AI Challenge Briefs

1. Housing maintenance triage
HMW: How might we help a housing maintenance coordinator prioritise urgent repairs across remote NT communities without “efficiency” quietly pushing remote tenants to the back of the queue?
User: a coordinator drowning in repair requests, with too few trades and huge travel distances. Build: something that reads free-text fault reports, ranks jobs by urgency + safety + logistics, and explains why each job sits where it does. Trust twist: it’s always “cheaper” to fix the town first — make the equity trade-off visible and let the human own it. A tenant should be able to ask why their repair was deprioritised and get a real answer.
2. Grant eligibility assessment
HMW: How might we help a grants officer assess applications consistently and produce reasons a rejected applicant could actually understand and challenge?
User: an officer checking applications against published guidelines, under deadline pressure. Build: a tool that maps each application to each eligibility rule, cites both the rule and the applicant’s supporting text, and drafts reasons for a decision. Trust twist: consistency is good, rubber-stamping is not. Require human sign-off, and don’t penalise applicants who write in plainer or second-language English instead of judging the substance.
3. Information-access (FOI) request triage
HMW: How might we help an information officer find the records relevant to a request while guaranteeing a human reviews anything sensitive before it’s released?
User: an officer processing access requests over large, messy document sets. Build: retrieval that surfaces relevant documents, flags likely personal/sensitive content for human redaction review, and summarises the set. Trust twist: nothing is released or redacted automatically. Reason about error costs — a missed exemption is a privacy breach; over-flagging buries the officer.
4. Risk-based inspection prioritisation
HMW: How might we help a regulator decide where to send limited inspectors first, without the model just re-checking wherever it looked last time?
User: an inspector choosing which premises to visit across a large area with few staff. Build: a risk-ranking model that orders premises and shows the factors behind each score. Trust twist: the feedback-loop trap — if you only inspect where you inspected before, the data will “prove you right” forever. Watch for proxies that unfairly target certain areas or business types.
5. Getting urgent messages to a human, fast
HMW: How might we make sure a vulnerable person’s urgent message never gets buried in a high-volume government inbox and reaches a human quickly?
User: a correspondence/complaints unit triaging a flood of public contact. Build: classification + urgency detection that routes and prioritises items and drafts acknowledgements, with an explanation for each call. Trust twist: the whole design goal is the escalation path — anything flagged serious should reach a human immediately, and the system should escalate, never try to respond itself.
6. Making sense of long documents without over-trusting the AI
HMW: How might we help a caseworker get through a backlog of long policy or case documents faster, without nudging them into trusting a summary they haven’t verified?
User: a caseworker who must read and interpret large volumes of policy or case material. Build: a summariser that links every claim back to the exact source passage, and flags where it’s unsure. Trust twist: speed can breed complacency. Design it so verification is easy and the human stays accountable for what they act on.