KiwiSaver Copilot — explains, never advises

Innovate AI Hackathon 2026 · BNZ × AWS × University of Auckland · Challenge 1: AI-First Customer Experience

The problem

When markets fall, KiwiSaver members see a red number with no explanation — and panic. During COVID-19 the FMA measured the damage: switching ran at seven times the normal rate, $1.2 billion moved into lower-risk funds and only $121 million ever moved back, locking in losses just before the recovery. Members aged 26–35 — with the longest horizons to lose — made roughly 31% of those switches. April 2025's tariff shock replayed the pattern. At the moment members most need context, the bank says nothing.

7×

normal switching rate, March 2020 (FMA)

$1.2b

switched to lower-risk funds at the low

$121m

all that ever returned

~31%

of switches made by ages 26–35

The solution

An AI copilot inside the BNZ app that speaks first when a balance drops, explaining what happened in plain English. Balance changes are decomposed deterministically — contributions, employer, fees, tax, markets — so the AI narrates only the market slice, grounded in the fund's real asset allocation. A scenario planner answers "what if I contributed 8%?" and "first home in three years?" with live, deterministic projections. And when a member asks the regulated question — "should I switch funds?" — the copilot educates on the trade-offs, then hands them to a BNZ adviser with the conversation context attached.

Why it's safe

Compliance is architecture, not vibes. The FMC Act advice boundary is enforced in Amazon Bedrock Guardrails: fund recommendations are structurally blocked and rerouted to humans. The LLM never performs arithmetic — every number comes from the deterministic engine, so hallucinated figures cannot reach members. Every response logs its grounding facts and guardrail decisions, supporting BNZ's CoFI fair-conduct obligations.

Feasibility & impact

Everything required — unit prices, transactions, allocations — BNZ already holds; the build is serverless on AWS (API Gateway, Lambda, Bedrock, Knowledge Bases, EventBridge) at cents per conversation. Impact is measurable from day one against the FMA's COVID baseline: fewer panic switches, thousands of dollars preserved per intercepted switch, every handoff a warm, qualified adviser lead. A working prototype accompanies this summary: deterministic engine, guardrailed live LLM chat, adviser handoff.

Simulated member data; the market events and FMA COVID-19 switching research are real. Nothing in the prototype or this summary is financial advice.