HR AI agent · Candidate screening
Candidate Screening AI Agent
Score applications against the written requisition, keep the ranking, the notice and the impact figures together, and hand the shortlist to the named recruiter who decides who is interviewed.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Named recruiter
- Notice first
- After launch
- Agent Care
What this agent does
Produces the rank, not the rejection
In
- 01
The written requisition sets the criteria, and the score records which of them moved it.
- 02
The output is a rank, and where the line falls under that rank is drawn by a person.
Reason
- 03
Selection rate, scoring rate and impact ratio go to the record, and so does the unknown count.
- 04
A fit score makes the tool an automated employment decision tool under NYC Admin. Code 20-870.
- 05
Candidate notice runs ten business days ahead of use, so a requisition without one is not scored.
Decide
- 06
A vendor-supplied score is treated as a consumer-report question, not as a product feature.
- 07
Postcode, school and gaps in work history are flagged as proxies before they are weighed.
Out
- 08
A ranking carries the day it ran and the adverse-impact benchmark in force on that day.
- 09
Write to the tracker only inside the approval boundaries agreed during implementation.
Product statement
Reading, scoring and record-keeping belong to the agent. The shortlist belongs to a named recruiter, who takes it into the hiring decision and owns it there.
Example workflow
One requisition, application to shortlist
- 1Applications receivedCVs, application forms, screening answers, referral notes and tracker fields
- 2Criteria fixed and versionedThe written criteria, the weights, the notice already served and the day that set was fixed
- 3Scored and ranked apartThe score, the rank, the features that moved it and the criteria it was read against
- 4Controls appliedProxy checks, notice checks, impact-ratio checks and scoring confidence
No human action required
Stages 1 to 4 run unaided, and nobody leaves the pool at any of them — the agent is ranking, and the recruiter lane opens at the shortlist gate.
Goes to the named recruiter to shortlist.
Adds a talent-lead read first.
Recruiter review
The ranking is held with its criteria, its impact figures and the applications it was drawn from.
Shortlist · Return to pool · Send to talent lead
Shortlisted — by the named recruiter
- 6Tracker and notice records updatedOnly where write access and hiring records policy allow it
- 7Outcome evaluatedImpact ratios, notice timing, recruiter corrections and what review found
Corrections
Each recruiter correction is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Rejecting an applicant on a score.
- Deciding who is interviewed.
- Setting where the shortlist line falls.
- Signing a bias audit or its published summary.
Automation boundary
Agent acts unaided
- ✓Read each application against the written criteria.
- ✓Record the selection rate and the scoring rate for each category.
- ✓Hold the ranking until the notice is evidenced.
- ✓Flag postcode, school and gaps in work history as possible proxies.
Nobody leaves the pool except by a named recruiter, inside the agreed boundaries.
- Judging whether a criterion is job related.
- Telling a candidate why they did not advance.
- Choosing which requisitions the agent scores.
- Changes to the criteria, weights or notice.
Example output
One scored application, annotated
This serves a talent team who may have to explain a rejection years after the requisition closed; below is one scored application exactly as the agent leaves it.
Scored application · single requisition
Illustrative example- Score
- Data engineer, London office
- Recorded as
- Scored above the sample median
- Requisition
- A rank, not a decision
- Evidence of record
- Application form, 3 August 2026
- Confidence
- Held unshortlisted
- Held for
- The named recruiter, by name
As received
Built from the application form and the written requisition alone, and it claims nothing beyond them.
What the record holds
- Application form
- Written criteria
- Notice served
Why no rejection here
Cutting a candidate from the pool is a judgement the recruiter makes.
What the score decides
Below the set threshold a ranking gets a talent-lead read before the recruiter sees it.
Action
ShortlistReturn to poolSend to talent leadValue
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
Each score
From the application that carries it
Where the evidence is used
The agent does not vouch for an applicant, only for what the application says, when it arrived, and how the score was built from it.
It learns who got hired
Train a model on a decade of past hires and it reproduces them fluently: change nothing on a CV but the name, and the rank moves.
What was checked, and not found
The four-fifths rule sits at 29 CFR 1607.4(D) and was read there on 20 August 2026, but OFCCP removes its own copy at 41 CFR 60-3 on 26 October 2026 and the EEOC has a rescission at final-rule stage, so a ranking records which benchmark it was measured against and the date that benchmark applied.
Build an evidence trail
The application, the score it drew and the recruiter who shortlisted it stay together.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Applicant tracking
- Greenhouse
- Lever
- Workday
- Applications, stages and outcomes
Requisition and criteria
- Job specs
- rubrics
- weights
- Written criteria of record
Candidate-supplied documents
- CVs
- forms
- screening answers
- What the applicant chose to send us
Agent
Candidate screening and ranking
- Reads the applications
- Scores and ranks
- Holds for the recruiter
Vendor scoring services
- Assessment vendors
- scoring APIs
- Third-party scores and summaries
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six sieves between the model and the recruiter
Six sieves down one column, the last the finest. What settles is drawn in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a ranking may stand
- L3
- Leaves the shortlist to a named recruiter
- L4 – L5
- Keep the shortlist and the score behind it
- L6
- Ranks nothing and returns the pool when signals degrade
How Nestack evaluates it
Evaluate the whole screen — not only the shortlist that comes out.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the shortlist a recruiter works
- E1Final-output evaluationDid the ranking record the criteria it was actually built on?
- E2Step-level evaluationDid the agent read the right requisition, the right criteria and the live weights?
- E3Tool evaluationDid it read and write the correct requisition and the correct application?
- E4Confidence calibrationDo low-confidence scores actually attract more recruiter corrections?
- E5Slice evaluationHow does performance change across specific requisition classes?
- E6Business outcomeHow many rankings needed a correction before the recruiter shortlisted?
Floor — the pool a ranking rests on
Failure modes
Where each failure originates in the agent
Seven failure modes, set at the stage where each one first shows itself.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- TH-03
Scored before notice served
The application is read with no notice on file.
Stage gathers
The forms, the criteria and the notice record
02 · Reasoning
2 modes
- TH-04
Proxy carries what the score drops
A postcode or a gap in work stands in for a person.
- TH-06
Rescinded benchmark read as live
A withdrawn measure is worked as the current one.
Stage proposes
The criteria, the score and the impact figures
03 · Tool / write
2 modes
- TH-02
Thin impact base passed forward
A ranking moves on with most categories unknown.
- TH-05
Vendor score acted on unflagged
A bought-in score reaches a rejection unmarked.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- TH-01
Shortlisted, ranking unrecorded
The record shows a shortlist but not the ranks under it.
Stage returns
The ranking a recruiter reads and the notice sent
05 · Change / Version
1 mode
- TH-07
Silent drift into a scored tool
A summariser gains a score and changes what it is.
Stage tracks
Model, prompt, criteria rules and score dates
- Sev-1 · a candidate cut on no record
- Sev-2 · a skewed ranking reaches a lead
- Sev-3 · signals degrade, pool held unranked
Affected slices
Entry-level volume absorbs the corrections
A requisition-level impact figure can read clean while high-volume entry-level roles carry most of the corrections. Nestack reports the correction rate by requisition class, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| High-volume entry-level roles | 14.3% | 3.7× | Review | |
| Multi-jurisdiction requisitions | 10.1% | 2.6× | Review | |
| Career-changer applications | 6.3% | 1.6× | Watch | |
| Established specialist roles | 2.9% | 0.8× | Normal |
Bar: correction-rate lift vs. specialist-role baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What a thinned pool costs
A cycle ends when the group thinned at one stage alone is a standing case. That suite is what the next requisition run is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Correction rate rises on high-volume entry-level roles.
- 02
Diagnose
The requisition whose shortlist quietly stopped holding anyone with a break in their work history is worked backwards until one cause is left standing.
- 03
Improve
The shortlist goes out numbered, and the scores beneath it ride with it.
- 04
Verify
One impact case still failing is enough to hold the shortlist back.
- 05
Learn
It is kept for good, and the scoring rules change in that same commit.
Learn → Detect
The return edge. The next ranking is measured 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, scoring and ranking, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Requisition criteria and automation-boundary scope.Week 1
- 02Tracker, requisition and vendor sources.Week 1, Week 2
- 03Application-to-score and criteria-coverage mapping.Week 1, Week 2
- 04Applicant source ingestion.Week 2
- 05Criteria, weight and score binding.Week 2, Week 3
- 06Impact scoring and review routing.Week 3
- 07Recruiter shortlist workflow.Week 3, Week 4
- 08Applicant-tracker integration.Week 5
- 09Impact and scoring cases.Week 4, Week 5
- 10Guardrails and shortlist controls.Week 4
- 11Score-trail instrumentation.Week 5, Week 6
- 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 tier | PilotOne requisition, one cycle | ProductionProduction hiring workflow | AdvancedMultiple requisitions / jurisdictions |
|---|---|---|---|
| Introduced at Pilot | |||
| Scoring to your written criteria | ✓ | ✓ | ✓ |
| Named recruiter shortlisting | ✓ | ✓ | ✓ |
| Applicant-pool baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by requisition class | — | ✓ | ✓ |
| Recruiter review workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Tracker-and-source integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-source applications | — | — | ✓ |
| Cross-requisition ranking packs | — | — | ✓ |
| Large applicant pools | — | — | ✓ |
| Multi-jurisdiction notice controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, application 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.
You bring
We build with it
01Your live requisitions and the criteria each is scored against
→Criteria capture and notice versioning
Week 102Representative applications, tracker fields and vendor scores
→Source binding, scoring logic and the applicant-pool baseline
Week 203Your hiring calendar and the recruiters it names
→Criteria mapping, notice binding and the automation boundary
Week 104Access to the tracker APIs, feeds or exports
→Tracker, requisition and vendor source assessment, then integration setup
Week 205Rankings you would not want audited
→Impact cases and the failure round
Week 406What no score may establish
→Impact scoring, review routing, guardrails and release controls
Week 307A named recruiter who shortlists
→Release to the named recruiter, 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 fifth week carries two bands because those two phases coincide, not to make the column read better.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Criteria discovery, notice versioning and the automation boundary
W2Source integration and the applicant-pool baseline
W3Scoring, ranking logic and release controls
W4Evaluation suite, impact cases and the failure round
W5Tracker integration, pilot rankings and targeted corrections
W6One hiring cycle run under the talent lead, then Agent Care handover
- Reading the band
- No bar runs past the weeks its own work is named for, and week five is shared by design.
- At the end of W6
- When the score record validates, Agent Care adopts the agent.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · HR AI agent
Build a candidate screening agent around the ranking your last requisition never wrote down.
Show us one requisition you run each quarter and the criteria behind the last shortlist. If nobody can name the benchmark it was measured against, the answer is a date rather than a rule: Title VII has not moved, and the instrument that measured it is leaving.
- Typical delivery
- 4–6 weeks
- Deployment
- Your live requisitions
- Named recruiter
- Notice first
- After launch
- Agent Care available
Nestack Agents · Candidate screening
AGT-HR-01 · Agent Care available after launch