← Back to AI Lab Faraz Khan

✦ Concept · Assisted Lending × AI Agents

FinVista, Agent Era.

In 2023 I designed the app where sales executives run an entire lending journey by hand: 11 steps, 6 products, ~60% faster than the paper flow it replaced. Shipped, measured, real. This page asks the obvious next question: what does that desk look like when AI agents do the processing, and the executive's job becomes judgment?

Built on shipped work Working demo below Runs on the 12 trust patterns

Honesty first: FinVista shipped and its numbers are real. The agent era version is a concept, and I won't invent adoption numbers for it. The design thinking is the claim, and you can operate it yourself below.

The shift: from running the journey to supervising it

The 2023 insight was that lending is assisted: an executive drives, and the customer takes the phone only for the moments the law says must be theirs. Agents don't change that insight. They move the executive one seat up.

2023 · Shipped

The executive runs every step

One screen, one action, 11 steps. The executive types, uploads, verifies, and hands the phone to the customer at consent and signature moments.

  • Speed came from focus: no scrolling, no branching
  • ~60% faster processing, ~40% less training (measured)
  • The customer's moments stayed sacred: photo, consent, e-sign
Concept · Agent Era

Agents run the steps, the executive runs the desk

Filing, verification, and drafting become agent work. The executive supervises a queue: approving what's gated, resolving what's escalated, and stepping in when confidence drops.

  • Every agent action is previewed, gated, or logged
  • Below 70% confidence, the console forces escalation
  • The customer's moments stay exactly where they were: human

The Loan Desk, live

This is the concept as working interface, not slides. Run the morning shift: three agents process four applications, and everything that needs your judgment comes to you. Turn Clerk's dial to change how much runs on its own.

Live demo · The Loan Desk · Tue 09:12 Clerk's autonomy:

Queue · Two-Wheeler · 4 applications

Needs you judgment

Nothing yet. Run the shift.

Action log every move

What stays human, on purpose

FinVista taught me that the handoff is the story. Agents make that lesson sharper, not obsolete. Three kinds of moments never become agent work:

Customer only

Live photo, KYC and bureau consent, OTP, e-signature. The law gives these to the customer, and the interface keeps them ceremonial: plain question, clear consequence, their own hand.

Executive only

Approving gated actions, resolving escalations, overriding an agent, and every conversation that carries the relationship. Judgment is the job description now.

Never automated

Credit decisions and anything touching a protected attribute. Agents prepare the file; they don't decide who deserves money. That line is drawn in the interface, not the policy PDF.

The patterns doing the work

Everything in the demo above is assembled from my AI Trust Patterns Field Guide, where each one runs as its own live demo:

02Intent Preview, before the bureau pull 03The Autonomy Dial, on Clerk 08The Escalation Threshold, at 70% 09The Audit Trail, growing as you play 12Who Did What, on every log line

How I Made This, Step By Step

A taste of the process: the steps, the thinking, and what was on the bench. Same honesty as the rest of the page.

Step 01
Start from the shipped thing

The question came from watching executives run FinVista's 11 steps by hand in 2023. Which of those steps were judgment, and which were just motion?

Bench: the shipped FinVista flows · 5 rounds of validation notes from 2023

Step 02
Sort every step into buckets

Agent work, executive judgment, customer-only. The three buckets became the three surfaces you see above: the queue, the needs-you panel, and the untouchable moments.

Bench: the 11-step journey map · one question per step: judgment or motion?

Step 03
Borrow my own guardrails

No re-deriving from scratch: the intent preview, the dial, the 70% threshold, the audit log all came straight from my Field Guide, already tested as standalone demos.

Bench: my AI Trust Patterns Field Guide · Crux's tested dial, threshold and audit demos

Step 04
Script a believable morning

Four applications, nine actions, three judgment calls. I tuned that ratio deliberately, because the ratio is the argument: agents do most of it, humans decide what matters.

Bench: a scripted event chain · a ratio tuned by hand: 9 actions, 3 human calls

Step 05
Build with AI, keep the wheel

I directed, AI built, and I played the whole shift in a headless browser both ways, Clerk on Auto and on Suggest, until every path held up.

Bench: Claude for the build · vanilla HTML/CSS/JS · Playwright to play the shift both ways

Step 06
Draw the honesty line

Shipped numbers stay real, the concept stays labeled a concept, and the page says out loud that no bank has deployed this. The thinking is the claim.

Bench: labels, links and a footer that says "no bank has deployed this" in plain sight

Built the way it preaches: the machine did the work, the human kept the judgment, and the process left a trail you just read.