Responsible AI

Broker Assist

Designing a broker-first AI experience that prioritised confidence over novelty.

After the CommBroker redesign, brokers could find information more easily, but they still needed help interpreting policy in specific scenarios. I led design on CommBank's first broker-facing Gen AI pilot, shaping an assistant that answered scenario-based questions with guardrails, attribution, and measurable business value, while helping define the trust model, source strategy, and evaluation criteria for the pilot.

Role

Lead product designer

Team

1 product designer, 1 product owner, 3 engineers, CommBank AI team

200

Brokers in pilot testing

85%

Rated the pilot experience as excellent

1st

Gen AI pilot shaping CommBank's responsible AI approach

Broker Assist interface answering a mortgage policy question with source attribution

The assistant was designed as a guidance layer inside the broker workflow, not a flashy standalone chatbot.

Challenge

Finding information had improved. Applying it still required a phone call.

The CommBroker redesign improved discoverability, but brokers still called support when they needed help interpreting policy in their specific scenario. The remaining problem was not where content lived. It was whether the interface could help translate policy into usable next-step guidance.

What drove support calls

Scenario-based policy clarification
Questions about what happens next in a process
Uncertainty about how to apply the rule correctly
Status and process guidance in edge situations

Why search alone did not solve it

Questions were interpretive, not purely factual
Each scenario had unique context
Static content could not cover every permutation
Wrong answers carried real operational risk

Context

The project only worked because it was built on stronger structure and better content first.

This pilot did not begin with a blank page. It built on the JTBD research, call-driver analysis, and content cleanup that came out of the CommBroker redesign. That sequencing mattered. A Gen AI assistant grounded in inconsistent or legacy content would have amplified uncertainty instead of reducing it.

JTBD foundation

Mapped the difference between finding policy and applying policy to a live scenario.

Call-driver analysis

Focused the pilot on question types that genuinely drove operational cost and broker frustration.

Content readiness

Ensured the assistant had cleaner, approved source material to draw from.

Design implication

The AI needed to act as a guidance layer, not a shinier version of search.

Principles

The product principles were clear from the start: accuracy before fluency, confidence before speed.

In a regulated environment, a confident but incorrect answer is worse than a slower, more careful one. I designed the pilot around strict product principles that prioritised trust, attribution, and operational value over anything that merely felt novel.

Guardrails Responses were constrained to approved content and policies rather than open-ended generation.

Attribution Source visibility acted as a primary trust mechanism, not an afterthought.

Call reduction Success was measured by independent resolution, not by engagement or novelty metrics.

Design

The assistant was embedded in the workflow and designed to feel useful, not magical.

I designed Broker Assist as an integrated product surface inside CommBroker. The experience needed to answer scenario-based questions quickly, show where the answer came from, and make brokers more comfortable acting on the information they received.

Broker Assist home screen with suggested prompts — Entry state designed around common intents rather than an empty chat box.

Broker Assist answer with expandable policy source — Attribution was surfaced as part of the answer because trust depended on seeing the policy source.

Broker Assist across desktop, tablet, and mobile — The responsive system kept the assistant useful in the same contexts where brokers already worked.

Guardrails over creativity

The design intentionally constrained the answer space so the experience could stay dependable in a high-risk domain.

Embedded, not standalone

Keeping the assistant inside CommBroker reduced context switching and preserved broker workflow continuity.

Attribution as a feature

Seeing the underlying policy source materially increased broker confidence in the AI output.

Operational definition of success

The north star was a broker who did not need to call support because the product answered the question well enough.

I don't need to call anyone for this.

Pilot feedback that tied directly to the core business objective

Results

The pilot proved there was room for AI when it was introduced with enough discipline.

Brokers described the experience in practical terms: helpful, usable, instant. That tone mattered. In this context, measured adoption language was stronger than excitement. It suggested the assistant was becoming a reliable tool rather than a novelty. That mattered because the point of the pilot was not excitement. It was reducing routine interpretation calls while keeping broker confidence high.

200 Brokers engaged during pilot testing and early evaluation.

85% Rated the pilot experience as excellent in pilot testing.

1st Gen AI pilot that informed CommBank's broader responsible AI posture.

Reflection

The strongest AI design move was not the chat interface. It was everything that made the answer worth trusting.

This project sharpened my view that good AI product design is mostly upstream work: clearer use cases, stronger source content, better trust mechanisms, and tighter success criteria. The model itself is only one part of the experience.

It also reinforced a principle that applies beyond AI. In regulated environments, credibility is not decorative. It is the product. If the user cannot trust the answer, the interface has failed no matter how elegant it looks.

Additional case study

Smart & Social

How I designed shared-money behaviours inside the bank experience, not around it.

Read the next case study