Sruti Gandreti
Starwood

Bringing AI Into Starwood's Loan Boarding Workflow

Accountants were spending 90 minutes per loan boarding deals by hand. This is the AI-assisted validation system I researched and designed that cut that time by 83–89%, while keeping 100% accuracy and every decision in the accountant's control.

Timeline

Jan – April 2026

Role

Design Engineer

With

Vladimir Subbotin

Hanumantha Evuri

SIF Accounting Team

Disciplines

AI Workflow Design

Interaction Design

UX Research

Usability Testing

The Brief

Every time Starwood closes a new infrastructure loan, loan boarding kicks off.

The Starwood Infrastructure Finance accounting team receives a 100-300 page legal binder and manually extracts 50+ data fields from 3 document types (Credit Agreement, Assignment Agreement, Funding Memo) into spreadsheets, later migrated into an internal database.

Old loan boarding process

Every value is entered by hand, and accuracy is non-negotiable. A single error cascades into payment miscalculations, compliance failures, and incorrect financial reporting.

The team wanted AI to help them automate a critical extraction workflow, and my role was to turn that into a system they could reliably depend on. This was a high-visibility initiative, closely observed by executive leadership.

PHASE ONE: Discovery

I started by observing. Three contextual inquiry sessions. One for each document type.

I shadowed an accounting manager as they boarded a real loan, asked minimal questions, and paid attention to everything: what they clicked, what they said, when they hesitated, when they reached for verification.

What I Found

Discovery research statistics

Artifact Analysis

Alongside the shadowing sessions, I reviewed the actual tools accountants worked with.

  • Loan Boarding Screen.xlsm — the primary input file
  • SPT Lenders List.xlsx — entity reference data
  • Sample Credit Agreements, Assignment Agreements, & Funding Memos

This surfaced inconsistent naming conventions, so any AI extraction would need normalization. Reviewing sample docs revealed structural differences between US and UK loans, meaning a one-size extraction logic would fail.

What looked like a data entry problem was actually an orchestration problem.

PHASE TWO: Specification

Before I designed a single screen, I had to become a subject matter expert.

I synthesized everything from Phase 1 into a 56-field deterministic Extraction Rulebook, the document that bridged UX research and engineering.

Extraction Rulebook

For every field the accountants needed to extract, the rulebook specified:

Source pattern 1
Source pattern 2
Source pattern 3

PHASE THREE: Design Exploration

I had to establish what trust actually required when crafting solutions.

My UX research unveiled three major principles that every design variant had to include.

Provenance is a must

Accountants reject any value that appears without a trail. Every AI extraction needed a visible citation, including the source document name and exact page number.

Interest Rate: 4.5%

Source Document

Credit Agreement.pdf — Page 47

Show uncertainty

Time is of the essence in a busy accountant's day. Confidence scores triage their attention.

Confidence: 97.5%

Always editable

Every AI extraction is a suggestion, never a final answer. Accountants can override any field by highlighting correct text in the source text or entering manual values. The system never commits without human approval.

JPMorgan Chase Bank, N.A.

With trust mechanisms locked in as baseline, the real design question emerged:

How should the information flow feel like?

Talking to accountants revealed two competing user needs.

highconfidencespeedcontrol

This lead me to explore two interaction models for testing.

Prototype A: Document-by-Document

Mirrors existing mental model, validating one document at a time

Hypothesis: Familiarity reduces cognitive load and onboarding friction

Prototype B: Consolidated All-in-One

Eliminate context switching, centralizing all extractions and documents in one space

Hypothesis: Reducing navigation overhead enables faster conflict resolution

PHASE FOUR: Evaluative Testing

I conducted a counterbalanced mixed design study with two domain experts via Teams.

Each accounting manager validated four deals using both Prototype A and Prototype B, which I built with v0, Claude Code, and React. The prototypes were tested in a counterbalanced order to minimize order effects and better isolate the impact of the design changes.

Study design

Both prototypes delivered significant gains over the 90 minute manual workflow.

Three validation approaches, three timed outcomes:

Outcomes visual: Manual 90 min, Prototype A 20–25 min, Prototype B 10–15 min

Prototype B was 40–50% faster than Prototype A, cutting an already-optimized workflow nearly in half.

Prototype B reduced validation time by 83–89%.

How did this happen?

I realized in the shadowed sessions that accountants spent most of their time hunting for document conflicts. In Prototype A, this meant:

  1. Validate first Assignment Agreement (v1) → review closing date
  2. Validate second Assignment Agreement (v2) → review different closing date
  3. System determines latest version based on document metadata
  4. Accept final value

Prototype B eliminated steps 1–2. Conflicts were flagged automatically, with the most recent value surfaced.

Prototype A
Prototype B
Prototype A
Prototype B

100% error detection rate across both prototypes

100% error detection rate gradient

Every seeded error was caught. Both proposed workflows made boarding complexity manageable.

  • 8-10 seeded errors per each deal
  • All 75 errors detected by accountants
  • 0 false acceptances

NASA-TLX ratings revealed the cognitive shift.

“I really like how I'm in review mode and not doing the grunt work.” – Accounting Manager

2060100MentaldemandPhysicaldemandTemporaldemandPerformanceEffortFrustration
Manual Workflow
Prototype B

Mental demand

↓45%

82 → 45

Temporal demand

↓55%

85 → 38

Effort

↓46%

78 → 42

Bonus: PM Notifications

When accountants weren't drowning in tab-switching, they had headspace to spot a new pattern: they always notify a Portfolio Manager (PM) after a loan is boarded. If the system could auto-ping, that's another 5–10 minutes saved.

PM Notified

This insight didn't come from interview questions. It emerged because the workflow removed friction.

PHASE FIVE: Implementation & Impact

Recommendation to Engineering: Ship the Consolidated View

1

Q2

Phase 1 (MVP)

Consolidated validation screen

Per-field source transparency

Inline editing

2

Q3

Phase 2 (Add-ons)

Confidence score determination

Automated PM notifications

What difference does this make?

~82 mins

saved per loan

$20k

in reclaimed accountant capacity

210–350 hrs

annual time savings

Reflections

Designing at the Speed of Thought

I learned so much experimenting with using AI for prototyping. Leveraging v0, Claude Code, and React allowed me to bring ideas to life far faster than a traditional Figma-only workflow. Faster prototyping meant quicker feedback loops, reduced design waste, and narrowed focus on what actually works for users. It made me more hopeful of design not being the bottleneck to problem-solving as AI becomes an accelerant for exploration.

Trust-building in Traditional Finance Systems

The unique challenge here was not convincing users to use AI but was designing enough trust into the system so that they don't revert to their old workflow. I spent time understanding emotional and operational safeguards needed during the design process. Studying user behavior closely to see how they validated info manually helped me ideate ways to mirror those behaviors with the product experience. This experience will always remind me that AI products built on giving users confidence will be the most successful.