HR AI agent · Performance reviews
Performance Review Drafting AI Agent
Draft the review with each assertion tied to the record it came from, keep what the manager changed, and hold the document for the named manager who signs it.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Manager signs
- Sourced lines
- After launch
- Agent Care
What this agent does
Drafts the review, never signs it
In
- 01
A review is drafted, and the record behind each assertion is named beside it.
- 02
Ames, No. 23-1039, 5 June 2025: pretext asks whether a stated reason is the true one.
Reason
- 03
Dominguez-Cruz, 1st Cir., 2 February 2000: shifting explanations let a jury infer pretext.
- 04
EO 14281, 90 FR 17537: agencies deprioritise disparate impact. Treatment claims are untouched.
- 05
29 CFR 1602.14: once a charge lands, relevant records are held until final disposition.
Decide
- 06
Reg (EU) 2026/1744, in force 27 July 2026: Annex III duties wait until 2 December 2027.
- 07
Article 5(1)(f) was not deferred: emotion inference at work is prohibited, live 2 February 2025.
Out
- 08
A record is too thin to evaluate, and that thinness is recorded as the finding, not written over.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
Drafting, sourcing and holding belong to the agent. The signature belongs to a named manager, who adopts each sentence and answers for it later.
Example workflow
One review, notes to signature
- 1Source records receivedGoal records, one-to-one notes, peer input, delivery telemetry and the prior review
- 2Records ranked and boundWhich note supports which claim, who wrote that note, and the day it was written
- 3Draft review builtThe narrative, the record behind each assertion, the lines no record reaches and the diff against last cycle
- 4Controls appliedSource checks, prior-review diffing, protected-material checks and release confidence
No human action required
Stages 1 to 4 run unaided, and nothing reaches an employee at any of them — the agent is drafting, and the manager lane opens at the confidence gate.
Goes to the named manager to sign.
Adds a business-partner read first.
Manager review
The draft is held with its source records, its unsourced lines and the confidence.
Sign · Attach record · Send to HR
Signed — by the named manager
- 6Review and case records updatedOnly where write access and records policy allow it
- 7Outcome evaluatedSource accuracy, edit depth, manager corrections and what review found
Corrections
Each manager amendment is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Signing a review of another adult.
- Setting the rating a narrative supports.
- Deciding a performance concern exists.
- Countersigning a plan or a discharge.
Automation boundary
Agent acts unaided
- ✓Cite the source record behind each drafted line.
- ✓Surface any sentence whose evidence the record cannot substantiate.
- ✓Keep protected-activity material out of the drafting context window.
- ✓Diff each draft against the review it follows.
Nothing goes into a personnel file except by a named manager, inside the agreed boundaries.
- Judging whether the record supports a rating.
- Telling anyone a review is defensible.
- Choosing what a calibration outcome means.
- Changes to the rating, the plan or the file.
Example output
One drafted review, annotated
This serves HR and line managers who may have to explain, years later, who wrote a given sentence in a personnel file; below is one drafted review exactly as the agent leaves it.
Drafted review · single cycle
Illustrative example- Report
- Mid-year review, one direct report
- Draft
- Two goals met, one carried over from the prior cycle
- Status
- Draft, unsigned
- Record cited
- Goal record, 4 August 2026
- Confidence
- Two lines unsourced
- Held for
- The named manager, by name
As received
Drawn from the goal records and the notes on file, and it asserts nothing that those records do not.
What the record holds
- Goal record
- One-to-one note
- Peer input
Why no signature here
Owning a sentence as your own judgement is an act the manager makes.
What the score decides
Below the configured threshold a draft gets an HR read before the manager sees it.
Action
SignAttach recordSend to HRValue
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 drafted line
From the record that carries it
Where the review is used
The agent vouches for what a record says and the day it was written, not for what a manager believed. C-203/22, 27 February 2025: which data were used, and how, has to be answerable.
Read back in a deposition
SCHUFA, C-634/21: a preparatory score becomes the decision where the human draws strongly on it — and how strongly is a question about manager edits, which a product can measure.
What was checked, and not found
No rule checked obliges an employer to keep the notes, the prompt or the drafts a manager rejected. But 29 CFR 1602.14 preserves relevant personnel records until final disposition once a charge is filed, and a pipeline cannot hold what it deleted at the close of the cycle.
Build an evidence trail
The review, the evidence under it and the manager whose name it carries stay together.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Performance management
- Workday
- Lattice
- Culture Amp
- Goals, ratings and past reviews
Manager notes
- One-to-one notes
- check-ins
- What was observed, and when
Delivery telemetry
- Jira
- Salesforce
- GitHub
- Zendesk
- Tickets closed, never judgement
Agent
Performance review drafting
- Reads the records
- Drafts the review
- Holds for the manager
Peer and calibration input
- Peer feedback
- calibration records
- What colleagues wrote, and when
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six passes between the record and the file
Six passes over one page, the last the closest. What survives is drawn in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a draft may be signed
- L3
- Leaves the signature to a named manager
- L4 – L5
- Keep the review and the evidence behind it
- L6
- Hands the draft back unsigned when signals degrade
How Nestack evaluates it
Evaluate the whole drafting run — not only the review that comes out.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the review an employee reads
- E1Final-output evaluationDid the review record the source each assertion was drafted from?
- E2Step-level evaluationDid the agent read the right period, the right notes and the right report?
- E3Tool evaluationDid it read and write the correct employee and the correct cycle?
- E4Confidence calibrationDo low-confidence drafts actually attract deeper manager edits?
- E5Slice evaluationHow does performance change across specific report classes?
- E6Business outcomeHow many drafts needed a correction before the manager signed?
Floor — the records a review rests on
Failure modes
Where each failure originates in the agent
Seven failure modes, each set at the stage it first becomes visible.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- UL-03
Protected material retrieved
A leave, a complaint or an accommodation enters the context.
Stage gathers
The notes, the goals, the peer input and the metrics
02 · Reasoning
2 modes
- UL-04
Invented specific detail
A project, a date or an incident that no record holds.
- UL-06
Prior review regenerated
Last year's narrative returns as this year's judgement.
Stage proposes
The narrative, its sources, the diff and confidence
03 · Tool / write
2 modes
- UL-02
Narrative against the rating
The prose and the rating tell different stories.
- UL-05
Affect signal reaches the draft
Tone or sentiment inference enters an evaluative line.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- UL-01
Signed with no drafting record
The file keeps the text and the name, not who wrote it.
Stage returns
The review a worker signs and a lawyer reads back
05 · Change / Version
1 mode
- UL-07
Silent wording drift by class
The register moves by group while each review reads fine.
Stage tracks
Model, prompt, wording rules and review dates
- Sev-1 · a sentence the manager never meant
- Sev-2 · protected activity reaches the file
- Sev-3 · source degrades, draft held back
Affected slices
Thinly recorded reports absorb the corrections
A cycle-level accuracy figure can read clean while thinly recorded reports carry the bulk of the manager corrections. Nestack reports the correction rate by report class, and not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Thinly recorded reports | 9.9% | 3.7× | Review | |
| Newly transferred reports | 7.0% | 2.6× | Review | |
| Rating-boundary cases | 4.4% | 1.6× | Watch | |
| Long-tenured steady reports | 2.0% | 0.7× | Normal |
Bar: correction-rate lift vs. the long-tenured baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
What an unmeant sentence costs
A loop closes when the sentence the manager never meant is a standing case. That suite is what the next review cycle is measured against.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Correction rate rises on thinly recorded reports.
- 02
Diagnose
The line that entered the file because the deadline was nearer than the evidence is worked backwards until one cause is left standing.
- 03
Improve
Reviews go out numbered, and the evidence beneath them rides with it.
- 04
Verify
One wording case still failing is enough to hold the cycle back.
- 05
Learn
It is kept for good, and the drafting rules change in that same commit.
Learn → Detect
The return edge. The next review 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, review drafting, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Review narratives and automation-boundary definition.Week 1
- 02Goal, note and telemetry sources.Week 1, Week 2
- 03Sentence-to-source and provenance-coverage tracking.Week 1, Week 2
- 04Review-record ingestion.Week 2
- 05Record, goal and employee binding.Week 2, Week 3
- 06Source checks and review routing.Week 3
- 07Manager signature workflow.Week 3, Week 4
- 08Review-record integration.Week 5
- 09Wording and evidence cases.Week 4, Week 5
- 10Guardrails and drafting controls.Week 4
- 11Review-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 team, one review cycle | ProductionProduction review workflow | AdvancedMultiple teams / cycles |
|---|---|---|---|
| Introduced at Pilot | |||
| Review drafting to your competencies | ✓ | ✓ | ✓ |
| Named manager signature | ✓ | ✓ | ✓ |
| Prior-review baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by rating class | — | ✓ | ✓ |
| Manager review workflow in your systems | — | ✓ | ✓ |
| Approved write-back | — | ✓ | ✓ |
| Goal-and-record integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Conflicting source records | — | — | ✓ |
| Cross-team calibration packs | — | — | ✓ |
| Large review populations | — | — | ✓ |
| Multi-manager wording controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, review 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 review cycle and the competencies each report is rated on
→Record capture and source versioning
Week 102Representative goal, note and telemetry sources
→Source binding, drafting logic and the prior-review baseline
Week 203Your review calendar and the managers it names
→Competency mapping, source binding and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Goal, note and telemetry source assessment, then integration setup
Week 205Sentences you would not want exhibited
→Wording cases and the failure round
Week 406What no review may assert
→Source checks, review routing, guardrails and release controls
Week 307A named manager who signs the review
→Manager sign-off workflow, 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
Two of these phases do share the fifth week, and no band was widened to make the column look neat.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Cycle discovery, competency mapping and the automation boundary
W2Source integration and the prior-review baseline
W3Review drafting, source logic and release controls
W4Evaluation suite, wording cases and failure-mode testing
W5Review-record integration, pilot drafts and targeted corrections
W6One review cycle run under the named manager, then Agent Care handover
- Reading the band
- Each bar spans only the weeks its own work is named for, and week five carries two.
- At the end of W6
- Once the review record validates, Agent Care picks the agent up.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · HR AI agent
Build a review drafting agent around the sentence your last cycle cannot account for.
Show us one review cycle your managers run and the notes behind the last review written. If nobody can say which sentences the manager composed, the file holds a document with a name on it and no author. A review no one can account for comes back as a finding.
- Typical delivery
- 4–6 weeks
- Deployment
- Your live cycle
- Named manager
- Record first
- After launch
- Agent Care available
Nestack Agents · Performance reviews
AGT-HR-09 · Agent Care available after launch