Will Amazon's Recruiting AI Work?

[Transcribed from episode above]

Brad Owens

Hey everyone this is Brad Owens with the transform recruiting podcast coming at you solo today because something came out here recently that I wanted to talk about um and wanted to start a conversation around so according to recode. That viewed a confidential Amazon internal document. The Amazon is pushing out a whole bunch of recruiters. They've actually laid off a number of people that I knew personally because they've been developing Ai software that screens job applicants now. To to those that know me and know how I feel about Ai in the recruiting industry duh. That's going to happen more? Um, they may not have done the right job of training up their recruiters to do the things that a I can't do so there's some. Fault here on Amazon for letting some of these people go just because they purely didn't train them the right way. But I'm going to leave that aside for now what I wanted to talk more about is maybe where ai should play what Amazon is seeing and how I think that this might turn out here. So Amazon. They have always been this kind of this golden child of recruiterless hiring I'm using air quotes for those of you that are just listening to this podcast. You can also watch it on the website at transformrecruiting.com but they have always been that golden child.

01:38.30

Brad Owens

They had this entire process which actually went through myself if you haven't you should try it just to see what this whole thing is like you can apply for and be offered a warehouse job purely through just a online interview quiz type thing that they have they even have this little. Um. Almost like a I I wouldn't call it a test. It's more like a a trial run where you are put into this environment online where you have to take different stock put them on different shells based on different things now is really interesting and it was a good way to see if they had the basic seals to be able to do this job. You know, granted, it's not a interpersonal skills type job. It's simply just a pick and pack kind of position so it doesn't it's one of those that lends itself well to the Ai type hiring of hey can we so screen for this basic skill set that someone's going to need to do this job. Good. That is a place where I think ai, it's not exactly ai. Ah, where automation could play a role where you're simply just using these if then statesmen if this person can do this then you can offer them this job now. We're not going to get the most amazing hires in the world because of this we're going to have to play the numbers game because of the likelihood of us. Actually getting those people to start is very low case in point myself um I got through the entire process I got offered the job. Woo me? Ah, but I didn't want it in the first place. So I just started ignoring. Everything.

03:05.62

Brad Owens

Um, that was the wrong thing for me to do for my email inbox because Amazon's bots and their email program just kept hitting me and hit me hit me hit me for hey you're starting. It's coming up hey we've got open shifts hey this ed eventually went away because the marketing your software they use realized that I wasn't going to be the person that responded but as an example. That's a good way to use automation I like that now the way that Amazon has tried to use this in addition to that they are trying to identify those that may be good promotion material who should we actually elevate to higher positions after they come in from this automated warehouse. Type hiring system that we've got this no touch hiring. So how can we take that to the next level. Well they tried this back in the 2010 s they had a lot of discoveries that um people have been realizing that ah correlation. So how much hey this number in this number are. Closely related. That's interesting that doesn't mean causation that might those 2 numbers might not tell the whole story. Yes, they might be related. They might 1 might influence the other but that doesn't tell you why that doesn't tell you the cause well back in 2010 they realize that because they were just focused on such tiny little. Um, examples tiny little What am I looking for um, tiny little datasets that they were looking at that there was a lot of bias involved in this so they had a lot of underrepresented minorities that weren't getting put through the hiring process. They weren't getting put up for.

04:42.28

Brad Owens

Ah, promotions. This is a downfall of Ai Ai will cause that if you don't look at it closely. There is just a bias purely based on um, underrepresented minorities women in the workforce. It's just going to be a problem if you don't pay attention to it. It will be a problem. If you're automating something right now and you haven't looked at your data for underrepresented minorities and the outputs of that Ai look at it. You're going to be surprised. You're just going to be It's going to happen but there are a couple different problems that go into that or a couple different reasons that we created that problem for ourselves. Um, one. There's been bias in this system forever. You can't put Ai into a system that was created a while ago and expected just to all of a sudden do better than we've done. We haven't done a good job that underrepresented minorities have not been pushed through the system as they should have women are. Vastly underrepresented people of color are vastly underrepresented. Not only just in the positions they hold but the money they aren't we're trying to do our best on all of those I feel like the software itself or the Ai or whatever is actually doing this screening can only work with the data that it has so. When we are asking for this type of data where we can put demographics on who is going through a hiring system as you know in the recruiting process when someone submits their application. They have the opportunity to put in demographic data but they don't have to.

06:14.95

Brad Owens

A lot of folks that are from the underrepresented minorities. They um, they don't want to put down their demographics. They feel like that in the past and it's totally true may have affected things negatively for them. So yes, it's possible. That you don't even have data for these people that you're trying to create this ai system from so think about what data you do have then if underrepresented minorities aren't putting that data in your system. You have only data from those that are which are more than likely to be um, you know. White individuals Caucasian individuals from um, good backgrounds that are um I don't know how do I put this um those that are proud of what they're putting down I don't know proud's wrong word. Um those that have entered in this data. Just happen to be not um, underrepresented minorities. How about that. That's that's a better way for me to put that. Um, since that is the only data your system has to work from it's going to be biased. It's just going to be so the problem that I'm seeing and the reason I'm going through this description here is. I'm not surprised that Amazon's original model was biased because the data that they were doing their work from was already biased. You can't take a biased data set and all of a sudden surprise. It's unbiased in the ai that we're doing it just doesn't work that way. So the first problem is you got to fix your data.

07:45.34

Brad Owens

You have to have clean data if you're going to try to create Ai like Amazon's trying to create now they say in this article which I'll link to the one I'm reading is from vox the original internal document came from recode. Ah, so Jason do Ray I appreciate this this article because it's creating a great conversation for our audience here. Um, apparently the last two years or so they have been turning over some of their recruiters task to Ai. Ah that again specifically targeted for. Um, taking job applicants from across corporate and warehouse jobs who have been successful in a given role. They're going to fast track them to a position that would give them an increase so it doesn't have to be just in warehouse they're doing office workers as well. Um. That's they're calling right now automated applicant evaluation. There is a specific group within Amazon's hr division that is the artificial intelligence recruitment team. Fantastic I'm really glad that they're doing this that they're putting people and resources behind this this is the use. It's fantastic. Um, but again. The 2010 one was biased so they stopped that one while they tried it again and now the hr division believes that these machine learning models that they have have been successfully guarding against biases. Well they say just on race and gender. That's fine. It's better. Nothing good. Um.

09:13.74

Brad Owens

I feel like I want to know more about where this data came from how they got this kind of data. Um, because here's the problem if you say hey here is the profile of a person who has done fantastic in this job. We should now find people that are like this. And promote them through what are we going to find we're going to find that underrepresented minorities aren't those ones that have been tagged as the people who are doing a fantastic job because there's bias. Ah there's implicit bias built into our system whether they know it or not. Whether it's unconscious bias or not it is built into the system. So if you try and go from completely clean data. We'll just say you have clean data of those that have been promoted over the past five years and now you're going to go try to find profiles of those that are similar and promote those through you're going to have bias again. So it's interesting to me that Amazon day chart team has not seen that bias and that to me tells me that they have done a fantastic job of cleaning their data sources now. There's a lot of issues that go into that because that all comes down to if someone were if someone wants to self-report their demographics can't force people to do that I don't want to force people to do that. But I'd be curious to go more in depth with the Amazon team if anyone knows anyone what data they're actually using.

10:47.33

Brad Owens

How they're getting it and how they're ensuring that it's clean I'm very curious about that one. Um I can say to those of you that are listening that this is intriguing that you know they're like well what can we do for ourselves first try and address your data problem try and address everything that you may need. To ensure that you do not have a biased data set. There are those out there that can have conversations around that one of my neighbors actually just focuses in that. That's all she does for her job. So there are folks that will help you with that I would start at least figuring out where your data sources are some of the things that you feel would predict. Someone being good in their job. Once you have those coming together start just running some analysis on them see if you can figure out if we were to just take this set of job performance and come come up with who of these people has been promoted just. See if that's biased or not just run it through really quick just through an excel spreadsheet. It doesn't have to be Ai or machine learning or anything else just run it through just in a standard excel spreadsheet and see if you've got bias in there see if you've got underrepresented minorities. Not. Coming up with a proportion amount of the roles. That's a very basic way for people to start to so you know really see where there may be bias is built in and what type of data they should be paying attention to you might come up with one. That's.

12:17.30

Brad Owens

Completely flips the script and it starts saying that hey only focus on this and we'll going to start pushing more underrepresented minorities through that's bad. Be good to know, um and start working through until you find a data set with the particular pieces of data that you need to create an unbiased.

12:37.37

Brad Owens

Promotion potential number right? You're going to need to start from somewhere. Everyone's like oh we need a I for this? No you just need an excel spreadsheet and someone willing to take the time to go through it so you're going to have to find the right pieces of data. You're going to have to get them all into the same system at once which is. Large problem for most but I would say that if you have a few people that really think that this should be ah the way that your organization should go and they should have more Ai build a group of it doesn't have to be many 3 people that can start pulling this type of data scrubbed. Of course. As much as it can be and start coming up with hey can we have a reliable algorithm or reliable pieces of data that will show who would be good in this role. Just let them run loose with some excel spreadsheets find some data analysis folks that are graduating that tons and tons of them that need jobs. Um. Bring in a couple of those start with a data analysis team that can really go through the data with a fine tooth comb pick out those that would not be biased once you have that then we can start talking about hey what ai what machine learning can do this um. Yes, there are machine learning algorithms that will do this all the data scientists out. There are like there's boost that can do I know I'm talking to people that don't have those types of skills. So let's focus on just the basics run it through some different excel spreadsheets run it through a few different formulas see if you can come up with some that.

14:10.47

Brad Owens

Would provide good data for yourself. That's going to be really hard I understand it's going to be really hard hire those that are good at doing this. Maybe you will have your own story of recruiting firm x comes up with algorithm that will guarantee unbiased. Hires for your organization and congratulations. You just printed a ton of money and all it starts with is getting your data all onto the same place. So if I were to come up with a summary of this one. Let's say Amazon claims they have done it. Maybe in 10 years from now we'll look back and say nope that one was biased too. My thought is it's likely will be in some way because there's bias built into our system. It's very difficult not to have that it will happen. However, at least they're thinking about it. At least they are putting data scientists on this and if you too would like to have your own headline of hey we've come up with automation that's going to solve for this. We know how to reliably um promote those that are are good at your organization or we have a way to predict what of our candidates are going to be good ones that we should reach back out to. If you want to do that start with your data streams start trying to understand where your data is coming from start collecting that all into 1 central repository. You can call it your data like you can call it your system of record. Whatever you need to you need to have all this data going through the same place if you're going to analyze something to automate it.

15:41.90

Brad Owens

Has to be clean data and it has to go through the same spot so start looking at what's called system integrators start understanding what software is out there that can deal with endpoint management. So that if you have um. 1 piece of data in 1 system and 1 and another and you need to pass them back and forth and point management's fantastic if you need to call out to different things to automate those and point management's great system integrators are perfect. You can do this. It is totally possible. There's technology out there that does this it just simply takes starting the conversation. Start the conversation internally you can automate things. But if you need to have a a goal of how to get here start the data conversation. Get your data into um, a usable format find some sort of system of record or that central repository. Once you start with good data. A lot becomes possible a whole lot becomes possible I have a another podcast that you should go and listen to and look up about the largest value of your recruiting firm. You know we started out with contact info and then we got into the. Relationships and the actual placements that we've done but those are transient I think all the data comes from your transactional data and if you can't have that in a single repository. You're you're not going to be able to make use of the most value that you've got at your firm. So let's focus on the data. All of this entire episode the past 17 minutes

17:12.95

Brad Owens

Hey y'all focus on data. Get it into the right spot you have recruiters out there that are using excel spreadsheets still you've got siloed data everywhere that is not helpful if you have any types of transactions occurring outside of a single system that is data that you could be using. To automate successful algorithms and successful processes so focus on your data, get it all into 1 single repository just have the conversation. 1 of the places that I've started with others before is to map out every single data source. You could possibly think of talk to your recruiters. Talk to sit down beside them figure out how they're doing their job where are they keeping their data sit down with your managers where do they need reporting to come from what sources are they looking at you will start understanding where all this data is coming from once you have that then we can start this conversation of let's integrate. It. So. I know a lot of you out. There are very nervous about where this data is coming from the only way to fix that is to start. It's just a start right? No one's holding you back except maybe the budget for the next year so work on it work that into the budget right? Let's figure out where all our data is coming from because that's our superpower that is what's gonna help us. So. Um, if you want to be the next Amazon focus on your data if you've got any questions about how to do that. You'd like to come have conversations around how you've maybe approached that or if you know anyone at Amazon that may want to come on and have a question about what data they're actually using and how they started this whole thing. Let me know reach out. My email is helloatbradowins.com.

18:46.11

Brad Owens

Ah, this is just one of many different podcast episodes. You can find at transformrecruiting.com where we talk about all of the thoughts ideas and tech that are and people I guess that are changing the future of recruiting so come have a listentransformrecruding dot com. Until next time my name's Brad Owens you can email me at hello@bradowens.com. Thanks so much for listening in and I'll talk to you soon.

Previous
Previous

What Should You Automate? with James Lawton

Next
Next

Help the Newbies with Michael Maison