Read applications against published criteria with name, school and postcode withheld — because each can stand in for race — and hold every file for the admissions officer, who decides.
Ingest applications, transcripts, essays and recommendations from supported admissions.
02
Normalise transcript formats and field labels, and carry each fact forward with the document it came from.
Reason
03
Withhold name, photograph, secondary school and postcode from the evaluative read.
04
Summarise what the applicant has demonstrated against the criteria the institution publishes.
05
Bind each line of the summary to a passage in the file, and mark what the file leaves ambiguous.
Decide
06
Constrain the output against racial and ethnic descriptors.
07
Route every file to the named admissions officer, and any aid question to the financial aid administrator.
Out
08
Retain the criteria applied, the summary, the reviewer's correction and the release against the application.
09
Execute write actions only inside the approval boundaries agreed during implementation.
→Product statement
The agent proposes a reading; the admissions officer decides the outcome, and an aid determination stays with the financial aid administrator.
Example workflow
One application, file to read
AgentHuman
1Application receivedApplication platform, admissions system, transcript service or SIS record
2Criteria assembledTranscripts, essays, recommendations and activity records, each with its source document
3Read proposedCriteria met, demonstrated strengths, open questions and confidence
4Eligibility checks runRedaction checks, output constraints on racial and disability content, criteria rules and confidence threshold
No human action required
Nothing is ranked or decided in the first four stages, and identifying fields are withheld before the read — the officer's lane opens at the confidence gate.
5DecisionSplits on the review threshold
Clear on criteria
Goes to the admissions officer to weigh.
Criteria unclear
Adds a second reader first.
Officer review
The file is held with its criteria, its evidence spans and the confidence.
Approve · Correct · Send to committee read
Approved — the read goes forward▼
6Admissions systems updatedOnly where write access and review policy allow it
7Read evaluatedCriteria consistency, reviewer corrections, redaction misses and errors found after a decision went out
Corrections
Every reviewer correction is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required8 items
Rendering an admit, deny or waitlist decision.
Exercising professional judgement on need analysis or cost of attendance.
Granting and signing a dependency override.
Deciding that verification documentation is acceptable.
Automation boundaryAgent acts unaided
✓Check a file for completeness and list what is missing for the named owner.
✓Send deadline and document-request reminders.
✓Answer published policy.
✓Deduplicate applicant records and normalise transcript formatting.
Any write happens inside the boundaries agreed at implementation, never ahead of review.
Authorising or releasing a disbursement.
Weighing an applicant's own account of racial experience.
Applying or waiving a legacy, athletic or institutional preference.
Signing off the release of a decision or award batch.
Example output
One applicant read, annotated
Everything the agent proposes is attached to the file it was drawn from.
Admissions output · single applicationIllustrative example
Application
Summary line
Fields withheld
Source of record
Confidence
Race in output
Transfer applicant
Held a term-time job through the programme and led the design team to a regional final
Name · school · postcode
Essay, as submitted
91%
No racial descriptor emitted
As receivedTaken from the file as submitted and read after redaction — nothing on this side is decided by the agent.
Evidence usedEssay, as submittedTranscript on fileRecommendation letter
Why this summaryIt reports what this applicant did and a person still decides and it is logged..
ActionApproveCorrectSend to committee read
What the score decidesBelow the configured threshold the file picks up a second reader before it reaches the officer.
Value
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value landsValue 01 – 04
Every applicationFrom the application file
03Reading
Read from the file
Postcode, school and name can each stand in for race with no protected field present — so the read never sees them.
01Approved path
Read every file the same way
Routine files arrive summarised against the criteria the institution publishes.
02Human review
Send the officer where it matters
Open questions, thin files and anything touching aid are marked, so a reader's time goes where a wrong call is expensive to unwind.
04Build an evidence trail
Each read keeps the criteria applied and the reviewer who signed it.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Admissions systemsSlate by Technolutions Ellucian Recruit
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and a decision
The layers nest. What each one misses is named in the map below it.
L6 · Outermost — last line of defenceInward → L1 · closest to the model
L6Rollback / safe modePull the agent back to completeness checks only when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt, criteria-rule and redaction-configuration changes.Track
L4TraceabilityRecord the file read, the criteria applied, the output, the reviewer and the time.Record
L3Officer reviewHolds the read for the named officer; it governs who decides, not whether a disbursed award can be pulled back.Gate
L2Policy guardrailsTest output against the redaction list and the bar on racial, ethnic and disability descriptors; a failure returns the read.Restrict
L1Confidence thresholdsRoute low-confidence reads to a second reader before the officer sees them.Require review
Model coreRead produced — criteria met, demonstrated strengths, open questions and confidence
L1 – L2Test whether a read may stand
L3Puts the decision in an officer's hands
L4 – L5Show which criteria applied and when they changed
L6Reverts to summary-only when signals degrade
How Nestack evaluates it
Evaluate the whole read path — not only the summary that reaches the officer.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the read the officer sees
Depth of coverage ▼
E1Final-output evaluationDid every line of the summary rest on a passage in the file?
E2Step-level evaluationDid the agent use the right file, criteria and redaction configuration?
E3Tool evaluationDid it read the correct application and write to the correct record?
E4Confidence calibrationDo low-confidence reads actually attract more reviewer corrections?
E5Slice evaluationHow far apart do correction rates sit across specific applicant pools?
E6Business outcomeHow many reads needed correcting, and how many surfaced only after a decision or an award had gone out?
Floor — the decision the applicant receives
Failure modes
Where each failure originates in the agent
Seven modes, each pinned to where it starts.
Agent lifecycleDirection of processing →
01 · Retrieval1 mode
AZ-03
Unredacted file retrieved
The store returns the original, and the read sees name and address.
Stage gathersApplication, transcripts, with the source each came from
02 · Reasoning2 modes
AZ-04
Race rebuilt from proxies
Surname and school become a note about background.
AZ-06
Judgement turned into a formula
A pattern across earlier files is applied as an aid rule.
Stage proposesCriteria met, demonstrated strengths and confidence
03 · Tool / write2 modes
AZ-02
Disability note written to record
Volunteered medical content lands in a structured field.
AZ-05
Read written to the wrong file
An update lands on another applicant's record.
Stage writesOnly where write access and approval policy allow it
04 · Output1 mode
AZ-01
Claim the file does not support
A summary line rests on no passage in the file.
Stage returnsThe read the officer weighs and acts on
05 · Change / Version1 mode
AZ-07
Mid-cycle regression
A model change moves summary phrasing while a cycle is open.
Stage tracksModel, prompt, criteria rules and redaction config
Sev-1 · an award or a decision writtenSev-2 · a wrong read reaches the officerSev-3 · source degrades, file routes to review
The metric is the share of reads a reviewer had to correct before the file moved on. Averaged over every application it looks small; concentrated in the pools whose files carry.
Slice performance — reported separately, not only in aggregateIllustrative example
Slice
Failure rate
Lift
Lift vs. threshold
Status
Files whose essays discuss race
7.6%
3.6×
Review
Files disclosing a disability
5.7%
2.7×
Review
Files needing professional judgement
4.0%
1.9×
Watch
Files in none of these pools
1.9%
0.9×
Normal
Bar: correction-rate lift vs. the all-application baseline · scale 0–4.0× · tick marks the 2.0× review threshold2 of 4 slices over threshold
Evidence-linked improvement
First the test, then the release
A cycle is done when the failure has become a test the next release has to survive.
Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn
01Detect
Reviewer corrections rise in one applicant pool.
02Diagnose
The files, the criteria version and the redaction log are read together until one of them explains it.
03Improve
Version-stamp the change and attach the files that exposed it.
04Verify
The affected cases run again, and a fail stops the release.
05Learn
It becomes a standing test, and the review rubric changes with it.
Learn → DetectThe return edge. Detection next cycle is measured against the longer suite.
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, reading workflow, evaluation, integration, then production validation and handover.
WorkstreamWeek 1Week 2Week 3Week 4Week 5Week 6
01Admissions workflow discovery and boundary definition.
02Application-platform and reader-tool assessment.
03Published-criteria and redaction-rule mapping and rule mapping.
04Application ingestion and transcript parsing.
05Reading logic and evidence-span binding.
06Confidence scoring and second-reader routing.
07Admissions officer review.
08Admissions- and CRM-system integration.
09Read-consistency cases.
10Guardrails and review controls.
11Review-trail instrumentation.
12Deployment, documentation and Agent Care handover.
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallelFinal 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 pool, one officeProductionProduction admissions systemsAdvancedMultiple campuses / systems
Introduced at Pilot
Reading to your published criteria✓✓✓
Officer review✓✓✓
Read-consistency baseline✓✓✓
Introduced at Production
Reporting by applicant pool—✓✓
Review workflow in your systems—✓✓
Approved write-back—✓✓
Application-system integration—✓✓
Introduced at Advanced
Multi-campus and multi-state rules——✓
Multi-stage committee review——✓
High application volume——✓
Enterprise review controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build priceConfirmed after discovery based on integrations, workflow complexity, application volume, review controls and deployment requirements.
Separate from buildBuild 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.
You bringWe build with it
01Your published criteria and how they are applied→Published-criteria mapping and read configurationWeek 1
02Representative files from a past cycle→Reading baseline, criteria extraction and evidence-span bindingWeek 2
03Your redaction list and where it has to hold→Published-criteria and redaction-rule mappingWeek 1
04Access to relevant APIs, feeds or exports→Application, CRM and transcript assessment, then integration setupWeek 2
05Reads that came out inconsistent→Read-consistency cases and the evaluation suiteWeek 4
06What must reach a human reader before any decision→Confidence scoring, second-reader routing, guardrails and review controlsWeek 3
07Named admissions officers to review reads→Officer review workflow, then pilot and production validationWeeks 5–6
Nothing else is requiredDeployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy, which is why week 5 doubles up rather than padding.
PhaseW1W2W3W4W5W6
DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6
Week focusW1Admissions workflow discovery, criteria mapping and the automation boundaryW2Source integration and the reading baselineW3Reading workflow, confidence logic and review controlsW4Evaluation suite, redaction checks and failure-mode testingW5System integration, a pilot pool and targeted correctionsW6One application cycle read under the admissions office, then handover
Reading the bandA band spans the weeks its work is named in and no others. The doubled fifth week is real, not padding.
At the end of W6The last checks clear on real applications and monitoring moves to Agent Care.
DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Education AI agent
Build an admissions assistant around your institution's review chain.
Show us your application platform, your published criteria and who reads a file. Your financial aid administrator names the point an aid question stops being data, and we map the review path from there, set the automation boundary and leave the decision with the officer.
