Govern · agent risk record · Human Resources
Attrition Signal Agent
depth-agent
Risk class
High risk
EU AI Act Annex III(4)(b) and GDPR Article 22
Status
Operating with a gap
A2
Oversight
HRBP for the population
HR business partner
Explanation
On request
Disclosed only if the employee asks or if a business partner raises it
Appeal path
Named human
240h service level
Conformity
Self-assessed
assessed 141d ago
Declared gap
Employees are profiled for flight risk without being told. The opt-out exists because one employee appealed and asked for it, not because it was designed in. Review is 51 days overdue.
Classification
Why this agent sits in this class, under which regime
Regime
EU AI Act Annex III(4)(b) and GDPR Article 22
Class
High risk — Affects access to employment, credit or a comparable material interest. Requires human oversight, logging and contestability.
Basis
Employment profiling. Predicting who will leave is profiling of workers, and the flag is capable of affecting how a person is treated.
People affecting
Yes — the output changes an outcome for a named person
Scope
human-resources · human-resources/exit-management
Registered
190d ago
Log retention
84 months of decision logs retained
Conformity
Self-assessed — Assessed by the operator. Not independent.
Human oversight
The measure, and the person it is attached to
Measure
Flags are visible only to the HR business partner, expire after 60 days, and are excluded from promotion, merit and slate contexts by deny list. Employees may opt out.
Named human
HRBP for the population · HR business partner
Autonomy
A2
Transparency notice
Not served
Explanation surface
On request — An explanation is produced when the affected person asks for it.
Explanation reference
Disclosed only if the employee asks or if a business partner raises it
How an affected person contests this
The route, the named human at the end of it, and the service level
1 appeal on record
Route
Priya Raman, Chief People Officer
Path type
Named human
The appeal lands with a person, not a queue.
Service level
240 hours
No appeal currently open against this agent.
| Reference | Subject | Grounds | Routed to | Age | Review | Outcome | Evidence |
|---|---|---|---|---|---|---|---|
APL-0007 Closed | EMP-12008 employee · US · HR case management | Employee learned of a retention flag during a conversation with their business partner and objects to the profiling. | Priya Raman Chief People Officer | 96h of 240h | 52m human time | Overturned Flag withdrawn and the employee opted out of future flagging. The opt-out is now offered to the whole population, which it was not before this appeal. | HR9A-1012 #1 |
Adverse-impact readings behind this agent
Attrition risk flagging · Aug 2026
Monitoring status
Review overdue
Population a month
3,840
Review cadence
Quarterly
Priya Raman
Last reviewed
141d ago
| Cohort | Compared with | n | Selection rate | Reference rate | Impact ratio | Reading |
|---|---|---|---|---|---|---|
Under 30 age-band | 30 to 39 | 1,037 | 17.1% | 13.5% | 1.26 | Outside threshold Under 30 is selected at 126 percent of the rate for 30 to 39. For this decision being selected is the adverse event, so the test is the reciprocal of the four-fifths rule. |
50 and over age-band | 30 to 39 | 538 | 12.3% | 14.1% | 0.87 | Within threshold 50 and over is selected at 87 percent of the rate for 30 to 39. For this decision being selected is the adverse event, so the test is the reciprocal of the four-fifths rule. |
40 to 49 age-band | 30 to 39 | 883 | 13.3% | 14.1% | 0.94 | Within threshold 40 to 49 is selected at 94 percent of the rate for 30 to 39. For this decision being selected is the adverse event, so the test is the reciprocal of the four-fifths rule. |
Women Gender | Men | 1,690 | 15.8% | 13.7% | 1.15 | Within threshold Women is selected at 115 percent of the rate for Men. For this decision being selected is the adverse event, so the test is the reciprocal of the four-fifths rule. |
Under 2 years tenure-band | 2 to 5 years | 1,306 | 18.0% | 13.7% | 1.31 | Outside threshold Under 2 years is selected at 131 percent of the rate for 2 to 5 years. For this decision being selected is the adverse event, so the test is the reciprocal of the four-fifths rule. |
Over 10 years tenure-band | 2 to 5 years | 653 | 10.5% | 14.2% | 0.74 | Within threshold Over 10 years is selected at 74 percent of the rate for 2 to 5 years. For this decision being selected is the adverse event, so the test is the reciprocal of the four-fifths rule. |
Open alarms on this decision point
Derived from the readings, never authored by hand
FAIR-0013Sustained driftMedium severityAcknowledged
The ratio for Women has moved from 1.06 to 1.15 across 6 periods without ever breaching. This is inside tolerance and heading out of it, which is the reading a threshold alone would miss.
Owner Nadia Kovac · opened 170d ago
FAIR-0014Never reportableLow severityMonitoring
No reading has ever been produced for Non-binary and self-described on this decision. The cohort is declared in scope and the population exists, but it has never once been large enough to report against Men at the confidence this program requires.
Owner Nadia Kovac · opened 170d ago
FAIR-0015Threshold breachHigh severityOpen
Under 30 is being selected at 126 percent of the rate for 30 to 39. For this decision selection is the adverse event, so anything above 125 percent breaches. The breach has held for 1 of the six periods.
Owner Priya Raman · opened 17d ago
FAIR-0016Threshold breachHigh severityOpen
Under 2 years is being selected at 131 percent of the rate for 2 to 5 years. For this decision selection is the adverse event, so anything above 125 percent breaches. The breach has held for 3 of the six periods.
Owner Priya Raman · opened 78d ago
FAIR-0022Review overdueMedium severityAcknowledged
The documented review cadence for this decision is quarterly. The review fell due 51 days ago and has not been held. Readings continue to be produced; nobody has looked at them.
Owner Priya Raman · opened 51d ago
What this record is, and what it is not
Read this before quoting anything above
Fairness readings computed on modeled data demonstrate the instrument, not a real audit result. The classification, the oversight measure, the explanation surface and the appeal route are design decisions recorded here so they can be argued with; the population behind the readings is modeled.
No agent on this register carries an independent third-party conformity assessment. Every assessment on file was made by the operator.
Appeal outcomes are recorded against the evidence spine where a unit reference exists. Where the evidence column reads “not recorded”, the appeal was resolved outside the spine and that is a gap, not a rounding.