An AI document portal that replaces the manual chaos of mortgage underwriting with intelligent, real-time document review.

The Problem

Thirty minutes of manual review per package, every single time.

When a lender submitted a loan package, the assigned underwriter received an Outlook notification and then spent 30 minutes to over an hour manually reviewing each document for completeness and data accuracy before drafting feedback. For full underwriting reviews, the process could stretch to a month, with multiple back-and-forth cycles between the underwriter and the lender.

The problem ran deeper than time. Org charts are complex, multi-modal documents that AI struggled to process reliably, yet they are critical to establishing borrower structure. High-risk documents like appraisals required a second set of eyes, but inconsistencies buried across multiple reports were easy to miss. There was no centralized place to see whether extracted values agreed across documents, or to understand where the AI's confidence was low.

Research

Starting with what the business already knew.

This project did not begin with a blank slate. My Product Owner had already done significant legwork, gathering direct input from business leaders, underwriting managers, and end users to understand which analytics, metrics, and data points needed to surface in the new tool. My role was to take that synthesized knowledge and translate it into a design that replaced the old way of doing things.

The old way was entirely manual. Underwriters tracked deal metrics in spreadsheets, flagged issues in email threads, and built their own informal checklists over time. There was no shared system and no consistent view of what mattered. What the PO's research made clear was not just what information was needed, but how scattered and fragile the existing process was for getting to it.

Service Blueprint

Mapping research into a system view

With the PO's insights in hand, I used service blueprints to map where each piece of information lived in the current process and where it needed to live in the future one. The current state made the problem visible: packages arrived via email, documents were opened one at a time, values were checked against memory, and feedback went back out through unstructured Outlook threads with no record in the platform. Every data point the business cared about had a manual step attached to it.

The future state blueprint reoriented the flow so the platform absorbed that coordination. Instead of underwriters hunting for information, the system would surface it. Instead of email feedback loops, structured requests would close inside MyOptigo. The blueprint gave us a shared language with engineering and product for what had to change and in what order.

Journey Map

Two actors, one broken loop

The journey map traced two parallel tracks: the underwriter and the lender. On the underwriter's side, the friction was tool-switching: receiving a notification in Outlook, opening documents one by one from a local file system, checking values against memory, then drafting feedback back through email. On the lender's side, the friction was opacity: waiting on feedback with no visibility into where their package stood or what specifically needed to change. The two tracks only intersected through unstructured email, creating a loop where context was lost every time it changed hands. The map made clear that fixing one side without fixing the other would just shift the problem.

Underwriter journey from package receipt through structured lender feedback

The Design

AI doing the manual work. Underwriters doing the judgment work.

The designs below focus on the underwriter's side of the journey: the Document Manager portal they use to review packages, surface discrepancies, and issue feedback. This is where the manual process was heaviest: opening documents one at a time, cross-referencing values by memory, and sending ad hoc Outlook emails that left no record in the platform.

The Document Manager replaced each of those steps with an AI-driven one. Extracted values were aggregated automatically. Discrepancies were flagged without the underwriter having to look. Feedback was structured and sent directly from the portal. The design kept underwriters in control throughout: every AI output was visible, editable, and explainable, with the goal of freeing them for the judgment calls only they could make.

The entry point

The Document Manager lives inside MyOptigo, the platform underwriters already use to manage deal lifecycles. From the deal overview, underwriters can see all active tasks, loan trackers, and team assignments before opening the Document Manager directly from the page. The goal was zero context switching: the tool lives where the work already happens.

MyOptigo deal overview: Document Manager accessible directly from the active deal page

Property Summary

Once inside, the Property Summary tab aggregates AI-extracted data from all submitted documents into a single table. Each data point shows an accuracy score, a source document count, and a flag for cross-document inconsistencies. Underwriters immediately see where the AI is confident and where values conflict, without opening a single document manually.

Property Summary: AI-extracted values with accuracy scores and inconsistency flags. Low accuracy warning surfaces for Zip Code and Vendor.

Document Comparison

The Document Comparison tab lets underwriters select which documents to compare side by side, filtered by role. Where extracted values conflict across the Appraisal, PCA, ESA, and Zoning reports, discrepancies are flagged automatically with a banner and highlighted inline. The Zip Code mismatch shown here would previously have required opening four separate documents and checking manually.

Document Comparison (Underwriting): Zip Code conflict surfaced automatically across Appraisal, PCA, ESA, Zoning, and MyOptigo System

Purchase document comparison across Letter of Commitment, Loan Agreement, and Note Rate.

Individual document view with AI-generated review checklist and extracted values panel for the Organizational Chart.

Components

Built for density and precision

The Document Manager required a component set designed for information-dense review environments where underwriters are processing high volumes of structured data under time pressure. Accuracy score badges, inconsistency flags, and cross-document comparison rows had to communicate confidence levels instantly without requiring interpretation. Feedback components were designed to attach directly to specific extraction rows, creating traceable, document-linked requests rather than free-form email threads.

Assumed Outcome

What we expected to move, if it shipped

The Document Manager was designed and validated through research but had not been pushed to production at the time of this case study. Based on what we learned from underwriters and lenders during the design process, the intent was to meaningfully reduce the manual effort tied to document review and cross-referencing.

The structured in-portal feedback loop was designed to replace the Outlook email chain entirely. Partial adoption was intentionally avoided the workflow would only work if it fully replaced the old one, not ran alongside it.

Key Learnings

What this taught me

AI confidence needs a visual language. Underwriters would not act on AI outputs without understanding where the system was certain versus where it was estimating. Accuracy scores and inconsistency flags became the trust layer, not a nice-to-have.

Role-based defaults matter more than feature completeness. Surfacing the right documents first by role reduced cognitive load significantly. The same comparison table felt overwhelming in a one-size-fits-all view and immediately manageable when pre-filtered to context.

Replace the old workflow entirely, not partially. Structured in-portal feedback only worked when it fully replaced the Outlook loop. Partial adoption would have created two parallel systems and more confusion than the original problem.