Data-Driven Recruitment

Nearly every recruitment dashboard reports activity: applications received, time to fill, interviews held, offers made, cost per hire. These are worth having and they describe how busy the function was, not how well it chose.

Outcome measures describe what the hiring produced: retention at twelve months, performance after a year, whether the hiring manager would hire the same person again, time to productivity. These are harder to collect, arrive late, and are the only ones that answer whether the hiring was any good.

The reason the distinction matters is that activity measures can all be improved by lowering standards. Time to fill drops if you accept the first adequate candidate. Cost per hire drops if you stop using the channels that produce the best people. A function optimising only on activity will make itself look better and the organisation worse.

The practical rule is to pair every activity measure with an outcome measure and never report one without the other.

If an organisation does only one piece of hiring analysis, it should be source quality.

The common version compares sources by volume and cost per hire, which rewards whichever channel produces the most applications most cheaply. The useful version compares them by what happened to the people hired: retention at a year, performance, and whether they progressed.

The result is usually uncomfortable, because the cheapest high-volume channel often has the worst retention and the most expensive channel frequently pays for itself. Referrals typically perform well on retention, which is a genuine finding and also a caution, since referral-heavy hiring tends to reduce the diversity of the intake.

The most common analytical error in recruitment is treating a difference between small groups as a finding.

An organisation that hired eleven people through one channel and nine through another, with retention of eight and six, has not learned that the first channel is better. That difference is comfortably within what chance produces, and acting on it is guessing with extra steps.

This affects nearly everything a mid-sized organisation might want to analyse: which university, which assessment, which interviewer, which source, broken down by role type and by year. The cells empty out quickly.

Three practices help. Aggregate across longer periods before comparing. Compare groups rather than individuals, since one interviewer's scores across forty candidates say more than four candidates across ten interviewers. And state the count alongside every percentage, because a percentage without a denominator invites exactly this error.

Where the numbers genuinely cannot support the question, the honest answer is that the data does not say, which is more useful than a confident answer that is noise.

The most consequential use of hiring data is scoring or ranking candidates, and it carries a specific hazard.

A model built on who was hired and who succeeded historically learns the historic pattern. Where past hiring favoured particular backgrounds, universities, career shapes or groups, that preference is in the training data and the model will reproduce it, at scale and with an appearance of objectivity that makes it harder to challenge.

Removing the obvious fields does not solve it, because proxies remain. Postcode, university, career gaps, the wording people use and the schools they attended all carry information about background, and a model will use them.

The disciplines worth applying are to check outcomes by group rather than inspecting the inputs, to keep a human decision in the loop with a recorded reason, and to be able to say what a scoring system is using and why. A system nobody can explain is one nobody can defend if the pattern of its decisions is questioned.

The same caution applies to keyword filters in an applicant tracking system, which are a simple version of the same thing and are rarely reviewed at all.

Most of the value here is available without a analytics programme, provided a few things are recorded consistently.

• Record source accurately at application, and resist the catch-all option, since it is usually the largest category and tells you nothing.
• Record rejection reasons against defined criteria, because a funnel without reasons cannot explain where or why candidates are lost.
• Keep the link between the hire and the employee record, so retention and performance can be read back against how they were hired.
• Record who assessed, so interviewer effects are visible.
• Report counts alongside rates, always.

The single highest-value item is the third. Organisations whose recruitment data ends at the offer can never answer whether their hiring is good, only whether it was fast. Connecting the hire to what happened afterwards is what turns recruitment reporting into recruitment analysis.