The database you already paid for
Most agencies hold thousands of candidate records built over years, and consultants search it by keyword. Good candidates placed two years ago are invisible because nobody remembers the right search term.
Improving retrieval over that existing database is usually the highest-value, lowest-risk application available. It is a search problem before it is a prediction problem, and better search on clean data beats a model on messy data every time.
Operational predictions worth making
- Time to fill - which roles will take longest, so resourcing matches difficulty rather than order of arrival
- Vacancy fill likelihood - whether a role is realistically fillable on the terms offered
- Client conversion - which enquiries become instructions
- Candidate availability - who is likely to be open to a move, from engagement signals
- Placement retention risk - where a placement may not survive the guarantee period
None of these makes decisions about individuals in a way that affects their employment prospects directly, which keeps them clear of the hardest legal territory while still improving the economics.
Candidate ranking needs real care
Ranking candidates for a role is the obvious application and the one carrying most risk. A model trained on past placements learns who was placed before, and if past hiring favoured particular groups, the model reproduces that.
| Risk | Mitigation |
|---|---|
| Learning historical bias | Test outcomes across protected groups before use |
| Proxy discrimination | Audit features that stand in for protected characteristics |
| No explanation available | Require per-candidate reasoning before deployment |
| Regulatory exposure | Take legal advice; several jurisdictions now regulate this directly |
Postcode, school, career gaps and years of continuous employment are all common features that can act as proxies. A gap penalty in particular disadvantages people who took parental or carer leave.
Our strong preference is that such a model surfaces candidates for consultant review rather than filtering anyone out. Expanding the shortlist is defensible; silently shrinking it is not.
Parsing CVs is harder than it looks
CV parsing is a solved-looking problem that is not solved. Layouts vary enormously, job titles mean different things across companies, and dates are written a dozen ways.
Expect meaningful error rates, and design so errors are correctable by the consultant rather than silently wrong. A parsed field a consultant can fix in one click is useful; one buried in a database is a data quality problem accumulating quietly.
Where to start
For most agencies the sensible order is: fix search over the existing database first, then add time-to-fill and conversion prediction to improve how consultants spend their week, and only then consider anything that ranks people - with legal input from the outset.
The candidate you already have on file and cannot find is the cheapest placement you will never make.