The core problem in one paragraph
A model that learns from your previous hiring decisions learns the patterns in those decisions, including patterns you would not defend. It then applies them efficiently and at scale, with a rationale nobody can fully explain.
That is the discrimination risk, and it is not solved by removing names from CVs — models find proxies.
What is genuinely safe and useful
- CV parsing into structured candidate records
- Search over your own database, surfacing candidates keyword search misses
- Interview scheduling across multiple diaries
- Drafting job adverts and communications for human review
- Summarising interview notes for the panel
- Compliance chasing — references, documents, checks
None of these decides anything about a candidate's progress, which is what keeps them safe.
Surfacing versus filtering
Ranking a longlist so a recruiter looks at promising candidates first is defensible. Automatically rejecting anyone below a score is not, and it removes candidates who would have been hired.
Keep humans making every decision that affects whether a candidate progresses, and make that real rather than a rubber stamp on a machine's ordering.
If you use scoring at all
- Test for disparate impact across protected characteristics before deployment and periodically after
- Document how it works in terms you could explain to a candidate or a tribunal
- Keep a human decision at every progression stage
- Tell candidates which parts of the process are automated
- Retain the records needed to defend a decision
Video and personality analysis
Automated assessment of facial expression, tone or personality from video has weak evidence behind it and has attracted regulatory attention in several jurisdictions.
We would not build it and would advise against buying it. The efficiency gain does not justify the risk or the candidate experience.
The candidate side
Candidates use AI to write applications. Application volume rises and average quality converges, which makes CV screening less informative rather than more.
The response is to weight structured assessment, work samples and conversation more heavily, rather than to escalate an arms race in screening.