HR AI agent · Engagement insights
Engagement Insights AI Agent
Count the responses, hold back any cell small enough to name the person who wrote it, and put the finding in front of a named officer who decides whether it may be published.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- A named officer
- Cell size first
- After launch
- Agent Care
What this agent does
Counts the responses, never publishes the finding
In
- 01
A wave closes, and the cell each answer was counted in is written down beside it.
- 02
A cell falls under the reporting minimum, and it is suppressed before a manager reads it.
Reason
- 03
The employer knows the roster, which is what GDPR Recital 26 means by means reasonably likely.
- 04
A survey runs on a platform, and BetrVG § 87(1) Nr. 6 turns on objective suitability alone.
- 05
In BAG 1 ABR 47/16 the survey escaped § 94 because it stayed anonymous and voluntary.
Decide
- 06
A retention call reads the score, and 29 CFR 1607.2(B) lists retention as a selection use.
- 07
Under 29 CFR 1607.4(D) the missing records can themselves evidence adverse impact.
Out
- 08
Consent is offered as the basis, and EDPB Guidelines 05/2020 paragraph 21 says it is not one.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
Counting, suppression and the cell record belong to the agent. Publication, and any decision taken on a finding, belongs to a named people-analytics officer.
Example workflow
One survey wave, response to publication
- 1Responses receivedPulse tool, annual census, free-text comments, exit surveys or a collaboration graph
- 2Cells counted and fixedThe wave, the cell each response fell in, the minimum in force and the cuts already published
- 3Finding drafted, cells namedThe draft finding, the cells beneath it, the responses it may not show and confidence
- 4Controls appliedCell-size checks, overlapping-cut checks, language-coverage checks and drafting confidence
No human action required
Stages 1 to 4 run unaided, and nothing is published at any of them — the agent is counting, and the officer lane opens at the suppression gate.
Goes to the named officer to publish.
Adds a works council read first.
Officer review
The finding is held with its cells, its response rate and the cuts it could be combined with.
Publish · Suppress · Send to works council review
Published — by the named officer
- 6Reporting and analytics systems updatedOnly where write access and the works agreement allow it
- 7Outcome evaluatedSuppression stability, response-rate movement, officer suppressions and what review found
Suppressions
Each officer suppression is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Publishing a finding to line managers.
- Setting the minimum cell size.
- Naming a team inside a readout.
- Releasing an individual risk score.
Automation boundary
Agent acts unaided
- ✓Suppress any cell falling under the minimum in force.
- ✓Flag any two cuts whose overlap rebuilds a suppressed cross-section.
- ✓Draft the works-council description of what the platform collects.
- ✓Hold the finding for the officer who publishes it.
Nothing publishes and no cell opens except by the named officer, inside the agreed limits.
- Judging whether a promise of anonymity still holds.
- Signing the works agreement the tool runs under.
- Deciding what a fall in candour means.
- Changes to the cell minimum, the cuts or the wording.
Example output
One finding, annotated
This serves a people-analytics team that may have to show a works council what the platform actually touches — and Annex III point 4(b) duties, deferred to 2 December 2027, will land on the same tool; below is one finding exactly as the agent leaves it.
Finding record · single team cell
Illustrative example- Finding
- Engagement index, one team cell
- Counted in
- Below the reporting minimum, unpublished
- Wave
- Counted in a cell of four
- Evidence held
- Response set, one free-text batch
- Confidence
- Held suppressed
- To be published by
- The publishing officer, by name
As received
Drawn from the survey platform and the wave record, and it says nothing about anyone the cell could name.
What the record holds
- Wave and cell
- Response rate
- Free-text batch
Why nothing publishes here
Saying a cell is safe to publish is a call the named officer makes.
What the score decides
Below the configured floor a finding picks up a works council read before the officer.
Action
PublishSuppressSend to works council 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
Every response
From the wave it was answered in
Where a score becomes a decision
Under 29 CFR 1607.16(Q) your flight-risk score is a selection procedure, and 1607.2(B) lists retention; 1607.15 wanted the impact records from the first score.
We said it was anonymous
A minimum cell size of five protects one cross-section; you are selling trends and drill-downs, which are many overlapping cross-sections.
What was looked for, and not found
A full-text search of the Code of Federal Regulations returns nothing for sentiment analysis, employee attrition or pulse survey, while adverse impact and collective bargaining both return hits on the same index, and no independent validation of individual attrition prediction was found.
Build an evidence trail
The response, the group it was counted in and the officer who published it stay together.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Survey platforms
- Qualtrics
- Culture Amp
- Glint
- Waves, items and response sets
Core HR records
- Workday
- Oracle HCM
- Personio
- Rosters, grades and tenure
Collaboration systems
- Microsoft 365
- Google Workspace
- Message and calendar metadata only
Agent
Engagement measurement
- Counts the responses
- Suppresses small cells
- Holds for publication
Works council and records
- Works agreements
- DPIA registers
- Cell minimums and retention terms
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 readout
Six filters along one train, the last the tightest. What survives is set out in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a finding may publish
- L3
- Leaves the publication to a named officer
- L4 – L5
- Keep the finding and the cell size behind it
- L6
- Suppresses the cell and reports nothing when signals degrade
How Nestack evaluates it
Evaluate the whole measurement — not only the score that comes out.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the score a leadership readout carries
- E1Final-output evaluationDid the finding record the cell it rests on and the rate that answered it?
- E2Step-level evaluationDid the agent read the right wave, the right roster and the live cell minimum?
- E3Tool evaluationDid it read and write the correct wave and the correct cell?
- E4Confidence calibrationDo low-confidence findings actually attract more officer suppressions?
- E5Slice evaluationHow does performance change across specific populations?
- E6Business outcomeHow many findings were suppressed before the officer published?
Floor — what people will tell their employer
Failure modes
Where each failure originates in the agent
Seven failure modes, each set at the stage where it first shows itself.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- VJ-03
Respondent mix never gathered
Who answered, and how many, is not carried.
Stage gathers
The wave, the roster, the cells and the rate
02 · Reasoning
2 modes
- VJ-04
Sarcasm scored as contentment
A bitter comment is counted as a warm one.
- VJ-06
Language read as a mood
A region lights up because it does not write in English.
Stage proposes
The cells counted, the cuts they allow and confidence
03 · Tool / write
2 modes
- VJ-02
Live with no works agreement
The platform runs before anybody signed for it.
- VJ-05
Risk score decides a retention call
A score becomes the basis of an employment decision.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- VJ-01
Published, and the writer is named
A manager recovers one person from the readout.
Stage returns
The score a readout carries and a board reads
05 · Change / Version
1 mode
- VJ-07
Connector widens, notice does not
A release note adds a surface nobody notified.
Stage tracks
Model, prompt, cell minimums and cut rules
- Sev-1 · a person named from a published cell
- Sev-2 · a cell publishes below the minimum
- Sev-3 · signal degrades, finding held back
Affected slices
Small teams absorb the suppressing
A population-level suppression figure can read clean while teams near the reporting floor carry most of the suppressing. Nestack reports the suppression rate by population, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Small teams near the floor | 8.3% | 3.7× | Review | |
| Non-English comment sets | 5.9% | 2.6× | Review | |
| Attrition-risk cohorts | 3.7% | 1.7× | Watch | |
| Large business units | 1.7% | 0.8× | Normal |
Bar: suppression-rate lift vs. large-business-unit baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What a nameable cell costs
A cycle shuts when the team small enough to name is a standing case. That suite is what the next survey wave is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Suppression rate rises on teams near the reporting floor.
- 02
Diagnose
The readout that let a manager work out who wrote the comment is worked backwards until one cause is left standing.
- 03
Improve
Findings go out numbered, and the cell sizes beneath them ride with it.
- 04
Verify
Each touched suppression case is run again, and one red holds it back.
- 05
Learn
The case is kept, and the reporting rules change in that same commit.
Learn → Detect
The return edge. The next finding 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, suppression logic, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Suppression rules and the automation-boundary work.Week 1
- 02Survey, roster and collaboration sources.Week 1, Week 2
- 03Response-set and collaboration-graph source mapping.Week 1, Week 2
- 04Survey response capture.Week 2
- 05Cell binding and suppression logic.Week 2, Week 3
- 06Confidence scoring and council routing.Week 3
- 07Officer publication workflow.Week 3, Week 4
- 08Reporting-system integration.Week 5
- 09Suppression and basis cases.Week 4, Week 5
- 10Guardrails and publication controls.Week 4
- 11Response-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 wave, one population | ProductionProduction reporting workflow | AdvancedMultiple waves / jurisdictions |
|---|---|---|---|
| Introduced at Pilot | |||
| Suppression to your cell minimums | ✓ | ✓ | ✓ |
| Named officer publication | ✓ | ✓ | ✓ |
| Population-and-cell baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by population class | — | ✓ | ✓ |
| Officer publication review in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Survey-and-directory integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-country works agreements | — | — | ✓ |
| Layered publication sign-off | — | — | ✓ |
| Large response volumes | — | — | ✓ |
| Multi-country consultation controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, population coverage, 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 survey waves and the officer each one names
→Response capture and cell counting
Week 102Representative past waves and how each was reported
→Roster binding, suppression logic and the population baseline
Week 203Your reporting calendar and the councils it consults
→Cell-minimum mapping, roster binding and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Survey, roster and collaboration-source assessment, then integration setup
Week 205Findings you would not want re-identified
→Suppression cases and the evaluation round
Week 406What no finding may reveal
→Confidence scoring, council routing, guardrails and release controls
Week 307A named officer who publishes the finding
→Release to the named officer, then pilot waves 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
These bands are counted, not spaced for appearance; the fifth week carries two because they coincide.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Wave discovery, cell-minimum mapping and the automation boundary
W2Survey and roster integration and the population baseline
W3Suppression logic, overlapping-cut checks and release controls
W4Evaluation suite, suppression cases and failure-mode testing
W5Reporting-system integration, pilot findings and targeted corrections
W6One survey wave run under the people-analytics lead, then Agent Care handover
- Reading the band
- A band covers only the weeks its own work is named for, and the fifth week doubles on purpose.
- At the end of W6
- When the response 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 an engagement insights agent around the promise you made when you asked.
Show us one survey wave and the readout that went to managers. If a team of eight was reported quarter after quarter and sliced two ways, the anonymity you promised was gone before anyone lied about it. Ask what cell each number was counted in.
- Typical delivery
- 4–6 weeks
- Deployment
- Your survey waves
- A named officer
- Cell size first
- After launch
- Agent Care available
Nestack Agents · Engagement insights
AGT-HR-12 · Agent Care available after launch