This AI recruitment startup wants to remove bias from the hiring process

Could AI help take the stress and time out of applying for a job, while also giving you tailored tips on how to perform better?

Job interviews aren’t fun — many find them daunting and often unsuccessful applicants are left not knowing where they went wrong.

But an Australian AI-powered recruitment platform is trying to change that, by using machine learning to reduce the amount of time applicants and hiring managers spend on the hiring process, and the likelihood of bias.

Sapia.ai is powered by machine learning and trained on 1.3 billion words from former job applicants. It boasts Qantas, Woolworths Group, Big W,? Suncorp and Medibank Private as its clients.

The startup raised $17 million last year — backed by Macquarie Capital & W23, the investment arm of Woolworths, to help it expand into the lucrative US jobs market.

Sapia.ai Founder and CEO Barb Hyman talks to Tech Wrap-Up Australia about how chatbots make applying for a job easier, how AI can reduce bias in the hiring process and why the resume is dead.

Barb, what is your technology trained on??

Five years ago we started to build this product based on chat as the medium, recognising that people feel more comfortable on their phones. We had to figure out the signals that could be extracted from that language data — your personality, your communication skills and your behavioural traits.

To train our model we use only first party data, which is data that we have collected from people using the product. This is done via a structured interview, much like you might get interviewed for a job.

The model has been trained on 1.3 billion words so far.

This is data that is proprietary to us — we own it as part of our partnerships with our clients. It doesn't have any Personal Identifiable Information (PII) data. It doesn't know you, your name, your skin colour or your demographics.

It is the purest and cleanest data set that you can use in order to inform an AI system, in contrast to third party data, which is social media data or resume data. Third party data contains a lot more latent signals of privilege that may lead to bias.

How does a job applicant use the software??

As an applicant, you effectively have a chat conversation in a web browser. It doesn't matter when you do it and it is untimed, so that people feel safe and don’t get too anxious. Five minutes after you submit your answers you get a response containing your personality profile. This also gives you coaching and advice on future applications, whereas in many cases job applicants are simply ghosted.?

If the applicant scores highly relative to what the company is looking for then they make it to the next round, which is a recorded video that gets uploaded and sent to the employer. From there it is up to the employer if they bring you in for a final interview. With this process, we have taken away all the steps that a recruiter might do for screening a resume. We are giving an employer a short list of qualified job applicants.

? Click here to subscribe?to the latest tech and startup news and analysis on LinkedIn.

Does the employer set the criteria or do you?

We are effectively a screening, interview, assessment and feedback tool all in one. The assessment incorporates traditional behavioural science methods, like psychometric testing. But we are using machine learning and the power of incredible processing speed to do it at scale for recruiting.

We use rule-based models defined by the client — for example, someone who is high in critical thinking or a great communicator. Those qualities are assessed based on the language an applicant uses. Because we use language data and no demographic data, we are starting off with a very clean, transparent, explainable model.

Anyone who is hiring applicants where people skills matter, such as soft and communication skills in retail, would be more likely to use this. But if you're hiring engineers, you wouldn't use us to assess someone’s coding skills.

In the past you headed up recruitment and HR at REA Group and Boston Consulting Group. How did those experiences inspire you to create an AI recruitment platform?

In all the companies I've worked with, your people are your most valuable asset and your most significant cost. But coming from consulting on the strategy side, it struck me how little data was used to make decisions around talent, whether it's who you hire or who you promote. I heard time and again the bias that came out of people's mouths around that.

They would judge applicants a lot on where they went [to school], or on their body language, or the amount of time they had spent in a previous job.

But in consulting, you don't make recommendations to a board without a lot of analysis and a lot of objective data. So I thought there must be a way to bring more objectivity, more data.

There’s a lot of concern that AI (especially generative AI) reinforces existing biases. How can you be sure Sapia.ai is removing bias from the hiring process?

Probably the best example [of bias in AI] recruitment is what happened with Amazon, where they experimented by building an algorithm to help with hiring. The algorithm factored in the CVs of those who had already been hired, and compared that data set to the CVs of those who applied. The problem there was that the algorithm was biased before they even started to use it.? As a result it turned out that they mostly hired men.

This is why we don’t use any historical decisions or data to build our models. We use rules-based models. You can identify where you might be impacting on equity before you build and deploy the model.

There is never any disparate impact or adverse impact in our models because you can do that testing before you release it.

We do still see bias in human decisions, but the bias in the process has been removed. At the end of the day it is still up to the employer, but the shortlist that is presented to them is an unbiased shortlist.?

Do you have technical knowledge in AI and machine learning? And do you need to have that to run a tech company?

I don't. I think it's about the team that you have around you. I have the best engineering team in Melbourne, if not Australia. I think the diversity of perspective is what has built the company.

Plenty of startups and tech companies are re-branding themselves as AI-related as awareness of the technology increases. Is that a problem?

It is a technology that definitely carries risk, so a big part of what I do is help leaders navigate the space and really understand it.

One question I would ask a company calling themselves AI-related is do they have any data scientists in their business? If they don't, there isn't really any AI there. Do they have a tech manual that describes how the AI works? Do they use model cards? Model cards were invented by Google as a way of providing transparency around predictive models. If they don't have those, I would suggest there aren't any models.?

I would also ask them to show me their bias testing regime and their AI governance regime. If you don't get satisfactory answers to that, then it's probably a bit of puffery going on.?

You launched Sapia.ai in the US earlier this year. What are the advantages of being there?

It is a market where diversity and inclusion really matters. From a scale perspective, it is good for our company growth and our data [collection]. It is also a really smart market. The smarter companies and the fast followers can really differentiate between true and ethical AI versus the puffery that I mentioned earlier.

How do you hire your employees? Can you share any advice for finding good staff?

Don't use a resume. Take the approach of a structured interview. You have to have something that is consistent where you're measuring people against a set of criteria, and you have to have more than one person do that in order to de-risk bias. That is effectively what we have scaled as a product.

?? Want to learn more about AI and recruitment? Follow? Barb Hyman on LinkedIn.

? Click here to subscribe?to the latest tech and startup news and analysis on LinkedIn.