The reference check nobody really does

As it is usually done, with two names supplied by the candidate, a ten-minute call and the question “would you recommend them?”, a reference check teaches you almost nothing, and the research can no longer even say how little. It becomes useful again on one condition: stop collecting an opinion and start verifying facts, with questions written in advance as you would for a structured interview.

It is the step everyone ticks and nobody prepares. It comes at the end of the process, when the candidate has said yes, the client is waiting and every day counts, which makes it the perfect formality: you call, you hear good things, you sign. The trouble is not that reference checks are useless by nature. It is that we run them in the one way guaranteed to make them useless.

What does a reference check actually predict?

Not much, and that little is no longer measured. Schmidt and Hunter’s meta-analysis, published in 1998 and quoted ever since in every recruitment textbook, gave reference checks a validity of 0.26, meaning a modest correlation with the performance observed afterwards.

In 2022, Paul Sackett and three co-authors revisited those estimates, correcting a methodological error that had inflated most of them. The most discussed result concerns general mental ability, which falls from 0.51 to 0.31. The most instructive result for us is quieter: reference checks receive no new estimate at all. They sit in the list of procedures set aside for insufficient information, next to graphology, because the original data no longer allow anyone to say what they were measuring.

The same revision does contain a gap that matters here. The structured interview, with questions fixed in advance and a scoring grid, comes out at 0.42 and ranks first; the unstructured interview drops to 0.19. These are not two different tools but the same conversation between two people, held with or without preparation. Nothing suggests that a call to a former manager escapes that rule.

Why does the usual call produce nothing?

Because each of the three people involved has a good reason to say nothing useful, and the question asked never forces them to.

The referee was chosen by the candidate, who has no reason to offer the name of someone likely to judge them harshly. The referee also knows one slice of the career, often an old one, and answers about what they saw. A study published in the International Journal of Selection and Assessment, covering more than 1,700 managers in the US and the UK, adds a less obvious reason: a manager rates a former team member more favourably for a recruiter than in private, because the team’s performance also reflects on them. The authors call it vicarious self-promotion.

Go around the chosen referee and call the company instead, and you hit the other wall. HR departments increasingly confirm dates of employment and job title and nothing more, to protect themselves against defamation claims, as SHRM noted in 2018. The instruction is written into an internal policy and the person applying it has no discretion, so calling back a second time changes nothing.

Then there is the question itself, which invites a yes and gets one every time you ask “would you recommend them?” of someone the candidate picked. What you end up with is a confirmation that confirms nothing, filed in the record with the same weight as a fact.

What French law requires before you pick up the phone

The candidate must know that references will be taken before they are, and the French Labour Code says so in two articles that are worth reading rather than paraphrasing.

Article L1221-8 provides that the candidate is “expressly informed, before they are used, of the recruitment methods and techniques applied to them”, and that those methods must be relevant to the purpose pursued. Article L1221-9 adds that no information about them personally may be collected through a process that has not first been brought to their attention. Article L1221-6, finally, limits what you may look for: information with a direct and necessary link to the job on offer.

The CNIL, France’s data protection authority, drew three practical consequences from this in its recommendation on recruitment. Collecting references from the candidate’s professional circle is allowed provided it is not done without their knowledge. Contacting their current employer without their express agreement is unfair practice. And a character investigation through personal referees is excessive.

The practice that holds up under these texts is simple: tell the candidate in writing that references are part of the process, ask them for the contacts, and never call the current employer without their agreement. The law does not forbid speaking to someone the candidate did not name, provided they know the step exists, but that is the area where a recruiter in a hurry most easily crosses the line.

Which questions actually produce information?

The ones about a fact the candidate asserted, which the referee can answer without having to give an opinion.

A candidate who said in interview that they led the migration of a billing system in a team of eight has made three verifiable claims. You do not ask the referee to confirm that they “led the migration”, which suggests the expected answer; you ask who led that project, how many people worked on it, and what the candidate did themselves each week. If the three answers match, you have learned something. If they diverge, you have learned more.

The second family of questions is behavioural, on the model of the structured interview: a specific situation, what the person did, what came of it. “Tell me about a time she had to announce a delay to a client” produces a story, with details that an obliging referee struggles to invent. The comparative question has its place too: “compared with the other people you managed in this role, where would you place her?” leaves less room for uniform praise.

These questions are written before the call, from the candidate’s file, and the generic part is the same for every candidate for the same role. That is what turns a pleasant conversation into a control, and it is exactly what project-based hiring processes almost never plan for, as Cobalt’s analysis of IT services firms hiring without a process fit for it shows.

A reference is also an identity check

The reference check sits at exactly the right moment to shut the door on fraud, and that is its most underrated function. We have argued elsewhere that a control belongs where it closes a door, before signature and before access is granted, and that 41% of companies have already hired a fake profile without realising it.

A fake profile often comes with fake referees, and the countermeasure needs no tool: you do not call the mobile number provided, you go through the company switchboard and ask for the person by name. A referee who does not exist in the organisation they claim to represent settles the question more surely than ten minutes of praise. For an IT services firm the stakes go beyond hiring, since the person placed will work inside a client’s systems.

What an agent prepares, and what a person calls

An agent can do everything around the call and nothing that gives it its value.

Before, it extracts the verifiable claims from the file, dates, scope, team sizes, the role declared on each project mentioned, and drafts the question grid for this particular candidate. It checks that the written notice was given and recorded, since that is the legal condition, and it looks up each company’s switchboard rather than trusting the numbers supplied. Afterwards, it records the answers against each claim and flags the discrepancies without interpreting them.

The call itself stays with a person, for two reasons that each stand on their own. It leaves the company and engages a third party’s word, which puts it on the side of what a person approves before it goes out. And the most useful information in a reference check is often the way someone hesitates before answering, which no transcript captures.

Two precautions follow from the same texts. An agent that went off on its own to gather information about the candidate from sources they know nothing about would be precisely the process article L1221-9 forbids. And the outcome of a reference check is not summed up as a score: it is written as refutable sentences, what was confirmed, what diverged, what could not be verified, for the same reasons that make the AI match score the wrong instrument.

That is the split we apply in Balt: the grid, the paper trail and the write-up on one side, the conversation on the other.

Should you still take references at all?

Yes, but fewer and better. The concession is real: a step whose validity is no longer measured costs everyone time, and a reference check run out of habit, two friendly calls at the end of the process, could disappear without the quality of hiring suffering for it.

What deserves to be kept is narrower and more useful. One or two references, chosen to cover the most recent experience and the one closest to the role, questioned on facts, at a point where a diverging answer can still change the decision. The rest, the general impression and the warm recommendation, is worth neither the recruiter’s time nor the referee’s.

That leaves a question of timing, which arises whenever a step is added at the end of a process: every day between the candidate’s yes and the signature is a day offered to another offer. So the reference check runs in parallel with the final stages, never after them, because the time it takes to fill a role is spent waiting far more than working.