lender valuation

Designing what AI is allowed to do

A property valuation workflow that automates repetitive work and keeps consequential decisions with the broker.

AI-nativeFintechInteractive experiment
Try the walkthrough ↓

01 / Context

Less administration. Clear responsibility.

The challenge

Brokers were spending approximately 1.5 hours per valuation request, and some broker groups used dedicated teams to manage the administrative workload.

The intervention

I designed a single workflow that reused TeamView data, translated it into lender-specific schemas, surfaced exceptions for broker review and orchestrated submission through a Computer Use Agent.

The responsibility boundary

Preparation can be automated. The broker resolves exceptions, approves the request and reviews the result. Reports return to the originating application.

02 / Approach

AI gets me to more possibilities. Taste decides which ones survive.

I use AI across exploration, prototyping and implementation to get to the point where I can make better decisions faster. Magic Patterns helped me experience rough ideas quickly. Claude Code helped move the design into working frontend, contributing to around 80% of what reached production, as described in the original case study.

Prompt → generate → experience → critique → refine → build → learn → repeat

The most important parts are experience and critique. Does this make the experience simpler? Does the user understand what is happening? Should AI make this decision, or should a human? What can disappear completely?

03 / Experience

Request a valuation without leaving the deal.

Brokers move the same application information between a CRM and lender portals. Interpret it once, translate it for each lender, and return the result with human oversight.

Illustrative simulation · Fictional data · No live submission
AussieAlex Morgan New purchase · Application 24.01.26
TeamView / Purchase / Loan $550,000 / LVR 62.86%

Comparison

Pepper Money

Aussie Activate Package

5.24%

$3,450 / month

Repayment savings $120 / month

Loan fees

Upfront fees $395. Ongoing fees $395 / year.

Loan features

Illustrative two-year fixed, principal and interest, owner occupied. Sample loan $700,000; maximum LVR 80%.

ANZ

ANZ Package

5.30%

$3,500 / month

Repayment savings $70 / month

Loan fees

Upfront fees $0. Ongoing fees $395 / year.

Loan features

Illustrative two-year fixed, principal and interest, owner occupied. Sample loan $700,000; maximum LVR 80%.

HSBC

HSBC Package

5.35%

$3,550 / month

Repayment savings $20 / month

Loan fees

Upfront fees $600. Ongoing fees $395 / year.

Loan features

Illustrative two-year fixed, principal and interest, owner occupied. Sample loan $700,000; maximum LVR 80%.

Sample figures from the reference prototype, not current rates or a loan offer.

04 / Workflow

AI in the product. AI in my process.

AI lowered the cost of trying an interaction. My job was deciding whether it deserved to stay.

AI / ChatGPT + Claude

Explore approaches and edge cases.

Puneet

Define the problem and responsibility boundaries.

Responsibility boundary

Agent may

  • Read documents
  • Structure fields
  • Translate per lender
  • Flag conflicts
  • Report events

Human decides

  • Resolve conflicts
  • Choose the lender
  • Approve submission
  • Review the result

Prompt → build → interact → critique → revise ↺

05 / Decisions

Where the easy answer wasn’t the right one.

Reconstructed trade-offs from the reference case study. These compare plausible AI defaults with my design decisions; they are not quoted model outputs.

01

Autonomous submission vs explicit approval

Prepare automatically. Submit only on explicit approval.

Preparation is reversible; sending to a lender crosses a responsibility boundary. The broker signs off on the exact request that leaves.

  • Prefill request · from the application
  • Fix missing items · by the broker
  • Compliance · confirmed by the broker
  • Ready · waiting for explicit approval

02

Review every field vs focus on exceptions

Lead with exceptions. Keep every field one click away.

Confirming values that already match trains people to click through. Make attention useful, with full details still available.

  • Add a property address
  • Select a property type
  • Confirm the SDR has been checked
  • Confirm the AUP review

03

One loading state vs visible steps and recovery

Show discrete steps and a safe way to recover.

When something fails, brokers need to know what happened, whether anything was sent, and what to do next.

  • Approved request locked
  • Formatted for Horizon Bank
  • Lender unavailable · request not submitted
  • Approved request preserved · retry or return to review

06 / Reflection

The faster I can build, the more carefully I choose what should happen.

Normalise → translate → review exceptions → approve → act → return. The intended benefits are less re-entry and clearer oversight. This prototype demonstrates the interaction model, not measured production impact.

Explore the original full prototype ↗

Next / daelibs design system

Giving a growing product a shared foundation ↗

Back to selected work

ChatPNT

A guide to Puneet’s work

Hi, I’m ChatPNT. Ask me about Puneet’s work, experience or approach to design.

Answers come from curated portfolio content. I don’t generate new facts or send your questions anywhere.

Static knowledge · Private to this session