If you work directly or indirectly with HR (and even more so in recruiting), you may recall this story…


#LongStoryShort… what is it really about?

A few years ago, Amazon conducted an experiment to automate CV pre-selection! The idea for Amazon was to develop an algorithm that could receive 100 CVs, rate each out of 5 stars (a bit like we evaluate books) so that the algorithm could produce, every time, the 5 candidates with the highest probabilities of succeeding in the company! #GuaranteedTimeSaver

And then… Bust! We find out that it doesn’t work! 😱

Worse still, the algorithm tended to rule out – almost systematically – women’s CVs. #NotFair

What lesson can we learn from this failed experiment?

An algorithm isn’t inherently sexist… it is taught to be.

First, it is crucial to understand that an algorithm is never biased/sexist/racist/homophobic/… itself. It still only constitutes the formalisation of a recruitment process parameterized by human beings.

In the specific case of Amazon, (1) the A.I. was trained using data from the CVs (2) of people already employed within the company, (3) who were recruited by a team of humans in HR.

#1 – First the question of CVs…

If we program AI using data from CVs, the “initial knowledge” that we provide it relates, on the one hand, to the academic backgrounds of the people (degree[s] obtained, which school[s]), and on the other hand, to their professional career (positions held, for how long, in which companies, in which fields). #EndOfStory

#2 – The nature of the sample…

Amazon – at least at the time – as a company had a 60:40 split of male vs female employees… Obviously, men being over-represented in the initial sample, this will affect the constitution of the rules for configuring the algorithm. #Logical

#3 – Finally, the HR team’s own biases.

Humans are fallible; we are subject to many biases (several dozen) both in our understanding of information and in the way we make decisions. The algorithm only modelled the recruiting process, as it had been applied by Amazon teams over several years. It therefore reproduced the biases that are inherently human.

⚠️ Important: I don’t pass specific judgment – whatever it may be – on the work provided by these teams. If we were to run the same AI on the CVs in most companies, we would arrive at a generation of algorithms that discriminate against women (at least with regard to their access to management positions and higher), people over 45 years old, disabled people, even people whose names “sound foreign”… (to name just a few examples).😏

So should we put an end to algorithms in recruitment?

No one will deny that there is a HUGE problem here. Depriving yourself of HALF the available talent is a HUGE problem for any company… Not to mention discriminating against women, solely on the grounds of their gender (a criterion absolutely unrelated to their ability to occupy the job(s) offered).

However, drawing the conclusion that algorithms are in inherently dangerous would be a terrible misconception. An error equivalent to saying that “it is inconceivable to drive a car” if one had read in the newspaper about a fatal car accident.

Wouldn’t the real problem actually be… selecting on the basis of CVs?

Today, almost everyone – recruiters included – agree that CVs are not a reliable predictor of people’s ability to succeed and thrive in the workplace. Additionally, many studies carried out on the subject have all come to the same conclusion!

And yet, in so many companies we still continue to preselect candidates on the basis of their CV… 🧐 #FindTheProblem

By poorly selecting information with little (to no) predictive value (the information contained in a CV), which moreover comes from a male-dominated sample and then running it in through A.I…. What else could we expect? 🤷🏻‍♂️

Yet TODAY, any company has the means to significantly improve their pre-selection…

Of all the companies that I frequent and collaborate with (and there are many of them), there is not a single one that has managed to make a significant improvement in the quality of their recruitment ( I mean in a objective, measurable way) relying solely on their internal resources. And I’m not saying that to point fingers at them… It’s just that they have other things to do with their already busy days! #Priorities

So, how should we go about doing things?

Numerous studies carried out all over the world attest that it is BY FAR more relevant to select your candidates (or employees when it comes to internal mobility) on the basis of “who they really are, meaning their potential. Particularly, by using assessment systems (recruitment tests). Note that these systems can perfectly complement the elements you already have in place if you wish…

Logical… Look at what we have hanging in the balance:

  • On the one hand, a CV: academic and professional background
  • On the other hand, who this person (really) is : intellectual agility, ability to learn, deep motivations, personality and behaviours in a work environment

In your opinion, what will allow you to best predict the ability of your candidates to succeed in a job? To integrate into the existing team? To adhere (or not) to the values of your company as well as to the company’s project?

You can automate (at least in part) your screening processes, whilst avoiding the “Amazon effect”.

But beware: using “recruitment tests” alone (to assess the characteristics of “Potential”) will not guarantee your recruitment success by itself. We will have to go a bit further than that…

Predictive models: the central part of predictive recruitment algorithms.

“Tests” are used to collect reliable data (as long as the tests you use are reliable) on each of your candidates. But this data, the “profile” of each candidate, must be used at some point in comparison to a model of expected criteria! It is precisely this comparison (bringing together the profile of candidates and a profile of expected criteria) that can be modelled through an algorithm.

This model of expected criteria is said to be “predictive” when there is evidence linking the criteria that are defined (personality traits sought for example) and the variable(s) that you wish to predict (such as performance or even adjustment to a company culture).

How do we create a predictive model ?

To create a predictive model, all you need to do is evaluate a sample of the population for whom you want to predict performance (for example). For example, you could ask each of your employees in any given post(s) to complete a personality, motivation and/or reasoning questionnaire. This is step #1.

Then (Step #2), you will assess each of them on the performance criteria that you want to anticipate in your future candidates. Basically, you have to say for each of your current employees how well they are performing. Typically, this should be rather obvious to score…

Except in some companies, we will sometimes hear:


“Well, we don’t know, we don’t have the data!”


⚠️ DISCLAIMER: Let’s be clear… If you cannot – internally – assess the performance of your employees (ie to say “who does the job well and who does not” in a given position), NOTHING or NOBODY could ever help you to significantly improve the quality of your recruitment process.

You will be able to improve your Candidate Experience and get stats on your candidate funnels but without this ability to objectively understand performance of teams, you will NEVER be able to improve your ability to predict the success and/or commitment of your candidates… #TheTruthHurts

If an expert comes to see you and explains – without asking for or having access to this type of information – that they have a formula that will allow you to recruit better, they’re just selling snake oil. #WalkAway

“If my company has up until now recruited mainly men (or women, or candidates only from the largest schools…) isn’t there a risk of creating an Amazon-like bias?”

Actually, NO… because psychological and behavioural criteria are extremely well distributed within the population, regardless of gender, age, origin or many other criteria.

As a result, this approach allows you to directly bypass this type of bias!

Results from predictive recruitment

On average, companies that use predictive recruiting select employees are 20% more likely to be successful than those they select using a more traditional approach.

They speed up their recruitment process by 30%, lower their costs by around 20% and cut their turnover in half for certain sensitive positions.


Is it magic? ✨🔮✨


Not in the least… it’s just that they started recruiting their employees by being interested in who they really are (and by measuring it precisely) in addition to (or rather than) focusing exclusively on their CV… Nothing more, nothing less!

And guess what: when we take an interest in who people really are, we get better results in the end!

Crazy, huh? 🤓☝️