Skip to content
Marcus Gonzalez

Agentic Loan Origination

When AI agents do the work, trust becomes the design problem.

Chase Home Lending is rebuilding loan origination around AI agents to fund loans in days instead of weeks, but AI efficiency gains won't be realized unless our employees trust their work.

Fig. 00 · Illustrative reconstruction · Operating model

AI agents replace manual reconstruction with evidence-based review.

Current workflow — reconstruct the loan by hand
Loan dataone form of many
Illustrative reconstructionSanitized capture pending
New workflow — review evidence and decide
Pipeline · Home equityactivity and actions
Loan Aagent · income verified
Loan BEvaluate condition
Loan Cagent · docs indexed
Loan Dagent · appraisal ordered
Illustrative reconstructionSanitized capture pending
Illustrative reconstruction · production UI and values differFrom manual reconstruction to evidence-based review
Before and after the redesign: a workday spent re-keying data across dozens of forms becomes a queue where AI agents report their work and an employee decides one escalated exception.
Context01

Agentic loan origination platforms are resetting the standard for speed and cost3. After vendor solutions failed to deliver the expected efficiency, Chase chose to build rather than buy.

A loan origination system carries a mortgage from application to closing and stores its documents, data, decisions, and history. AI has expanded what Chase could build now that models can reliably read documents, check data, and enter information to move a loan forward.

I own the employee experience and UI engineering for the loan origination system, and lead two designers and an engineer on the project.

Problem02

Mortgage competitiveness comes down to rates. Today, process and coordination inefficiencies force employees to spend most of their time verifying work instead of moving loans forward — and that overhead shows up directly in our rates.

The process scatters loan information across forms, documents, comments, and personal files. Every touch on a loan starts with piecing together what's already been done, which caps each person's capacity.

Fig. 01 · Illustrative reconstruction · Context fragmentation · ~100 forms1

The system keeps loan data separate from the work.

Loan data spans ~100 forms1
Loan dataone form of many
Illustrative reconstructionSanitized capture pending
The context lives somewhere else
Where the context lives
Documentssupporting evidence
Commentsdecision history
Email chainspartner context
Personal filesworking context
Illustrative reconstructionSanitized capture pending
1
1

The system logs what changed, never why. Employees have to reconstruct that context before they can act.

Illustrative reconstruction · production UI and values differLoan data sits apart from the context it needs
Two panes: loan data distributed across ~100 forms, and the decision context scattered somewhere else across documents, comments, email chains, and personal files.

The workflow depends on employees to advance each loan, but the system doesn't show who's active, what's blocked, or what needs action. Handoffs go unnoticed, loans sit idle, and closing takes longer.

Fig. 02 · Illustrative reconstruction · Invisible queue state

Work can wait without an owner.

Pipeline · Home equityno visible work state
Loan A—
Loan B—
Loan C—
Loan D—
Illustrative reconstructionSanitized capture pending
Illustrative reconstruction · production UI and values differA loan stays idle until someone checks the queue
A pipeline of loans shows no activity while a task waits unseen.
Research03

Field observation showed how both failures shaped the workday.

Fig. 03 · Research · Field study

Finding work and piecing together the loan history took longer than the decisions themselves.

46observed hours

23 sessions × 2 hrs · in person

AdvisorUnderwriterProcessorCloser

1

Locating the next task

Employees monitored the pipeline because the system did not direct tasks to them.

2

Reconstructing the loan

Before acting, employees rebuilt the loan's history from comments, documents, and personal records.

Locating the workRebuilding historyDeciding

The imbalanceThe decisions only took minutes.

1

One processor spent ~90% of a two-hour session monitoring the pipeline. They could neither complete their only task nor schedule a follow-up in the origination system, so they copied the task into Salesforce.

2

One underwriter spent ~75% of a two-hour session documenting the loan in a personal spreadsheet, including custom income and debt calculations, before making the initial decision.

Session strip and drawn session are characterizations, not measurementsCobalt marks the human decisions
Forty-six observed hours across 23 two-hour sessions with advisors, underwriters, processors, and closers converge on two findings, locating the next task and reconstructing the loan, while a representative session drawn to scale shows the decision landing as minutes at the end.
Decisions04

I designed the architecture to embed agent evidence throughout the application, so employees have the context they need wherever they pick up the work.

The new system replaces 100+ separately maintained forms with governed task types built from shared components. Four connected surfaces carry that evidence at every level of the work.

Fig. 04 · Method · Four surfaces

From finding the next task to verifying a single value

From finding the next task to verifying a single value

  1. PipelineAcross all loans · find the work

    Shows which loans need attention and the latest agent activity on the rest.

  2. TimelineAcross one loan · understand it

    Turns agent and human activity into a readable history.

  3. TaskAt one decision · review it

    Puts the evidence, rationale, and expected impact behind one required action.

  4. DetailsAt one value · verify it

    Links verified information to its source and shows what still needs verification.

Squares are the system; the cobalt circle is where a person decidesFind · understand · review · verify
The four surfaces drawn as stations on one line of narrowing scope: the Pipeline finds work across all loans, the Timeline explains one loan, the Task presents one decision for a person to review, and Details verifies a single value.

Pipeline: make waiting work visible

The current pipeline makes employees open a loan to see whether anything has changed. I added the latest agent activity, who is working on the loan, and any open tasks to each row. Loans that need a person appear in a separate section, so the queue directs attention instead of demanding constant monitoring.

Fig. 05 · Illustrative reconstruction · Pipeline

Static loan rows become a visible work queue.

Before — no visible work state
Pipeline · Home equityno visible work state
Loan A—
Loan B—
Loan C—
Loan D—
Illustrative reconstructionSanitized capture pending
After — activity and required action
Pipeline · Home equityactivity and actions
Loan Aagent · income verified
Loan BEvaluate condition
Loan Cagent · docs indexed
Loan Dagent · appraisal ordered
Illustrative reconstructionSanitized capture pending
1
1

The design shifts responsibility for surfacing each handoff to the queue, instead of asking employees to poll for work.

Illustrative reconstruction · production UI and values differCobalt marks the one row that needs a person
Before and after the pipeline redesign: flat rows with no status become rows carrying agent activity, with one urgent task that requires an employee.

Timeline: let the loan explain itself

The current system's comments are hard to follow. Showing every agent event would recreate the same problem at a larger scale. I kept each event as a structured record of who acted and what changed, then used AI to summarize older activity while keeping the full record one click away. Before-and-after values show how the loan changed over time.

Fig. 06 · Illustrative reconstruction · Timeline

Structured events preserve the facts while AI makes the history readable.

Loan timelineagent and human activity
  1. Agent · summarized
    Several earlier events, condensed.
  2. Agent · updated income
    statedverified
  3. Advisor · re-locked rate
    expiredre-locked
Illustrative reconstructionSanitized capture pending
1
stated verifiedeach event records the change
1

The summary reduces reading without becoming the record. Employees can still inspect the event that changed the loan.

Illustrative reconstruction · production UI and values differThe summary condenses; it never becomes the record
A loan timeline where an AI summary condenses earlier events and each remaining event records who acted and how a value changed.

Task: put the evidence before the decision

Giving employees more evidence without structure would still leave them to decipher the loan. I standardized the task layout and limited the agent to supplying loan-specific content. Every task follows the same sequence: why a person is needed, what the agent used, how it reached its conclusion, and what the decision will change.

Fig. 07 · Illustrative reconstruction · Task

One task turns scattered context into a reviewable decision.

Before — context spread across the system
Where the context lives
Documentssupporting evidence
Commentsdecision history
Email chainspartner context
Personal filesworking context
Illustrative reconstructionSanitized capture pending
After — task · one reviewable decision
Required actionfixed layout
Evaluate income conditionneeds judgment
Evidence, sources, and rationaleloan-specific
Verified incomesource linked
Application valuediffers
What the decision will changefixed layout
DTI ±Rate ±LTV ±
Illustrative reconstructionSanitized capture pending
1
1

Unplanned work enters the same governed structure as agent-generated tasks, so it does not create a second, undocumented path through the loan.

Illustrative reconstruction · production UI and values differRequired action · evidence · rationale · expected impact
Before and after: loan data, evidence, and history spread across separate parts of the system become one task holding the required action, the evidence and rationale, and what the decision will change.

Details: make every value answerable

I kept the complete field list out of the everyday workflow and reserved Details for verification. Employees can inspect every value, open its source, or filter to the information that has not been verified.

Fig. 08 · Illustrative reconstruction · Details

Verified information links every value to its source.

All fields and their verification status
Detailsverification status
Annual incomeverified · income doc
Property valueverified · appraisal
Reserves documentationnot received
Credit scoreverified · bureau pull
Illustrative reconstructionSanitized capture pending
One click — what still needs verification
Detailsone click
▾ filter: unverified
Reserves documentationnot received

verify before decision

Illustrative reconstructionSanitized capture pending
1
1

Verification becomes an exception path instead of the default way employees navigate the loan.

Illustrative reconstruction · production UI and values differVerified links to a source; open items stay visible
Two panes of the Details surface: every field linked to its verification source, and a one-click filter showing only the information that still needs verification.
Status05

Agent work is only as trustworthy as the record it leaves behind. The team is building toward the first home equity release at the end of October 2026.

Scope
I own the employee experience and serve as the UI engineering lead for the loan origination system.
In motion
The team is defining the trace schema: the contract that determines what AI agents report about their actions. I am designing how that information appears across the interface.
Trajectory
The first release supports home equity lines of credit. The same surfaces and task types will later expand to support purchase and refinance mortgages.
Measuring
After release, the team will measure cost per loan, loan cycle time, and the number of loans each employee can manage.
Evidence and sources

1~100 forms — The current origination system contains more than 100 separately maintained form screens.

246 hours / 23 sessions / 4 roles — We conducted 23 two-hour in-person sessions with advisors, underwriters, processors, and closers. I helped design the study, conducted ~25% of the sessions, and helped synthesize the findings.

3Platform speed and cost claims — Vesta published the supporting efficiency research with partner lenders.

Contact

How do you get someone to trust work they didn't do themselves? Working on this? Let's talk.

marcus@marcusjg.comLinkedIn

Next case →

Employee Pattern Library

The threadEvery screen in this case is built from that system's components, down to the type treatment that distinguishes AI-generated text.