Back to work
SP Sri Harshita Pilla

[ Bind-Ready // 120-Minute Design Challenge ]

Bind-Ready

Bind-Ready is an AI underwriting assistant where the system prepares and the human decides. In 120 minutes I designed and prototyped an AI workbench that prepares insurance submissions for binding: it gathers evidence, flags conflicts, drafts broker outreach, and holds every final call for the human.

[ Role ]
Product design and prototype
[ Timeline ]
120 minutes, solo, 2026
[ Format ]
Ajaia LLC Design Lead assessment
[ Client Context ]
Marsh & Colby — concept prototype
Explore the live prototype
[ Bind-Ready // Submission Overview ]
Submission Overview. Concept prototype — click to enlarge.
00 / Context

Problem and solution

The Problem Statement

Before a policy can bind, underwriting assistants at Marsh & Colby reconcile policy documents against a coverage checklist in a spreadsheet tool called Bindwise. The sponsor, Owen Restrepo, Director of Operations, believed the checklist itself needed no changes, and arrived with a confident view of the solution. The checklist was never where the time was going.

The Solution

Bind-Ready moves the work to where the time actually goes. One trusted view gathers the scattered evidence, mismatches surface side by side instead of being caught by careful reading, broker corrections start sooner, and every final decision stays an explicit, attributable human act.

01 / Brief

A confident wishlist met a different operational problem.

The brief

The Ajaia LLC Design Lead assessment allowed 120 minutes for everything, including the walkthrough video. I had to decide quickly what deserved to be designed.

Marsh & Colby is a mid-market commercial insurance brokerage. Before a policy can bind, underwriting assistants reconcile policy documents against a coverage checklist in a spreadsheet tool called Bindwise.

The sponsor, Owen Restrepo, Director of Operations, believed the checklist itself needed no changes. He arrived with a confident view of the solution.

[ Owen's Wishlist ]

The initial request pointed toward feature additions around the existing checklist.

  • Smart checklist autofill
  • A cleaner interface
  • An AI chat assistant
  • Mobile access
02 / Evidence

What the evidence said.

Phase 1: Evidence over requests

Challenge

The wishlist pointed at the checklist, and the clock allowed 120 minutes for everything, including the walkthrough video. The risk was designing the confident request instead of the real problem.

Goals

Test the wishlist against the people doing the work: interviews with two underwriting assistants, Anika Shah and Colin Doyle, read against 30 days of usage logs.

Key Outcomes

The interviews and logs made the turning point clear. The checklist was not where the time was going: the hours went to hunting documents, and the days went to waiting on brokers.

“It can take days waiting on a reply.”

[ Anika Shah — Underwriting Assistant, Springfield ]

She hated the broker email loop when a document did not match, and asked for smart autofill.

“The checklist goes fast, I basically know it by memory.”

[ Colin Doyle — Underwriting Assistant, Hartford ]

The slow part was finding documents across broker email, client portal, and fax. He was “hunting through my inbox half the time.” Mismatches hit about a third of submissions, and he caught them by reading carefully.

[ The Logs // 30 Days ]
2.1 days

Median wait for broker replies after Anika flagged a mismatch

9 waits exceeded 4 days

23 / 61

Anika's submissions flagged for coverage-limit mismatch

30-day sample

11 vs 3 min

Colin's average document search time versus checklist completion

Source finding cost 3.7× more

19 / 54

Colin's submissions with coverage-limit mismatches

17 also had documents outside the portal

The Synthesis

The interviews and logs agree that mismatches are frequent and the broker wait is the real cost. They disagree with the wishlist: nobody's bottleneck is filling in the checklist. Anika asked for autofill because that was the fix she could picture. The logs show her actual days were lost to mismatches and broker waits.

03 / Direction

Why this is not the wishlist.

[ Not Built ]
  • No autofill-first checklist
  • No AI chat assistant
  • No mobile interface
01 — The Move

Gather scattered evidence

Bring documents from broker email, the client portal, and other channels into one trusted submission view.

Broker emailClient portalOne trusted view
02 — The Move

Surface conflicts directly

Show mismatched values side by side instead of relying on someone to catch the difference through careful reading.

Side-by-sideSource values intact
03 — The Move

Start broker correction sooner

Draft the correction request for human review so the multi-day back-and-forth begins earlier and with clearer evidence.

Drafted outreachHuman review
Colin's “I just read carefully” is exactly the kind of silent, uninspectable risk a binding decision should not depend on.

[ Why The Human Checkpoints Exist ]

Explicit mismatch resolution, document verification, and checkbox-gated confirmation make the decision visible.

04 / Concept

The system prepares.
The human decides.

One principle runs through everything. Every resolution is an explicit human action, attributed by name, and nothing is ever marked ready on its own.

05 / Flow

Four beats from evidence to decision.

The workbench keeps the preparation fast and the moments of judgment unmistakably human.

01

Name the block, not just the status

A readiness banner states “2 issues block bind-ready.” Stat cards track documents, exceptions, and broker requests, while the evidence table keeps every requirement tied to its source and verification state.

System matched is not Verified. System matched means the system associated a document with a requirement. It is not the same as a human Verified review.
Readiness bannerEvidence tableVerification states
[ Flow 01 // Submission Overview ]
01. Submission Overview. Click to enlarge.
02

Compare first. Never choose silently.

Two documents disagree on the coverage limit. Both values appear exactly as received, and the interface says plainly that the system does not choose which one is authoritative.

Application
$2,000,000
Certificate
$1,000,000

A four-step trail follows the broker request from Needs action through Resolved. The draft waits for a human to send, the original comparison remains in history, and the issue can be reopened.

Side-by-side valuesNo silent choiceReopenable
[ Flow 02 // Exceptions ]
02. Exceptions. Click to enlarge.
03

Confirmation is a human act

The evidence checklist must be fully met before the final action becomes available. An unchecked statement gates the button: “I have reviewed the required evidence and confirm this submission is ready to proceed to binding.”

Until the issues are resolved and the underwriter confirms, the action stays disabled.

Checkbox-gatedDisabled until resolved
[ Flow 03 // Final Review ]
03. Final Review. Click to enlarge.
04

Keep behavior inspectable

Every action is designed to land in a reverse-chronological activity log, so resolution work and final decisions have an attributable trail.

The prototype screen shown here is the empty state before actions are taken. It tells the user exactly which events will appear and where.

Activity logAttributable trail
[ Flow 04 // Activity ]
04. Activity. Click to enlarge.
06 / Decisions

What I decided.

These rules turn a screen flow into an authority model. The system can accelerate preparation without inheriting the human's right to decide.

01

Verification stays human

The system never marks evidence verified on its own.

02

Source values stay intact

Every value appears exactly as it came from the source document, with no interpolation.

03

Drafting is not sending

A human sits between every generated draft and every outbound request.

04

Actions leave a trail

The activity log records every action so the path to bind-ready remains inspectable.

07 / Scope

Honest by design, honest about scope.

Prototype boundaries

This concept prototype demonstrates how AI assistance and human judgment can coexist in a consequential workflow. It is a behavior model, not a production underwriting system.

  • Runs on deterministic fixture data
  • Parses no real documents
  • Simulates broker outreach locally
  • Stores no user data
08 / Reflection

The constraint made the idea sharper.

A 120-minute constraint forced the sharpest version of the idea. I cut everything that was not the core loop of prepare, flag, draft, and confirm.

The result is a small system with a clear moral architecture, and that is the part I would carry into production.

“Anika asked for autofill because that was the fix she could picture. The logs show her actual days were lost to mismatches and broker waits. My job was to design for what the assistants need, measured in their own data, and be ready to defend it.”

[ My Answer To Owen — “Why Are You Showing Me Something Else?” ]

[ After Confirmation // Bind-Ready State ]
After human confirmation. The Bind-Ready state, reached only by a person.
Explore the live prototype
↑