Main site

Banking AI agent · HMDA data

HMDA & Lending-Data-Integrity AI Agent

Reconcile every register field back to the system it came from, run the recording and submission clocks all year, and leave the certification with the representative who knows the data.

Typical delivery
4–6 weeks
Deployment
Your stack
Named certifier
Year-round
After launch
Agent Care

What this agent does

Reconciles the register, not the decision

In

  1. 01

    An application is taken, and it is ingested from the LOS beside the pricing, geocoding and credit-file records.

  2. 02

    A field is coded — tract, rate spread, action taken — and each is reconciled to the system it came from.

Reason

  1. 03

    A final action is taken, and the record reaches the register inside the thirty days § 1003.4(f) allows.

  2. 04

    A file is priced, and the rate spread is checked against the APOR table in force that week, not a cached one.

  3. 05

    A quarter is counted against the filing trigger — covered loans and applications combined, purchases excluded.

Decide

  1. 06

    An edit fires in the platform, and it routes to a person to verify against the file, never cleared on a score.

  2. 07

    A small-business register is scoped for the rebuilt 2028 rule — far fewer lenders, first filing in 2029.

Out

  1. 08

    A field is retained with its source system, the exception raised on it and the person who cleared it.

  2. 09

    Execute write actions only inside the approval boundaries agreed during implementation.

Product statement

The agent reconciles and assembles; a representative with knowledge of the data certifies, submits, and resubmits if a value proves wrong.

Example workflow

One register, field to certification

  1. 1Final action recordedLOS record, pricing engine, geocoder or credit-file entry
  2. 2Fields reconciledEach register field tied back to the system of record it was drawn from
  3. 3Register assembledField values, exceptions, edits and confidence
  4. 4Controls appliedSource reconciliation, the § 1003.4(f), sixty-day and 1 March clocks, edit checks and confidence threshold

No human action required

Stages 1 to 4 run unaided and certify nothing — the agent records the decision a person already made, and the certifier's lane opens at the gate.

5DecisionBranches at the confidence threshold
High confidence

Goes to the named certifier to read.

Low confidence

Adds a compliance-manager read first.

Certifier review

The register is held with its field sources, its open exceptions and the confidence.

Approve · Correct · Send to compliance review

Certified — the annual filing only

  1. 6Register systems updatedOnly where write access and approval policy allow it
  2. 7Outcome evaluatedField-defect counts, correction distance, edit outcomes and resubmissions

Corrections

Every certifier correction is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Certifying accuracy and completeness of the register.
    • Submitting the register to the federal agency.
    • Verifying a quality edit against the loan file.
    • Retesting coverage — a March 2026 order may reopen it.

  • Automation boundary

    Agent acts unaided

    • ✓Reconcile each register field to its system of record.
    • ✓Track the § 1003.4(f) recording and § 1003.5 filing clocks for the named owner.
    • ✓Assemble the evidence each platform exception turns on into the review queue.
    • ✓Flag the exceptions and the unmatched fields, and hold.

    Any write happens inside the boundaries agreed at implementation, never ahead of certification.

    • Making or revisiting the credit decision on a file.
    • Determining the action taken where the file is unclear.
    • Deciding what a resubmission covers and changes.
  • Changes to field mappings or automation thresholds.

Example output

One register field, annotated

Everything the agent reconciles is attached to the system it was drawn from — that is the § 1003.6(b) evidence.

Register output · single record

Illustrative example
Loan
Closed-end refinance
Register field
Rate spread, derived against the APOR in force
Value on the LAR
1.87
Source of record
Pricing engine
Confidence
96%
Clock position
Recorded within the quarter

As received

Taken from the pricing engine and the APOR table in force — nothing on this side is written by the agent.

Source facts used

  • Pricing engine record
  • APOR table, that week
  • Credit file disclosure

Why this value

It ties to the pricing engine, not to a value the register re-derived for itself.

What the score decides

Below the configured threshold the field picks up a compliance-manager read before the certifier.

Action

ApproveCorrectSend to compliance review

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 register field

From the system of record

03Reconciliation

Reconcile against source

Draw on the LOS, the pricing engine and the geocoder — HMDA was among the FDIC's most-cited findings in 2025.

01Approved path

Knowledge, not rank, certifies

The register arrives reconciled, so whoever certifies can point at the system behind any field on it.

02Human review

Send review to the exceptions

Exceptions are marked, and § 1003.6(c) shelters a good-faith quarterly error only if it clears before 1 March.

04

Build an evidence trail

The field, the system it was drawn from and the representative who certified stay on the register.

Integrations

Typical integrations

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

  • Origination systems

    • Encompass
    • Empower
    • MeridianLink
    • Byte

  • Core and servicing

    • Fiserv
    • FIS
    • Jack Henry
    • Servicing and portfolio data

  • Filing and edits

    • HMDA Platform submissions
    • LAR format
    • edit reports

Agent

HMDA & lending data integrity

  • Reads the record
  • Reconciles the fields
  • Holds for certification

  • Pricing and geocoding

    • Pricing engine
    • APOR
    • FFIEC geocoder
    • tracts

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

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

Agent controls

Six checks between the model and the register

Six checks, narrowing as they go. What survives the last one is named in the map below.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeNarrow the agent to field reconciliation when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt, field-mapping and clock-configuration changes.Track
L4TraceabilityRecord the source system, the value, the exceptions, the corrections and the certification time.Record
L3Certifier reviewHold the register for the named certifier; it governs release, not whether a certified value is right.Gate
L2Policy guardrailsTest fields against the configured § 1003.4 rules and mappings; a failure returns the field.Restrict
L1Confidence thresholdsRoute low-confidence fields to a compliance-manager read before the certifier sees them.Require review
Model coreRegister assembled — field values, exceptions, edits and confidence
L1 – L2
Test whether a field may stand
L3
Puts the certification in a person's hands
L4 – L5
Keep the field and the system behind it
L6
Reverts to field reconciliation when signals degrade

How Nestack evaluates it

Evaluate the reconciliation workflow — not only the finished register.

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

Surface — the register the agency reads

  1. E1Final-output evaluationDid every field on the register match the system it was drawn from?
  2. E2Step-level evaluationDid the agent use the right systems, field mappings and clock configuration?
  3. E3Tool evaluationDid it read and write the correct loan and the correct field?
  4. E4Confidence calibrationDo low-confidence fields actually attract more certifier corrections?
  5. E5Slice evaluationHow does performance change across specific data fields?
  6. E6Business outcomeHow many fields needed a correction or a resubmission after filing?

Floor — the outcome the institution answers for

Failure modes

Where each failure originates in the agent

Seven failure modes, placed at the stage each one originates.

Agent lifecycle

Direction of processing →

  1. 01 · Retrieval

    1 mode

    • HW-03

      Superseded system of record

      Built on an address the credit file has since replaced.

    Stage gathers

    Loan files, LOS records, pricing and geocode data

  2. 02 · Reasoning

    2 modes

    • HW-04

      Stale benchmark table

      Rate spread derived against a cached APOR vintage.

    • HW-06

      Action-taken mismapping

      A changed LOS workflow maps a denial to withdrawal.

    Stage proposes

    Field values, exceptions, edits and confidence

  3. 03 · Tool / write

    2 modes

    • HW-02

      Edits cleared in bulk

      Quality edits closed without a person verifying them.

    • HW-05

      Unreviewed field drift

      A resubmission changes fields nobody asked to change.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • HW-01

      Signed without knowledge

      The register is certified by someone who has not opened a file.

    Stage returns

    The register a named representative certifies

  5. 05 · Change / Version

    1 mode

    • HW-07

      Silent mapping regression

      A rule change quietly alters how a field is derived.

    Stage tracks

    Model, prompt, field rules and mapping config

  • Sev-1 · acts outside the boundary
  • Sev-2 · a wrong value reaches the register
  • Sev-3 · a source degrades, the field is held
See the complete Banking failure-mode catalogue

Affected slices

One weak field can force a full-register fix

A register-level accuracy figure can read as clean while a few data fields carry most of the defects — and resubmission is directed field by field. Nestack reports the defect rate by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Rate spread on priced loans7.3%3.7×Review
Action-taken and reason codes5.2%2.7×Review
Census tract and geocoding3.4%1.7×Watch
Retail closed-end originations1.5%0.8×Normal

Bar: field-defect-rate lift vs. retail-origination baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

The loop closes on a regression case

The loop shuts when the resubmission is a regression case, not when it has been explained. That suite is what the next register assembled is measured against.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    A field's defect count crosses the resubmission threshold.

  2. 02

    Diagnose

    The rate-spread column that was wrong for a whole quarter is read back until one cause holds.

  3. 03

    Improve

    Any change goes out numbered, with the fields that caused it attached.

  4. 04

    Verify

    A register case still failing is enough to hold the release back.

  5. 05

    Learn

    One case added to the suite, one line added to the field record.

  6. Learn → Detect

    The return edge. The next detection runs against a suite one case longer.

Typical build scope

Twelve workstreams across six weeks

The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, reconciliation workflow, evaluation, integration, then production validation and handover.

Workstream

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

  1. 01HMDA workflow discovery and boundary definition.Week 1
  2. 02Origination and pricing source review.Week 1, Week 2
  3. 03Field mapping, edit rules and submission clocks.Week 1, Week 2
  4. 04Loan-record ingestion and normalisation.Week 2
  5. 05Reconciliation logic and source binding.Week 2, Week 3
  6. 06Confidence scoring and exception routing.Week 3
  7. 07Certifier review workflow.Week 3, Week 4
  8. 08Filing and origination-system integration.Week 5
  9. 09Field and reconciliation cases.Week 4, Week 5
  10. 10Guardrails and certification controls.Week 4
  11. 11Register-trail instrumentation.Week 5, Week 6
  12. 12Deployment, documentation 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 portfolio, one registerProductionProduction filing systemsAdvancedMultiple registers / entities
Introduced at Pilot
Reconciliation to your systems✓✓✓
Certifier review✓✓✓
Field-accuracy baseline✓✓✓
Introduced at Production
Reporting by data field—✓✓
Certifier review workflow in your systems—✓✓
Approved write-back—✓✓
Origination-system integration—✓✓
Introduced at Advanced
Multi-channel field mappings——✓
Multi-stage compliance review——✓
High application volume——✓
Multi-institution register controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on integrations, workflow complexity, transaction volume, 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. 01Your loan systems and field mappings

    Loan-record ingestion and field mapping

    Week 1

  2. 02Representative filed registers

    Reconciliation baseline, source binding and clock modelling

    Week 2

  3. 03Your edit-resolution and correction rules

    Field mapping, edit rules and submission-clock modelling

    Week 1

  4. 04Access to relevant APIs, feeds or exports

    Origination, pricing and geocode assessment, then integration setup

    Week 2

  5. 05Registers you would not want resubmitted

    Edit cases and failure-mode testing

    Week 4

  6. 06What no certification may cover

    Confidence scoring, exception routing, guardrails and certification controls

    Week 3

  7. 07Named certifiers with knowledge of the data

    Certifier review workflow, then pilot and production validation

    Weeks 5–6

Nothing else is required

Deployment, documentation and Agent Care handover are ours.

Delivery timeline

Four phases across six weeks

Each band spans the weeks its phase actually takes, which is why the fifth carries two of them.

Phase

W1

W2

W3

W4

W5

W6

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

Week focus

W1HMDA workflow discovery, field mapping and the automation boundary

W2Source integration and the reconciliation baseline

W3Reconciliation workflow, confidence logic and certification controls

W4Evaluation suite, edit checks and failure-mode testing

W5Filing integration, a pilot quarter and targeted corrections

W6One filing season run under the compliance manager, then Agent Care handover

Reading the band
Each bar sits on the weeks its work is named in and no others. The week 5 doubling is real work, not padding.
At the end of W6
Once the season validates, Agent Care assumes the running agent.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Banking AI agent

Build a HMDA agent around the person who certifies.

Regulation B's effects test was repealed in July 2026; Regulation C did not move. Show us your loan systems, your field mappings and who certifies — the certification stays with them.

Typical delivery
4–6 weeks
Deployment
Your loan data
Named certifier
Year-round
After launch
Agent Care available

Nestack Agents · HMDA & lending data integrity

AGT-BK-15 · Agent Care available after launch

Build a HMDA agent

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

About you
Your register

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