Main site

HR AI agent · Policy and entitlements

Policy & Entitlements Copilot AI Agent

Answer policy and entitlement questions with the clause named and the carve-out attached, then hold the draft for the named adviser who states what the employer owes.

Typical delivery
4–6 weeks
Deployment
Your stack
Adviser decides
Clause named
After launch
Agent Care

What this agent does

Drafts the answer, never states the entitlement

In

  1. 01

    A question about holiday or notice arrives, and the clause it is answered from is named beside it.

  2. 02

    Varity, 516 U.S. 489, 1996: answering a benefits question is itself a fiduciary act.

Reason

  1. 03

    Amara, 563 U.S. 421, 16 May 2011: a summary describes a plan and is not its terms.

  2. 04

    Eddy, D.C. Cir., 23 November 1990: complete and correct material information, not merely true.

  3. 05

    29 CFR 2520.102-2(b): a limitation may not be minimised, so the carve-out rides with the benefit.

Decide

  1. 06

    29 CFR 790.13: the good-faith defence needs a written instrument this agent cannot issue.

  2. 07

    A commencement instrument moves a qualifying period, and the stored answers are re-derived.

Out

  1. 08

    An entitlement turns on a fact no policy holds, and that dependency is recorded as the finding.

  2. 09

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

Product statement

Drafting, sourcing and holding belong to the agent. Release belongs to a named adviser, who states the entitlement and owns what the employer has said.

Example workflow

One entitlement question, corpus to release

  1. 1Question receivedHandbook clauses, plan documents, summaries, the contract file and the applicable particulars
  2. 2Documents identified and rankedWhich instrument governs, which is a summary of it, and the day each version took effect
  3. 3Draft answer builtThe entitlement stated, the clause cited, the carve-out attached and what the corpus could not reach
  4. 4Controls appliedJurisdiction checks, commencement checks, summary-against-plan checks and release confidence

No human action required

Stages 1 to 4 run unaided, and nothing is said to an employee at any of them — the agent is drafting, and the adviser lane opens at the confidence gate.

5DecisionBranches at the release threshold
High confidence

Goes to the named adviser to release.

Low confidence

Adds a benefits-counsel read first.

Adviser review

The draft is held with the clauses under it, the carve-outs and the confidence.

Release · Attach clause · Send to counsel

Released — by the named adviser

  1. 6Case and answer records updatedOnly where write access and records policy allow it
  2. 7Outcome evaluatedClause accuracy, carve-out coverage, adviser corrections and what review found

Corrections

Each adviser correction is counted in the evaluation.

What should not run autonomously

Human approval stays in control

Outside the boundary — human approval required

8 items

  • Telling an employee what they are owed.
    • Deciding which document governs the entitlement.
    • Refusing a request that is in substance a claim.
    • Signing a written statement of particulars.

  • Automation boundary

    Agent acts unaided

    • ✓Cite the clause that supports each drafted answer.
    • ✓Set the disqualification clause beside the benefit that it limits.
    • ✓Check each quoted period against the commencement instrument in force.
    • ✓Hold the draft where the jurisdiction is unstated.

    Nothing reaches an employee except by a named adviser, inside the agreed boundaries.

    • Judging whether a policy is a term of the contract.
    • Assuring anyone that a benefit will be paid.
    • Ranking a summary above the plan it describes.
  • Changes to the handbook, the plan or the particulars.

Example output

One entitlement answer, annotated

This serves an HR team who may have to explain, years later, why a worker was told what they were told; below is one drafted answer exactly as the agent leaves it.

Entitlement answer · single question

Illustrative example
Question
Paternity leave in the first year
Answer
No qualifying period applies to a GB worker on these facts
Status
Draft, unreleased
Instrument cited
ERA 2025 s.16, 6 April 2026
Confidence
Summary, not plan terms
Held for
The named adviser, by name

As received

Drawn from the commencement instrument and the handbook clause, and it claims nothing beyond either.

What the record holds

  • Handbook clause
  • Commencement date
  • Contract particulars

Why no release here

Saying what a worker is owed is a judgement a named adviser makes.

What the score decides

Below the configured threshold an answer gets a counsel read before the adviser sees it.

Action

ReleaseAttach clauseSend to counsel

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

Each answer

From the clause that carries it

03Drafting

Where the answer is used

The agent vouches for what a policy says and the day it took effect, not for the outcome. Pearce, 6th Cir., 20 June 2018: unambiguous plan terms defeated estoppel — capable of binding is not bound.

01Approved path

Saying it can make it so

On remand the Second Circuit affirmed reformation of the plan to adhere to representations made by the plan administrator, 23 December 2014.

02Human review

What was checked, and not found

No regime checked requires an employer to keep a record of the advice it gave a worker about an entitlement. But 29 CFR 825.500(c)(4) retains the written notices given to employees, and (c)(5) the documents describing policies and practices, which at scale is the same file.

04

Build an evidence trail

The answer, the document it came from and the adviser who released it stay together.

Integrations

Typical integrations

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

  • Policy and handbook stores

    • SharePoint
    • policy libraries
    • Handbook clauses and their dates

  • Benefits administration

    • Plan documents
    • summaries
    • Eligibility and its carve-outs

  • Core HR records

    • Workday
    • SAP SuccessFactors
    • Hire dates, the entity and jurisdiction

Agent

Policy and entitlement answering

  • Reads the clauses
  • Drafts the answer
  • Holds for the adviser

  • Contract and statutory files

    • Offer letters
    • s.1 particulars
    • Notice, holiday and sick-pay terms

  • Observability & evaluation

    • OpenTelemetry
    • Langfuse
    • Supported monitoring/evaluation sources

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

Agent controls

Six gates between the corpus and the employee

Six gates down one corridor, the last the narrowest. What clears is set out in the map below.

L6 · Outermost — last line of defence

Inward → L1 · closest to the model

L6Rollback / safe modeNarrow the agent to clause retrieval when evaluation or production signals degrade.Roll back
L5Version monitoringTrack model, prompt and clause rules, and note the version each answer was drafted under.Track
L4TraceabilityRecord each answer, the clauses under it, the version it cited and every read of the file.Record
L3Adviser releaseHold the draft for a named adviser; the hold governs release, not whether the answer is right.Gate
L2Instrument guardrailsTest each answer against the instrument in force, and refuse one whose clause has been superseded.Restrict
L1Confidence thresholdsRoute a low-confidence draft to a counsel read before the answer reaches a worker.Require review
Model coreAnswer drafted — the entitlement, the clause, the carve-out and its sources
L1 – L2
Test whether an answer may be released
L3
Leaves the statement to a named adviser
L4 – L5
Keep the answer and the document behind it
L6
Answers with the document alone when signals degrade

How Nestack evaluates it

Evaluate the whole draft — not only the answer that comes out.

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

Surface — the answer a worker reads

  1. E1Final-output evaluationDid the answer record the clause it was actually drafted from?
  2. E2Step-level evaluationDid the agent read the right document, the right version and the right entity?
  3. E3Tool evaluationDid it read and write the correct worker and the correct entitlement?
  4. E4Confidence calibrationDo low-confidence answers actually attract more adviser corrections?
  5. E5Slice evaluationHow does performance change across specific entitlement classes?
  6. E6Business outcomeHow many answers needed a correction before the adviser released?

Floor — the corpus an answer rests on

Failure modes

Where each failure originates in the agent

Seven failure modes, each at the stage where it first becomes visible.

Agent lifecycle

Direction of processing →

  1. 01 · Retrieval

    1 mode

    • TI-03

      Wrong-entity document read

      The policy pulled belongs to another entity or another jurisdiction.

    Stage gathers

    The clauses, the plan text, the dates and the entity

  2. 02 · Reasoning

    2 modes

    • TI-04

      Summary answered as plan

      A benefit is stated from a document that is not the plan.

    • TI-06

      Withdrawn letter cited as live

      A rescinded interpretation is worked as current authority.

    Stage proposes

    The entitlement, the clause, the carve-out and confidence

  3. 03 · Tool / write

    2 modes

    • TI-02

      Carve-out left off the answer

      The benefit is stated and the condition on it is not.

    • TI-05

      Answer worded as a refusal

      A no becomes a determination nobody wrote up.

    Stage writes

    Only where write access and approval policy allow it

  4. 04 · Output

    1 mode

    • TI-01

      The answer becomes the term

      A released statement is read later as what the employer promised.

    Stage returns

    The answer an employee acts on and a court reads

  5. 05 · Change / Version

    1 mode

    • TI-07

      Silent commencement drift

      A qualifying period is removed and the stored answers keep it.

    Stage tracks

    Model, prompt, clause rules and in-force dates

  • Sev-1 · an entitlement promised in error
  • Sev-2 · a superseded rule reaches a worker
  • Sev-3 · source degrades, answer held back
See the complete Human Resources failure-mode catalogue

Affected slices

Cross-border questions absorb the corrections

A department-level accuracy figure can read clean while cross-border entitlements carry most of the adviser corrections. Nestack reports the correction rate by answer class, not only in total.

Slice performance — reported separately, not only in aggregate

Illustrative example
SliceFailure rateLiftLift vs. thresholdStatus
Cross-border entitlements13.7%3.7×Review
Plan-governed benefits9.7%2.6×Review
Recently commenced rights6.1%1.6×Watch
Settled handbook questions2.8%0.8×Normal

Bar: correction-rate lift vs. the settled-handbook baseline · scale 0–4.0× · tick marks the 2.0× review threshold

2 of 4 slices over threshold

Evidence-linked improvement

What a promised benefit costs

The loop shuts when the assurance no plan document carries is a standing case. That suite is what the next answer released is measured against.

Improvement cycle · five stages

Switchback — the path turns at Improve and returns at Learn

  1. 01

    Detect

    Correction rate rises on cross-border entitlements.

  2. 02

    Diagnose

    The benefit that existed only because somebody said it did is worked backwards until one cause is left standing.

  3. 03

    Improve

    Every answer goes out numbered, with the clauses that support it attached.

  4. 04

    Verify

    Nothing releases while one touched entitlement case is red.

  5. 05

    Learn

    The case stays on, and the answering rules are rewritten alongside it.

  6. Learn → Detect

    The return edge. The next answer 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, answer drafting, evaluation, integration, then production validation and handover.

Workstream

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

  1. 01Entitlement questions and automation-boundary work.Week 1
  2. 02Handbook, plan and contract sources.Week 1, Week 2
  3. 03Source-to-clause and entitlement-coverage tracking.Week 1, Week 2
  4. 04Policy-corpus ingestion.Week 2
  5. 05Clause, plan and worker binding.Week 2, Week 3
  6. 06Carve-out checks and review routing.Week 3
  7. 07Adviser release workflow.Week 3, Week 4
  8. 08Case-record integration.Week 5
  9. 09Entitlement and source cases.Week 4, Week 5
  10. 10Guardrails and answering controls.Week 4
  11. 11Answer-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 entitlement class, one cycleProductionProduction answering workflowAdvancedMultiple classes / jurisdictions
Introduced at Pilot
Answer drafting to your documents✓✓✓
Named adviser release✓✓✓
Policy-corpus baseline✓✓✓
Introduced at Production
Reporting by entitlement class—✓✓
Adviser review workflow in your systems—✓✓
Approved write-back—✓✓
Handbook-and-plan integration—✓✓
Introduced at Advanced
Conflicting instruments——✓
Cross-entity answer packs——✓
Large policy libraries——✓
Multi-jurisdiction answer controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build price
Confirmed after discovery based on integrations, workflow complexity, corpus 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 live entitlement classes and the document each is answered from

    Document capture and clause versioning

    Week 1

  2. 02Representative handbook, plan and contract sources

    Source binding, clause logic and the policy-corpus baseline

    Week 2

  3. 03Your benefits calendar and the advisers it names

    Entitlement mapping, clause binding and the automation boundary

    Week 1

  4. 04Access to relevant APIs, feeds or exports

    Handbook, plan and contract source assessment, then integration setup

    Week 2

  5. 05Answers you would not want reformed

    Entitlement cases and the evaluation round

    Week 4

  6. 06What no answer may promise

    Carve-out checks, review routing, guardrails and release controls

    Week 3

  7. 07A named adviser who releases the answer

    Release to the named adviser, 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

Widths track what a phase costs and not the grid, so the fifth week draws one band beneath another.

Phase

W1

W2

W3

W4

W5

W6

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

Week focus

W1Entitlement discovery, clause versioning and the automation boundary

W2Source integration and the policy-corpus baseline

W3Answer drafting, carve-out logic and release controls

W4Evaluation suite, entitlement cases and failure-mode testing

W5Case-record integration, pilot answers and targeted corrections

W6One benefits cycle run under the plan administrator, then Agent Care handover

Reading the band
Each bar runs only across the weeks its own work names, and week five is shared by design.
At the end of W6
Once the answer record validates, Agent Care assumes the agent.
Duration
Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.

Next step · HR AI agent

Build an entitlements copilot around the document your last answer never named.

Show us one entitlement your team answers week after week and the clause behind the last answer. If it went out without naming the document it came from, what the employer said is now the only record of what the employer meant. A benefit no plan carries comes back as a finding.

Typical delivery
4–6 weeks
Deployment
Your live policies
Named adviser
Document first
After launch
Agent Care available

Nestack Agents · Policy and entitlements

AGT-HR-02 · Agent Care available after launch

Build an entitlements copilot agent

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

About you
Your policies

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