T3
Demo workspace · fictional company and sample data · not client results

Impact Assessments

How each AI system could affect the people it touches, and what Kestrova does about it.

4 systems assessed
SYS-01

Kestrova Shortlist

Signed off by system owner (week 19, after internal audit finding)

Help recruiters prioritise applicants by ranking them against a role profile, with an AI summary of each application. A recruiter makes every hiring decision.

Affected groups
  • Job applicants
  • Customer recruiters and hiring managers
Foreseeable misuse
  • Using auto-decline to reject applicants with no human review
  • Relying on the AI summary instead of reading the CV
Potential effects
  • Applicants lose an opportunity without human review
  • Unequal outcomes if rankings or summaries are less accurate for some groups
Controls
  • Auto-decline off by default, with warning and audit log
  • Summaries labelled as AI-generated with a link to the CV
  • Sales and product wording corrected to describe how the system actually works
Residual concerns
  • Customers can still enable auto-decline after the warning
  • Summary accuracy depends on quarterly sampling
Linked risks
SYS-02

Kestrova Retain

Signed off by system owner (week 19, after internal audit finding)

Help customer HR teams decide who to prioritise for retention conversations by predicting the likelihood of an employee leaving in the next 6 months.

Affected groups
  • Customers' employees
  • Customer HR teams and line managers
Foreseeable misuse
  • Using scores in redundancy, disciplinary or pay decisions
  • Treating a prediction as a fact about an individual
Potential effects
  • Discrimination through proxy features (health, sex)
  • Job loss or unfair treatment if scores are misused
Controls
  • Sick-day and parental-leave features removed; model retrained
  • Feature approval rule for protected-characteristic proxies
  • Terms of use and in-product warning on permitted use
  • Training-data register
Residual concerns
  • Contract wording on training data still with General Counsel
  • Misuse can only be monitored, not fully prevented
Linked risks
SYS-03

Kestrova Interview Notes

Signed off by system owner

Save interviewers time by transcribing video interviews and drafting a structured scorecard that the interviewer completes and confirms.

Affected groups
  • Job applicants (interviewees)
  • Interviewers and hiring managers
Foreseeable misuse
  • Submitting AI-drafted scores without reviewing them
Potential effects
  • Lower scores for candidates whose speech is transcribed less accurately
Controls
  • Drafts labelled as AI-generated
  • Interviewer must confirm each score before submission
  • Quarterly sampling of 50 transcripts against recordings
Residual concerns
  • Transcription accuracy depends on the vendor's model
Linked risks
SYS-05

Kestrova JobAd Writer

Signed off by system owner

Generate a first draft of a job advert from role details entered by a customer recruiter, who edits and publishes it.

Affected groups
  • Job seekers reading published adverts
  • Customer recruiters
Foreseeable misuse
  • Publishing a draft without review
Potential effects
  • Adverts with wording that can signal discrimination (e.g. by age), deterring applicants
Controls
  • Blocked-terms check and bias-wording review step (added after the week 12 incident)
  • Beta expansion paused until review complete
  • Drafts labelled as AI-generated
Residual concerns
  • Blocked-terms lists need regular updating
Linked risks
None in the register excerpt