Main site

Banking AI agent · Loan origination

Loan-Application & File-Completeness AI Agent

Establish what an application is missing, pull income and obligations from the documents cited to source, and hand an underwriter a complete file — the credit decision and its reasons stay with a person.

Typical delivery
4–6 weeks
Deployment
Your stack
Credit decision
Underwriter
After launch
Agent Care

What this agent does

Completes the file, decides nothing

In

  1. 01

    Take the application as it arrives — branch, broker, portal or origination system — and record the date it was received.

  2. 02

    Take in the documents supplied with it — pay records, tax and trading statements, bank statements and asset evidence.

Reason

  1. 03

    Read income and obligations out of the documents and keep the page, date and line each figure came from.

  2. 04

    Check the figures against each other and against what the applicant stated, and mark where they disagree.

  3. 05

    Measure the file against the product's stated criteria and the documents that product requires.

Decide

  1. 06

    List what is missing, stale or inconsistent, and what a named underwriter has to look at first.

  2. 07

    Hold anything that would amount to a credit judgement, a score, a rank, a price or a reason for decline.

Out

  1. 08

    Hand the underwriter a file with each figure traceable to its document, and the open items named.

  2. 09

    Retain the application, the documents read, the extraction, the checks run, the clock and every correction.

Product statement

The agent completes and evidences the file. Approving, declining, pricing, ranking an applicant and the reasons on any notice stay with named credit staff.

Example workflow

One application, end to end

  1. 1Application receivedBranch, broker, digital channel or the origination system, with the date it arrived recorded
  2. 2Documents readPay records, tax and trading statements, bank statements, asset evidence and identification
  3. 3Income and obligations extractedEach figure kept with the document, page and date it came from, and the applicant's stated figure beside it
  4. 4File checked against the criteriaThe product's stated criteria and required documents, cross-document consistency, staleness and the clock

No human action required

Stages 1 to 4 run before an underwriter opens the file — intake, extraction and the checks finish first. Nothing in that stretch scores the applicant, ranks them or forms a view on the credit.

5DecisionSplits on whether the file is complete and internally consistent
Complete and consistent

Goes to an underwriter as a file ready to work.

Incomplete or contradictory

Held, with the gap named and the clock visible.

Underwriter or credit officer

Reads the file, each figure and the document behind it, then decides the application and owns the reasons on any notice that follows.

Take the file · Correct · Request information

File taken — handed back

  1. 6File handed over or the gap set outPassed to a named underwriter, or the outstanding items are written to the applicant's file with the clock
  2. 7Outcome evaluatedExtraction accuracy against the document, items missed, underwriter corrections, clock breaches and gaps by cohort

Corrections

Figures an underwriter re-reads from the document are counted.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Approving, declining or withdrawing an application.
    • The rate, the limit, the term or the price.
    • The reasons stated on an adverse-action notice.
    • Scoring or ranking an applicant.

  • Automation boundary

    Agent acts unaided

    • ✓Read income and obligations, and cite the document.
    • ✓Check the figures against each other and the form.
    • ✓Measure the file against the product's stated criteria and required documents.
    • ✓Name what is missing or stale and keep the notification clock visible.

    Write actions run only inside the approval boundaries agreed during implementation. The credit call is not one of them.

    • Deciding an income source counts, or how much.
    • Accepting a document that could not be verified.
    • Counteroffers, conditions and exceptions to policy.
  • Changing criteria, extraction rules or clock settings.

Example output

One application file, annotated

Everything the agent extracts is attached to the document and the page it came from.

Completeness output · single application

Illustrative example
Applied for
Fixed-term personal loan
Stated on the form
Salaried, one employer
Documents
4 of 6
File status
Incomplete — two items
Confidence
88%
Credit decision
Not made by the agent

As received

The application as it was submitted and the documents attached to it — nothing on this side is inferred or rewritten.

Evidence used

  • Pay record, page 2
  • Bank statements, 90 days
  • Stated income on the form

Why it is held

Two required documents are absent and the pay record disagrees with the statements.

What the score decides

How hard the underwriter re-reads a figure against its source, not how the credit is decided.

Action

Take the fileCorrectRequest information

Value

Where AI adds value

The same four claims, placed at the point in the workflow where each one applies.

Where the value lands

Value 01 – 04

Every application received

Branch, broker, portal or origination

03Completeness

Apply your own criteria

Measure the file against the product's stated criteria, the documents it requires and the figures the applicant supplied.

01Approved path

Hand over a worked file

Income and obligations arrive extracted and cited to the document, so the underwriter opens on the credit question instead of on data entry.

02Human review

Name the gap while there is time

Missing, stale and contradictory items surface early enough that the applicant can still supply them and the notification duty is met.

04

Build an evidence trail

Retain the application, the documents read, each figure and its source, the checks run, the clock and the underwriter's correction — on both paths.

Integrations

Typical integrations

Five system groups connect to the same agent. Which of them are in scope is decided in discovery.

  • Loan origination

    • Encompass
    • nCino
    • Blend
    • Temenos
    • Origination APIs

  • Core & servicing

    • FIS
    • Fiserv
    • Jack Henry
    • existing customer records

  • Income & obligation data

    • Payroll records
    • employer VOE
    • Tax transcripts
    • bank-statement feeds

Agent

Loan application & file completeness

  • Extracts and cites
  • Names what is missing
  • Hands the file over

  • Documents & credit files

    • Portal and broker upload
    • scanned packs
    • Credit bureau files
    • case management

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

Integration availability depends on the client's existing systems and API access.

Agent controls

Six layers between the model and the credit file

Each control wraps the one inside it. A file clears every layer before an underwriter is asked to work it, and the credit decision sits outside all six.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeReturn the intake queue to unaided staff if evaluations or production signals degrade.Roll back
L5Clock and recordRecord the date received, the documents read, each figure and its source, and every correction.Record
L4Credit-decision gateApproval, decline, price, limit and the reasons on a notice stay with named credit staff.Gate
L3Attribute exclusionFields on your prohibited-basis and proxy list are excluded from what the agent reads and writes, and verification disparity is a standing test.Exclude
L2Consistency checksDocuments are read against each other and against the stated figures, and disagreements are marked.Compare
L1Source citationA figure is not presented unless the document, page and date sit beside it.Cite
Model coreFile assessed — the extracted figures, the document behind each one, what is missing and the confidence
L1 – L2
Decide whether a figure may be shown
L3
Excludes listed fields and tests for disparity
L4 – L5
Keep the decision human and the clock visible
L6
Pulls automation back when signals degrade

How Nestack evaluates it

Evaluate the file the underwriter gets — not only the figures on it.

Coverage runs the whole depth of the workflow, and every layer is cut by slice.

Surface — the file the underwriter opens

  1. E1Extraction evaluationDid each figure match the document it was taken from?
  2. E2Completeness evaluationDid it name every item the product's criteria require?
  3. E3Consistency evaluationWere contradictions between documents raised rather than averaged?
  4. E4Disparity testingDo incompleteness and verification outcomes differ by group?
  5. E5Slice evaluationHow does file quality change across specific application cohorts?
  6. E6Business outcomeHow often was a file sent back, and did any clock run out?

Floor — the decision a credit officer has to defend

Failure modes

Where each failure originates in the agent

Seven failure modes plotted against the five stages of the agent lifecycle.

Agent lifecycle

Direction of processing →

  1. 01 · Intake

    2 modes

    • LA-01

      Application not logged as one

      A worked-up enquiry starts no clock and is never reported.

    • LA-02

      Altered document accepted

      A regenerated pay record reads cleanly and enters the file.

    Stage gathers

    The application, the documents and the date received

  2. 02 · Extraction

    1 mode

    • LA-03

      Variable income averaged wrong

      Bonus, overtime or commission is annualised over the wrong period.

    Stage reads

    Income and obligations, each cited to its document

  3. 03 · Completeness

    2 modes

    • LA-04

      Uneven proof burden

      A verification step falls harder on one group of applicants.

    • LA-05

      Clock spent re-checking

      The file sits in re-extraction while the notification window runs down.

    Stage checks

    The file against the product's stated criteria

  4. 04 · Handover / write

    1 mode

    • LA-06

      Gap handed over as a reason

      A missing document is passed on as though it were the credit reason.

    Stage hands

    The file to an underwriter, or the gap to the file

  5. 05 · Change / Version

    1 mode

    • LA-07

      Silent criteria drift

      A product criterion changes and stale document rules stay in use.

    Stage tracks

    Model, prompt, criteria and document-rule changes

  • Sev-1 · a duty owed to the applicant is missed
  • Sev-2 · a wrong figure reaches the underwriter
  • Sev-3 · the file degrades and work is redone
See the complete Banking failure-mode catalogue

Affected slices

Rework concentrates in the files without a payslip

A salaried file with one employer and a clean pay record is close to solved. What comes back to be redone is the self-employed application, the seasonal one and the photographed pack. Nestack reports by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Self-employed and variable income4.5%2.9×Review
Multiple or seasonal employers3.4%2.2×Review
Photographed and scanned documents2.9%1.9×Watch
Salaried, single employer1.2%0.8×Normal

Bar: returned-file lift vs. single-employer baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

A wrong figure is found by the underwriter, not by the agent

What tells you intake went wrong is the person who works the file next — a figure re-read from the document, an item asked for twice, a file that came back.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Returned files, re-read figures or repeat requests rise in one application type.

  2. 02

    Diagnose

    Read back against the document itself — the page, the extraction rule, the criteria list or the intake route.

  3. 03

    Improve

    The extraction rule or the required-document list is re-approved by credit policy and version-linked.

  4. 04

    Verify

    Replayed over past applications, including the files the underwriter sent back.

  5. 05

    Learn

    The document that beat the extractor is kept, and the disparity test is re-run before release.

  6. Learn → Detect

    The return edge. Changing what the agent asks for changes who finds the process hard, so fair-lending testing runs on the change before it ships.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, intake and extraction, the criteria and clock checks, evaluation, integration, then production validation and handover.

Workstream

Week 1Week 2Week 3Week 4Week 5Week 6

  1. 01Intake workflow and automation-boundary definition.Week 1
  2. 02Origination and document-system assessment.Week 1, Week 2
  3. 03Product criteria and required-document lists.Week 1, Week 2
  4. 04Application intake and receipt-date capture.Week 2
  5. 05Document classification and citation.Week 2, Week 3
  6. 06Income and obligation extraction rules.Week 3
  7. 07Cross-document consistency and staleness checks.Week 3, Week 4
  8. 08Protected-attribute field exclusion.Week 4
  9. 09Notification-clock capture and gap reporting.Week 4
  10. 10Extraction accuracy and disparity evaluation.Week 4, Week 5
  11. 11Underwriter handover and integration.Week 5, Week 6
  12. 12Observability, deployment and Agent Care handover.Week 6

12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel

Final scope and sequence confirmed in discovery

Engagement tiers

What each tier includes

Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.

Capability✓ in scope · — not at this tierPilotOne product, one channelProductionOrigination integrationAdvancedMulti-product / multi-entity
Introduced at Pilot
Extraction cited to the document✓✓✓
Cross-document consistency checks✓✓✓
Completeness against your criteria✓✓✓
Protected-attribute exclusion✓✓✓
Credit decision left to an underwriter✓✓✓
Receipt date, clock and audit trail✓✓✓
Baseline evaluation✓✓✓
Introduced at Production
Origination and workbench integration—✓✓
Observability and evaluation—✓✓
Introduced at Advanced
Multiple products and criteria sets——✓
Multi-entity and enterprise controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on products and criteria sets in scope, origination and document-system integrations, application volume, verification sources, approval controls and deployment requirements.
Separate from build
Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.

What we need from you

What you bring, and what we build with it

Each input maps to a piece of build scope and a week in the delivery timeline.

  1. 01The products in scope and the criteria each one states

    Product criteria and required-document lists

    Week 1

  2. 02Access to origination, core and document systems

    Origination and document-system assessment, then integration

    Week 2

  3. 03A representative set of what applicants actually send you

    Document classification, page-level citation and the extraction baseline

    Week 2

  4. 04How you calculate income for each employment type

    Income and obligation extraction rules

    Week 3

  5. 05Your notification clocks and how a file is closed or withdrawn

    Receipt-date capture, clock tracking and gap reporting

    Week 4

  6. 06The applications your underwriters sent back last year

    Evaluation suite, regression cases and failure-mode testing

    Week 4

  7. 07Named underwriters and whoever owns fair-lending testing

    Handover workflow, disparity testing, then supervised files

    Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Phases are drawn over the weeks they actually occupy. Week 5 carries both the last of the evaluation work and the first rebuilt files.

Phase

W1

W2

W3

W4

W5

W6

  1. DiscoveryW1
  2. BuildW2 – W3
  3. EvaluateW4 – W5
  4. Pilot & LaunchW5 – W6

Week focus

W1The products in scope, their criteria and the automation boundary

W2Intake, receipt-date capture and document classification

W3Income and obligation extraction, cited to the page

W4Consistency and clock checks, evaluation suite and disparity testing

W5Files rebuilt from applications you decided last year, nothing sent

W6Underwriters work from live files under review, then handover

Reading the band
Extraction accuracy and disparity testing are measured in week 4. Week 5 rebuilds applications you already decided, so nothing the agent produces reaches an applicant.
At the end of W6
Underwriters have worked live files the agent prepared, with each figure checked against its document, then Agent Care takes over monitoring.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Banking AI agent

Build an application agent around your own credit criteria.

Show us one product's criteria, the documents your applicants actually send and the files your underwriters sent back. We'll rebuild a month of applications you have already decided, show you what the agent would have called missing, and where the notification clock would have stood.

Typical delivery
4–6 weeks
Deployment
Your origination stack
Credit decision
Always an underwriter
After launch
Agent Care available

Nestack Agents · Loan application & file completeness

AGT-BK-02 · Agent Care available after launch

Build a loan-application agent

Four questions so the call starts from your actual setup — then pick a time that suits you.

About you
Your setup

No obligation. We’ll map the workflow, name the automation boundary and tell you if an agent is the wrong answer.