Main site

Real Estate AI agent · Tenant screening

Tenant-Screening & Adverse-Action AI Agent

Set what the consumer report actually says against the applicant file, flag duplicates, out-of-window records and identity mismatches, run the notice clock the state sets — a named human decides and types the reason.

Typical delivery
4–6 weeks
Deployment
Your stack
Human decides
Never scores
After launch
Agent Care

What this agent does

Assembles the record, never scores the applicant

In

  1. 01

    An application arrives, and the file behind it is pulled from the system of record with each source named.

  2. 02

    A report comes back, and what it says is set line by line against what the applicant file already holds.

Reason

  1. 03

    A proceeding appears twice in one report, and the duplicate is marked rather than counted a second time.

  2. 04

    A record falls outside the lookback its jurisdiction allows, and it is marked out of window, not weighed.

  3. 05

    A middle name is all that ties a record to this applicant, and the identity mismatch is raised first.

Decide

  1. 06

    A portable report is offered in Colorado, and its thirty-day window and no-cost condition are checked.

  2. 07

    A conditional offer stands in New Jersey, and the six factors are assembled as evidence, never as a view.

Out

  1. 08

    A denial is entered by a named human, the notice shell opens on that clock, and disputes join the record.

  2. 09

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

Product statement

The agent assembles the record and refuses to score it; a named human decides the application and types the substantive reason.

Example workflow

One application, report to notice

  1. 1Application and report receivedApplicant file, consumer report, disclosed criteria and the fee record
  2. 2Record assembledReport entries set against the file, each with its source and the version it came from
  3. 3Discrepancies markedDuplicates, out-of-window records, identity mismatches and confidence
  4. 4Controls appliedDuplicate and lookback checks, jurisdiction rules, sequencing tests and confidence threshold

No human action required

Stages 1 to 4 run unaided, and none of them decides anything — the agent is checking the record, and the human lane opens at the confidence gate.

5DecisionForks at the confidence threshold
High confidence

Goes to the named decision-maker.

Low confidence

Adds a compliance read first.

Human decision

The record is held with its flagged entries, its open gaps and the confidence.

Decide · Annotate · Send to compliance review

Decided — notice on the clock

  1. 6Applicant systems updatedOnly where write access and approval policy allow it
  2. 7Outcome evaluatedRecord accuracy, flag outcomes, notice timeliness and post-notice corrections

Annotations

Each annotation made in review is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Scoring, ranking or tiering an applicant.
    • Ordering applicants against one another.
    • The substantive reason typed into a denial notice.
    • Designing or tuning the screening criteria.

  • Automation boundary

    Agent acts unaided

    • ✓Set the consumer report against the applicant file, entry by entry.
    • ✓Flag duplicates, out-of-window records and identity mismatches.
    • ✓Run the notice clock the jurisdiction sets, and mark what has slipped.
    • ✓Assemble the evidence a fair-chance assessment needs, and hold the file.

    Writes land inside the boundaries agreed at implementation, and never ahead of the human.

    • Any accept-or-decline outcome on an application.
    • Predicting denial, late payment or eviction.
    • Inferring a protected characteristic or its proxy.
  • Changes to jurisdiction, clock or criteria rules.

Example output

One screening record, annotated

Everything the agent assembles stays attached to the report version the decider saw, and stays after the notice.

Screening-record output · single application

Illustrative example
Applicant
Two-bed application
Report entry
Eviction filing appears twice under two case numbers
Record date
Filed 2019-03
Source of record
Consumer report
Confidence
88%
Jurisdiction rule
Colorado — seven-year window

As received

Taken from the consumer report and the applicant file — nothing on this side is written by the agent.

Checks run

  • Duplicate-filing check
  • Lookback-window check
  • Identity-match check

Why this is flagged

One proceeding is counted twice, and it sits outside the Colorado window — the human weighs both.

What the score decides

Below the configured threshold the record picks up a compliance read before the decision-maker sees it.

Action

DecideAnnotateSend 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 application

From the applicant file

03Checking

Check the report against the file

Ask what the report says and whether the record belongs to this applicant at all.

01Approved path

It never scores an applicant

Routine applications arrive with the report already set against the file.

02Human review

Send review where the record is wrong

Duplicates, out-of-window records and identity mismatches are marked, so the read starts where the record fails.

04

Build an evidence trail

The report, the mismatch found in it and the person who decided stay on the record.

Integrations

Typical integrations

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

  • Property management

    • Yardi Voyager
    • MRI
    • AppFolio
    • Rent Manager

  • Screening providers

    • TransUnion SmartMove
    • Experian Connect
    • CRA APIs

  • Leasing and applications

    • Entrata
    • RealPage Leasing
    • Online application portals

Agent

Screening record and notice

  • Reads the report
  • Checks the record
  • Holds for decision

  • Notices and delivery

    • Email
    • print-and-mail
    • Document stores
    • e-sign

  • 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 decision

Six filters, tightest last. What none of the six catches is drawn in the map beneath this.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modePull checking back to record assembly only when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt, jurisdiction-rule and clock-configuration changes.Track
L4TraceabilityRecord the report version, the flags, the annotations, the decision and its time.Record
L3Human decisionHold the record for a named decision-maker who can change the outcome; it governs release, not correctness.Gate
L2Policy guardrailsTest records against lookback, sequencing and notice-content rules; a failure returns the record.Restrict
L1Confidence thresholdsRoute low-confidence records to a compliance read before the decision-maker sees them.Require review
Model coreRecord assembled — report entries, flagged defects, open gaps and confidence
L1 – L2
Test whether a record may stand
L3
Puts the decision in a person's hands
L4 – L5
Keep the report and the mismatch behind it
L6
Narrows to record checking when signals degrade

How Nestack evaluates it

Evaluate the whole screening record — not only the notice that leaves.

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

Surface — the notice the applicant receives

  1. E1Final-output evaluationDid every flagged entry trace to a line in the consumer report?
  2. E2Step-level evaluationDid the agent use the right jurisdiction rules, lookback and clock?
  3. E3Tool evaluationDid it read and write the correct application and report version?
  4. E4Confidence calibrationDo low-confidence records actually attract more reviewer annotations?
  5. E5Slice evaluationHow does performance change across specific application and criterion types?
  6. E6Business outcomeHow many records needed an annotation, or a correction after the notice went out?

Floor — the outcome the provider answers for

Failure modes

Where each failure originates in the agent

Seven failure modes, set at the stage where each one first appears.

Agent lifecycle

Direction of processing →

  1. 01 · Retrieval

    1 mode

    • FJ-03

      Superseded report read

      An entry is taken from a report version later replaced.

    Stage gathers

    Applicant file, consumer report and jurisdiction rules

  2. 02 · Reasoning

    2 modes

    • FJ-04

      Checklist reads as a ranking

      Record completeness is read as an ordering of applicants.

    • FJ-06

      Proxy surfaced in the record

      A surfaced field lets voucher status be inferred anyway.

    Stage proposes

    Duplicates, out-of-window records and confidence

  3. 03 · Tool / write

    2 modes

    • FJ-02

      Reason pre-populated

      The notice shell arrives with a reason already in it.

    • FJ-05

      Clock started at the report

      The Oregon fourteen days run from receipt, not denial.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • FJ-01

      Notice element omitted

      A mandated FCRA or state field is missing from the notice.

    Stage returns

    The record the human decides on and the notice sent

  5. 05 · Change / Version

    1 mode

    • FJ-07

      Silent clock drift

      A configuration change shifts a state clock unannounced.

    Stage tracks

    Model, prompt, jurisdiction rules and clock config

  • Sev-1 · a notice leaves outside the boundary
  • Sev-2 · a wrong record reaches the decider
  • Sev-3 · report degrades, record goes to review
See the complete Real Estate failure-mode catalogue

Affected slices

One criterion can carry most of the defects

A portfolio-level record-accuracy figure can read well while one criterion or application type carries most of the defective entries. Nestack reports accuracy by slice, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Criminal-record entries7.6%3.6×Review
Prior-eviction entries5.5%2.6×Review
Common-name applications3.2%1.5×Watch
Credit-only applications2.0%1.0×Normal

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

2 of 4 slices over threshold

Evidence-linked improvement

A cycle is not closed until it leaves a test

The loop shuts when the late notice is a regression case. That suite is what the next application worked is measured against.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Defect rate rises in a criterion slice.

  2. 02

    Diagnose

    The record that belonged to somebody with a similar name is read back through the report until the cause narrows to one.

  3. 03

    Improve

    Any change ships with a number, and the applications attach to it.

  4. 04

    Verify

    One application case still failing is enough to hold the release.

  5. 05

    Learn

    One case joins the suite, one line joins the screening record.

  6. Learn → Detect

    The return edge. Next time detection meets 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, record workflow, evaluation, integration, then production validation and handover.

Workstream

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

  1. 01Screening workflow discovery and boundary definition.Week 1
  2. 02Applicant and report source assessment.Week 1, Week 2
  3. 03Jurisdiction rule, clock and notice-content mapping.Week 1, Week 2
  4. 04Applicant-file ingestion and normalisation.Week 2
  5. 05Report-to-file matching and defect checks.Week 2, Week 3
  6. 06Confidence thresholds and flag routing.Week 3
  7. 07Human decision workflow.Week 3, Week 4
  8. 08Applicant and notice system integration.Week 5
  9. 09Accuracy and notice cases.Week 4, Week 5
  10. 10Guardrails and decision controls.Week 4
  11. 11Application-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 property, one jurisdictionProductionProduction applicant systemsAdvancedMultiple states / portfolios
Introduced at Pilot
Checking to your report and rules✓✓✓
Human decision hold✓✓✓
Record-accuracy baseline✓✓✓
Introduced at Production
Reporting by criterion—✓✓
Decision workflow in your systems—✓✓
Approved write-back—✓✓
Applicant-system integration—✓✓
Introduced at Advanced
Multi-jurisdiction rule sets——✓
Multi-stage compliance reviews——✓
High application volume——✓
Multi-jurisdiction notice 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 application intake and file structure

    Applicant-file ingestion and field mapping

    Week 1

  2. 02Representative decided applications

    Checking baseline, report matching and defect detection

    Week 2

  3. 03Your jurisdiction rules and notice templates

    Jurisdiction rule, clock and notice-content mapping

    Week 1

  4. 04Access to relevant APIs, feeds or exports

    Applicant and report source assessment, then integration setup

    Week 2

  5. 05Denials you would not want re-read

    Notice cases and failure-mode testing

    Week 4

  6. 06What no adverse notice may omit

    Confidence thresholds, flag routing, guardrails and decision controls

    Week 3

  7. 07Named decision-makers to review records

    Human decision 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

The calendar underneath sets every band here, and that calendar is what puts two into week five.

Phase

W1

W2

W3

W4

W5

W6

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

Week focus

W1Screening workflow discovery, rule mapping and the automation boundary

W2Source integration and the checking baseline

W3Record workflow, confidence logic and decision controls

W4Evaluation suite, notice-content checks and failure-mode testing

W5Applicant-system integration, pilot records and targeted corrections

W6One leasing cycle run under the property manager, then Agent Care handover

Reading the band
A bar sits only on the weeks its work is named in. The week 5 overlap is real work, not padding.
At the end of W6
Once the notice cycle validates, Agent Care assumes the agent.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · Real Estate AI agent

Build a screening-record agent around the person who signs the notice.

Show us your applicant file, your screening report and who decides. Not a score, not a rank, not a recommendation — a record checked against the report, a notice shell with the mandated fields in place, and an empty reason field. That is where a federal court landed: the Louis v. SafeRent settlement, finally approved on 20 November 2024, took the screening score off voucher applicants in this market, and California requires the human reviewer of an automated housing decision to hold real authority to change it. The checking is the point, because duplicate filings, out-of-window records and identity mismatches are what the FTC alleged against a tenant-screening bureau in July 2026. Then the clock. FCRA 15 U.S.C. 1681m(a) sets no deadline at all on a landlord adverse-action notice — it says the notice must be given, not when. The thirty days everyone cites is Regulation B, 12 CFR 1002.9(a)(1), a credit rule, and a lease generally is not credit. In the jurisdictions we verified the real clocks are state law: Oregon fourteen days from the denial, Colorado twenty calendar days from the decision to deny with a copy of any consumer report enclosed, Washington a notice with no day count. We run the clock that jurisdiction sets and name what stays with the person on the notice — including what they owe the applicant when a denial already sent rested on the wrong record: a reconsideration, a restored offer, a returned fee, a corrected notice.

Typical delivery
4–6 weeks
Deployment
Your applicant file
Human decides
Never scores
After launch
Agent Care available

Nestack Agents · Tenant-screening records

AGT-RE-17 · Agent Care available after launch

Build a screening-record agent

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

About you
Your applications

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