviewable via the audit log.
• Consent withdrawal is enforced and its effect verified.
• Document access links expire automatically (approximately five minutes).

14. AI Review Process

• This policy and approved AI uses are reviewed at least annually, and whenever a new AI use is proposed.
• Proposed new AI uses undergo a privacy and bias assessment before approval.
• Participant, provider and staff feedback informs continuous improvement of AI behaviour and wording.
• Material changes are documented and reflected in Trust Centre and policy updates.

Platform Positioning & Language

We use language that reflects this: AI assists, supports, identifies compatible options and provides decision support. We avoid language such as "AI decides", "AI recommends care" or "best provider". Housing Match uses automated matching, scoring and AI-assisted decision-support tools, but does not use these systems to independently make final housing placement, provider acceptance or support eligibility decisions. Final decisions remain with participants, providers and authorised professionals.