HR AI agent · Compensation benchmarking
Compensation Benchmarking AI Agent
Draw the range, name the survey it came from and count the employers standing behind the cut, then hold that number, with its provenance attached, for a named reward officer to approve.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Named officer
- Source traced
- After launch
- Agent Care
What this agent does
Draws the range, never approves it
In
- 01
A range is drawn, and the survey beneath it is named on the face of the output.
- 02
A source cannot be traced to a survey, and it is marked unusable rather than blended in.
Reason
- 03
A benchmark thins to a few reporters, and that reporter count is printed beside the number.
- 04
OFCCP EO 11246 rules were rescinded by final rule on 21 August 2026, so no audit duty is assumed.
- 05
An EEO-1 rescission reached OIRA on 14 May 2026, so no form is assumed to still be there.
Decide
- 06
Art. 9(6) puts accuracy on the employer's management, so the agent puts its name to nothing.
- 07
A gap is explained by a control that sits downstream of it, and the model is held, not shown.
Out
- 08
A number resists sourcing, and the failure to source it is what goes into the record.
- 09
Write back to range records only inside the approval boundaries agreed at implementation.
Product statement
Drawing, matching and holding belong to the agent. The number a pay decision rests on belongs to a named reward officer, who approves it and owns where it came from.
Example workflow
One market range, sources to approval
- 1Sources receivedPublished survey cuts, job records, internal pay history or a vendor extract
- 2Matched and datedWhich survey, which cut, how many employers reported into it and the day it was collected
- 3Range drawn and cells countedThe benchmark, the internal spread, the reporter count and the vintage carried forward
- 4Controls appliedProvenance checks, cell-count checks, vintage checks and match confidence
No human action required
Stages 1 to 4 run unaided, and no range is approved at any of them — the agent is drawing, and the officer lane opens at the approval gate.
Goes to the named reward officer to approve.
Adds a senior reward read first.
Reward review
The range is held with the surveys under it, the reporter counts and the sources it could not trace.
Approve · Attach source · Send to reward review
Approved — by the named reward officer
- 6Range and job-architecture records updatedOnly where write access and retention policy permit it
- 7Outcome evaluatedProvenance capture, source currency, officer corrections and what review found
Corrections
Each reward correction is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Approving a range for use in a pay decision.
- Confirming the accuracy of a pay-gap report.
- Deciding a source is lawful to bring into the corpus.
- Setting the market position a job family is paid at.
Automation boundary
Agent acts unaided
- ✓Write the survey and its vintage beside the range.
- ✓Characterise each source by who collected it and how aggregated it is.
- ✓Print how many employers reported into a cut.
- ✓Carry the vintage of each survey forward into the range recommendation.
No range is approved except by a named reward officer, inside the agreed boundaries.
- Judging whether a job matches a survey benchmark.
- Telling a works council a category gap is justified.
- Choosing which market a job family is benchmarked to.
- Changes to a range, a grade or a pay decision.
Example output
One market range, annotated
This serves a reward team who may have to reconstruct, six months after a pay-gap report was submitted under Art. 10(1)(c), which survey a benchmark came from; below is one range exactly as the agent leaves it.
Market range · single job family
Illustrative example- Range
- Range review, senior engineering family
- Recorded as
- Traced to a named survey cut
- Survey
- Reporter count, printed
- Evidence of record
- Survey cut, 3 August 2026
- Confidence
- Held unapproved
- Held for
- The named reward officer
As received
Drawn from a third-party survey cut and the internal job architecture, and it claims nothing past them.
What the record holds
- Third-party survey
- Job-architecture record
- Internal pay ledger
Why no approval here
Approving a range is a judgement that the named reward officer makes.
What the score decides
Below the configured threshold a range gets a reward read before the officer sees it.
Action
ApproveAttach sourceSend to reward reviewValue
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 range
From the survey that carries it
Where the evidence is used
The agent does not vouch for a survey, only for who ran it, when it was collected and how many employers stood behind the cut; the same Guidelines reach a tool generating wage recommendations.
Whose numbers are these
Read in the January 2025 Guidelines: an exchange may be unlawful whether or not that effect was intended, and a competitor stands behind the cut.
What was checked, and not found
No safety zone for information exchange appears in the January 2025 Guidelines. The 1996 statements that carried one were withdrawn by DOJ on 3 February 2023 and rescinded by the FTC on 14 July 2023, and the 2016 guidance they sat under is still downloadable — availability is not currency.
Build an evidence trail
The range, the survey it came from and the officer who approved it stay together.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Third-party survey data
- Survey vendors
- published cuts
- Benchmark ranges by job family
Job architecture
- Grades
- families
- job records
- Matches and scope of work
Internal pay ledger
- Payroll
- pay history
- headcount
- Current pay by job and grade class
Agent
Compensation benchmarking
- Reads the sources
- Draws the range
- Holds for the officer
Reporting and filing records
- Pay-gap reports
- filing archive
- Category cuts and past submissions
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six rules between the survey and the range
Six rules against one edge, the last the finest. What measures true is set out in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a range may stand
- L3
- Leaves the approval to a named reward officer
- L4 – L5
- Keep the range and the survey behind it
- L6
- Reports the spread and no recommendation when signals degrade
How Nestack evaluates it
Evaluate the whole derivation — not only the range that comes out of it.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the range a pay decision uses
- E1Final-output evaluationDid the range record the surveys it was actually drawn from?
- E2Step-level evaluationDid the agent read the right survey, the right cut and the live match?
- E3Tool evaluationDid it read and write the correct job family and the correct market?
- E4Confidence calibrationDo low-confidence ranges actually attract more officer corrections?
- E5Slice evaluationHow does performance change across specific job families?
- E6Business outcomeHow many ranges needed a correction before the officer approved?
Floor — the sources a range rests on
Failure modes
Where each failure originates in the agent
Seven failure modes, each pinned to the stage where it first surfaces.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- UP-03
Tainted source ingested
A shared folder enters the corpus as market evidence.
Stage gathers
The surveys, the vintages and the job matches
02 · Reasoning
2 modes
- UP-04
Job matched on title alone
A whole survey distribution is imported wholesale.
- UP-06
Withdrawn safe harbour read as live
A rescinded survey condition is worked as shelter.
Stage proposes
The survey, the match and the reporter count
03 · Tool / write
2 modes
- UP-02
Thin cell passed forward
A range moves on without the senior reward read.
- UP-05
Range bound to wrong family
The number is filed against another job family.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- UP-01
Approved, provenance unrecorded
The record shows approval but no survey beneath it.
Stage returns
The range a pay paper carries and an offer uses
05 · Change / Version
1 mode
- UP-07
Silent vintage drift
Aged survey data is re-dated as though newly collected.
Stage tracks
Model, prompt, sourcing rules and survey dates
- Sev-1 · a range approved on no source
- Sev-2 · a competitor figure reaches a range
- Sev-3 · source degrades, range held back
Affected slices
Hot-market families absorb the corrections
A function-level correction figure can read clean while hot-market technical families carry most of the corrections. Nestack reports the correction rate by job family, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Hot-market technical families | 9.4% | 3.7× | Review | |
| Newly created hybrid roles | 6.6% | 2.6× | Review | |
| Multi-country job families | 4.1% | 1.6× | Watch | |
| Established core job families | 1.9% | 0.7× | Normal |
Bar: correction-rate lift vs. core-job-family baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What an untraced number costs
A cycle ends when the number traced to a competitor is a standing case. That suite is what the next review round is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Correction rate rises on hot-market technical families.
- 02
Diagnose
The benchmark that came from a peer spreadsheet rather than a survey, and read exactly like one, is worked backwards until a single cause is left standing.
- 03
Improve
The range goes out numbered, and the sources beneath it ride with it.
- 04
Verify
One provenance case still failing is enough to hold the round back.
- 05
Learn
The case stays on, and the sourcing rules are redrawn beside it.
Learn → Detect
The return edge. The next range 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, range drawing, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Job-matching review and automation-boundary scoping.Week 1
- 02Survey, job and pay-ledger sources.Week 1, Week 2
- 03Source-to-range and provenance-coverage mapping work.Week 1, Week 2
- 04Survey source ingestion.Week 2
- 05Survey, match and job-family binding.Week 2, Week 3
- 06Cell-count scoring and review routing.Week 3
- 07Reward officer approval workflow.Week 3, Week 4
- 08Reward-system integration.Week 5
- 09Provenance and matching cases.Week 4, Week 5
- 10Guardrails and recommendation controls.Week 4
- 11Source-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 job family, one round | ProductionProduction reward workflow | AdvancedMultiple families / markets |
|---|---|---|---|
| Introduced at Pilot | |||
| Range drawing to your job architecture | ✓ | ✓ | ✓ |
| Named officer approval | ✓ | ✓ | ✓ |
| Job-architecture baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by job-family class | — | ✓ | ✓ |
| Reward review workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Survey-and-ledger integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-source provenance rules | — | — | ✓ |
| Cross-market range packs | — | — | ✓ |
| Large survey libraries | — | — | ✓ |
| Multi-market range controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, survey 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 job families and the survey each is benchmarked to
→Source capture and vintage versioning
Week 102Representative survey cuts, job records and pay history
→Source binding, match logic and the job-architecture baseline
Week 203Your pay calendar and the reward officers it names
→Source mapping, provenance capture and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Survey, job and pay-ledger assessment, then integration setup
Week 205Ranges you would not want subpoenaed
→Provenance cases and the market round
Week 406What no range may be called
→Cell-count scoring, review routing, guardrails and approval controls
Week 307A named reward officer who approves the range
→Handover to the named reward officer, 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 follow the work rather than the grid, which is why one band is drawn beneath another in week five.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Source discovery, vintage versioning and the automation boundary
W2Survey integration and the job-architecture baseline
W3Range drawing, match logic and approval controls
W4Evaluation suite, provenance cases and the failure round
W5Reward-system integration, pilot ranges and targeted corrections
W6One pay round run under the reward lead, then Agent Care handover
- Reading the band
- A bar spans only the weeks its own work is named for, and week five carries two.
- At the end of W6
- When the source record validates, Agent Care takes the agent on.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Reward AI agent
Build a compensation benchmarking agent around the provenance your last range never carried.
Show us one job family you benchmark each year and the survey the last range came from. In Dorrell, No. 1:25-cv-02251 (D. Md.), one defendant settled in May 2026 and the case was then dismissed largely without prejudice on 5 August 2026. Cost arrived before vindication did.
- Typical delivery
- 4–6 weeks
- Deployment
- Your live job families
- Named officer
- Source traced
- After launch
- Agent Care available
Nestack Agents · Reward benchmarking
AGT-HR-10 · Agent Care available after launch