An AI that explains your KiwiSaver in plain English at the exact moment you panic — and hands you to a human at the exact moment the law requires one.
Innovate AI Hackathon 2026 · BNZ × AWS × University of Auckland
Challenge 1 — AI-First Customer Experience
The problem
A KiwiSaver balance drops, the app shows a red number with no explanation, and members make the one move that reliably hurts them: switching to cash at the bottom. The FMA measured it during COVID-19:
7×
normal switching rate at the March 2020 low (FMA)
$1.2b
moved into lower-risk funds — locking in losses just before the recovery
$121m
is all that ever moved back
~31%
of lower-risk switches were made by members aged 26–35 — the ones with the longest horizons
April 2025 replayed the pattern: tariff announcements knocked ~10% off global shares in a week, then markets rebounded within days of the pause. Everyone who switched at the low missed the bounce.
2 / 10The insight
Members don't need someone to pick their fund at 11pm on a Tuesday. They need someone to tell them what just happened, whether it's normal, and what it means for their goal — before they act on fear. Today, at exactly that moment, the bank says nothing.
Regulation shapes the answer: recommending a fund is regulated financial advice under the FMC Act. Explaining a balance isn't. The product lives precisely on that line — and treats it as a feature.
3 / 10The solution
Embedded in the BNZ app, the copilot speaks first when something worth explaining happens, and answers the questions members actually ask:
The live prototype next to this deck implements all four — try it.
4 / 10AI approach
Deterministic coreBalance decomposition and projections come from a calculation engine — same inputs, same answer, every time. The LLM never does arithmetic and never sees a number it could get wrong.
Grounded narrationAmazon Bedrock (Claude) receives the computed facts plus retrieved context — the fund's asset allocation, indexed market moves, curated market news — and writes the explanation. RAG, not recall.
Guardrails as complianceBedrock Guardrails encode the FMC Act advice boundary as policy: fund recommendations are structurally blocked and rerouted to adviser handoff. Compliance is architecture, not a system prompt asking nicely.
Auditable by designEvery response logs its grounding facts, retrievals and guardrail decisions — reviewable against CoFI fair-conduct obligations.
Technical architecture
EventBridge watches unit prices and triggers the proactive intercept; DynamoDB holds member context; every layer is serverless, so cost scales with conversations, not headcount.
Full diagram: architecture.html (linked from the demo). The prototype mirrors this shape — static app + serverless LLM proxy — swapping Bedrock for a hosted model only because the hackathon sandbox has no AWS account.
6 / 10Compliance
FMC Act — the advice line“Should I switch funds?” is regulated financial advice. The copilot educates on the trade-offs, then hands to a licensed adviser — it is structurally incapable of recommending.
CoFI — fair conductLetting members panic-switch unaided is a conduct risk. Proactively explaining losses, in plain English, is fair treatment made tangible — and logged.
No hallucinated numbersEvery figure on screen comes from the deterministic engine. The LLM's output is narration over verified facts, so a wrong number can't be generated — only a wrong sentence, which guardrails and evals police.
Human in the loopThe adviser handoff isn't a failure mode, it's the product working: the copilot warms up the conversation and attaches context, so advisers start where the member actually is.
Feasibility
The data exists todayUnit prices, transactions, contributions, fund allocations — the decomposition is a query, not a data project. No new data capture required.
Serverless economicsBedrock + Lambda means cost per conversation measured in cents, no idle infrastructure, and BNZ's existing AWS relationship.
Phased rolloutPhase 1: explain-only copilot in the app (lowest risk, immediate value). Phase 2: proactive intercepts on market events. Phase 3: scenario planning and adviser handoff integration.
Proven in miniatureThis working prototype — deterministic engine, guardrailed LLM, advice-boundary handoff — was built by students in days. The hard part isn't the technology; it's the design discipline, and that's what we're showing.
Impact
$1000s
preserved per intercepted panic switch — compounding for decades in members' balances
26–35
the age group most likely to panic-switch becomes the age group most reachable by an in-app copilot
↑ NPS
the bank that explained the crash beats the bank that sent a red number
Warm leads
every advice-boundary handoff is a qualified adviser conversation with context attached
Measurable from day one: switching rate during drawdowns vs control, copilot engagement after intercepts, adviser conversion, complaint rates. The COVID baseline means the counterfactual is already quantified.
9 / 10The next market shock is a certainty. Whether BNZ members meet it with an explanation or a red number is a choice.
KiwiSaver Copilot · Innovate AI Hackathon 2026 · Challenge 1
Simulated member data · real market events · FMA COVID-19 switching research