TRANSCRIPT 🔗
WEBVTT
00:00:01.679 --> 00:00:04.240
Hello and welcome to another coffee with developers.
00:00:04.320 --> 00:00:06.400
And today we thought we'd do something different.
00:00:06.639 --> 00:00:14.480
I had a lot of questions lately from people that are applying for jobs and actually are losing their job or getting uh getting like downsized.
00:00:14.640 --> 00:00:18.399
And there's a lot of questions in the world of uh work right now.
00:00:18.559 --> 00:00:33.280
So, as you know, part of the Dev Digest of the newsletter has a work section as well, and there's interesting news every every week, like companies asking for token maxing, then companies not giving people access to their AI services, people being asked to use AI although they will want to.
00:00:33.439 --> 00:00:34.640
There's all kinds of stuff there.
00:00:34.799 --> 00:00:40.960
So I thought instead of you asking somebody from a technology company who actually will say that, of course, AI is a great thing to do.
00:00:41.119 --> 00:00:44.320
Um, we talked to somebody who's actually from the HR business.
00:00:44.399 --> 00:00:48.000
And that's why I'm joined today by Philip Richter from Persone.
00:00:48.799 --> 00:00:52.960
Uh and you're a CPO, like uh that's a cop in another world, isn't it?
00:00:53.280 --> 00:00:54.560
Hey, thanks, thanks, Chris.
00:00:54.719 --> 00:00:55.679
Excited to be here.
00:00:55.920 --> 00:01:01.679
And uh yeah, uh being chief people officer in a tech company is uh definitely an interesting role right now.
00:01:02.000 --> 00:01:04.319
We also got Daniel Cranny uh, of course, with us.
00:01:04.400 --> 00:01:07.840
So we we both have questions for you about HR and about work.
00:01:08.079 --> 00:01:10.400
Well, first of all, what does what does Personio do?
00:01:10.480 --> 00:01:11.920
Like what services do you offer?
00:01:12.239 --> 00:01:13.120
Are you a platform?
00:01:13.280 --> 00:01:14.959
Are you uh should we know about you?
00:01:15.120 --> 00:01:18.560
I think I saw you in the news the other day that there seems to be some things going on.
00:01:18.879 --> 00:01:20.159
Yeah, you you definitely should.
00:01:20.239 --> 00:01:31.200
So Personio is the intelligent HR platform for small and medium-sized organizations, and we're like European headquartered and based, and we serve predominantly European companies.
00:01:31.280 --> 00:01:40.879
So if you're a company with anywhere between two and five thousand employees, you can do all your HR management operations through Personio.
00:01:41.040 --> 00:01:49.120
So from hiring people throughout managing them, paying them, performing uh performance management surveys, uh it's all in one platform.
00:01:49.359 --> 00:01:54.480
And you're based in Munich because you you you don't like paying low rent.
00:01:55.680 --> 00:01:56.239
Exactly.
00:01:56.400 --> 00:02:03.280
No, it's uh I I grew up I grew up around here and been around the block, but uh always find myself coming back.
00:02:03.439 --> 00:02:04.959
So this is this is home for me.
00:02:05.280 --> 00:02:08.719
So uh one of the biggest things, of course, is AI changing.
00:02:08.879 --> 00:02:14.240
I mean, we heard of it, uh, we we we're not gonna call it superintelligence, we're gonna call it AI and keep calling it AI.
00:02:14.319 --> 00:02:15.360
That's a good thing, I think.
00:02:15.840 --> 00:02:18.400
We chatted earlier about pre-preparation of the talk.
00:02:18.560 --> 00:02:26.960
You said that in currently you internally have a change to be an AI first company or to to put AI into your work as well.
00:02:27.199 --> 00:02:27.919
How's that going?
00:02:28.000 --> 00:02:30.240
And what are the main challenges that you're facing with that?
00:02:30.560 --> 00:02:32.960
We've definitely been on our own journey.
00:02:33.039 --> 00:02:41.840
So earlier in the year, we decided we're going to make uh Claude an anthropic enterprise available to everyone at the company.
00:02:42.159 --> 00:02:54.080
That was our first big push to say, okay, let's change the way we work, and accompanied that with a full week where we're all immersed in just trainings, building, figuring out the technology.
00:02:54.159 --> 00:03:06.159
And that was like our first big step that really led to the mass adopting AI and starting to build workflows, agents, and so on.
00:03:06.319 --> 00:03:18.240
And of course, as probably everyone else who's been on that journey, like you figure out, okay, once you are over the initial hump of getting people to adopt, like you run into subsequent problems that you have to solve.
00:03:18.400 --> 00:03:21.759
So, of course, you create some controlled chaos.
00:03:22.000 --> 00:03:23.759
Uh, people are building all sorts of stuff.
00:03:23.919 --> 00:03:28.080
At some point, you realize, okay, not all of it is like super productive and efficient.
00:03:28.159 --> 00:03:30.479
So you try to rein it back in.
00:03:30.560 --> 00:03:41.439
And I would say by now, like we've reached a state where we're investing a lot more in building the platform that runs all of the automations, the agents make it more scalable.
00:03:41.520 --> 00:03:44.960
We're also taking a slightly different approach, saying not everybody has to build everything.
00:03:45.120 --> 00:03:46.159
Everybody still has access.
00:03:46.400 --> 00:03:47.680
We want everybody to use it.
00:03:47.840 --> 00:03:52.240
But we also say there are like use cases that should be built centrally.
00:03:52.479 --> 00:04:07.919
And we're right now, over the last couple of months, like we've also taken a much more people-centric approach or organizational point of view, because there are so many studies out there that all will all quote tech is only 30% of the change.
00:04:08.080 --> 00:04:13.439
Uh, the other 70% has to do with how fast your people and your or can adopt.
00:04:13.599 --> 00:04:21.279
And so we're spending much more time now on the 70% figuring out how do roles change as a result, how does work
Load full transcript
change?
00:04:21.439 --> 00:04:25.199
How does like our or structure change as a result?
00:04:25.279 --> 00:04:28.639
And what do we need to do to support all of that?
00:04:29.600 --> 00:04:36.000
Well, oftentimes I find that there's AI solutions that are there to try to find a problem that doesn't exist.
00:04:36.319 --> 00:04:46.720
So, have you seen that like uh things that cost your company a lot of time, that basically were things that were frustrating for people, repetitive and boring, uh, they can be done with AI right now.
00:04:46.800 --> 00:04:49.120
Have you seen some of those use cases?
00:04:49.279 --> 00:04:54.319
And uh have you been doing it long enough to see if there's an efficiency increase in numbers?
00:04:54.639 --> 00:04:55.040
For sure.
00:04:55.120 --> 00:04:57.439
So there are, I got I totally get what you mean.
00:04:57.600 --> 00:04:59.199
I think we've all seen it.
00:04:59.360 --> 00:05:10.560
Just the power of AI and like letting you build things also leads to people being excited about building and building things that you realize afterwards were not that useful in the first place.
00:05:10.800 --> 00:05:15.199
But definitely, like we see a lot of use cases where it's just genuinely helpful.
00:05:15.360 --> 00:05:18.560
And I can also give examples from the HR world.
00:05:18.800 --> 00:05:23.600
There is a lot of manual operational work in HR, or there used to be.
00:05:23.680 --> 00:05:39.519
Like we're getting hundreds of tickets every month by employees who have like general questions, who need a reference letter, who need a certificate for something, and like we're in the process of automating most of that work, not just because it's more efficient, but also because it's much better employee experience.
00:05:39.600 --> 00:05:45.759
If you don't have to file a ticket and wait for things to be done, but uh rather to be able to do that live.
00:05:45.839 --> 00:05:55.920
And I think in the same way, there are use cases everywhere across the company where things are just getting either much more efficient or the results are a lot better.
00:05:56.160 --> 00:06:03.759
But I can definitely relate to the trap that you're citing of things getting built that don't really serve any purpose and just cost us money.
00:06:04.079 --> 00:06:11.360
Now, Dan, you build a lot of stuff for us lately using AI, and you basically went through your uh allotment of tokens in like an hour.
00:06:11.759 --> 00:06:12.399
Do you see that?
00:06:12.480 --> 00:06:16.800
Uh do you see that something where the company has to has to make it easier for you?
00:06:16.959 --> 00:06:18.399
Do you see that is a problem?
00:06:18.639 --> 00:06:22.079
Is like something like the token usage of of people.
00:06:22.240 --> 00:06:26.639
Is that a new uh feature that we need to track, much like we used to track days off?
00:06:27.120 --> 00:06:28.000
It's a really good point.
00:06:28.160 --> 00:06:32.240
I mean, it it is kind of a um it's a part of
a package now, I feel like, isn't it?
00:06:32.319 --> 00:06:35.439
Because I think Philip, you said then we made it available to all of the staff.
00:06:35.600 --> 00:06:46.160
And I think that like anything when when you know companies are trying to attract staff, uh the the the perks they can offer are going to be you know the things that go do go into our decision-making process.
00:06:46.240 --> 00:06:55.600
Um and so I I think that probably the the having a good allowance of tokens or at least um an understanding that you're not going to be like restricted in your role.
00:06:55.680 --> 00:07:01.199
Um and I think that that that of course puts the pressure on the company, right, to have the budget to allow for that.
00:07:01.360 --> 00:07:10.959
Um and I'd just be curious to know if like, you know, the the from from your side, Philip, you know, have the have the things that these that different companies offer to developers changed?
00:07:11.120 --> 00:07:14.800
Or um uh is that something you can get across in your platform?
00:07:14.959 --> 00:07:25.199
Or you know, is it something that is um more so about managing those kind of nitty-gritty, um not so enjoyable tasks, which is what a lot of us use AI for or use platforms for or use tools for?
00:07:25.360 --> 00:07:28.160
Um I'd just be curious, you know, how has that changed at all?
00:07:28.560 --> 00:07:28.879
Yeah.
00:07:29.040 --> 00:07:29.199
Yeah.
00:07:29.360 --> 00:07:36.079
So the to your to the point around like the the tokens, like it's obviously something that we also encountered ourselves.
00:07:36.240 --> 00:07:37.519
Like you give everyone access.
00:07:37.600 --> 00:07:42.879
At some point you run into the question, how much are you paying for tokens and who's using how much?
00:07:43.120 --> 00:07:57.839
And we started first building our own internal solution to track that, monitor that, allow people to like request top-ups in an easy way, and then realized like that's a problem that probably every other company has and that many of our customers have.
00:07:57.920 --> 00:08:06.240
So we've started building that into our product because also as like as the chief people officer, I'm not only interested anymore in like what am I paying for people.
00:08:06.319 --> 00:08:11.519
I'm interested in like what am I paying on top, not only in terms of benefits, but now also in tokens.
00:08:11.680 --> 00:08:15.519
And so we've built that into Personio as a product, the platform.
00:08:15.600 --> 00:08:22.959
That's where we track it now, and that's where we control it, and that's also where our people can track their own usage, see patterns, and so on.
00:08:23.120 --> 00:08:26.879
What we've learned is that yes, you need a nuanced approach.
00:08:27.040 --> 00:08:33.679
Like, of course, engineers have a much higher budget than the people team has, for example.
00:08:33.759 --> 00:08:45.679
But also in the people team, like we're taking an approach where I say there is a there is
a cap on your tokens per month, but also there is a super slim process to request more.
00:08:45.759 --> 00:08:49.039
And my stance pretty much is beyond a certain limit.
00:08:49.200 --> 00:08:52.320
I just want to know what you're building and what you're using the tokens for.
00:08:52.399 --> 00:09:00.399
But I'm pretty liberal in approving top-ups because I want people to use it to build and to get better results.
00:09:00.639 --> 00:09:11.279
But I also want to understand what are you using it for because we're getting a lot of learnings for things where four people in a given month will request all will request top-ups for pretty much the same thing.
00:09:11.440 --> 00:09:16.399
And so that's the insight that we then take and say, okay, here's something that we can build centrally and roll up back to all of you.
00:09:16.720 --> 00:09:23.360
Like we said, token maxing and even measuring lines of code that are written or these things that kind of happen as trends in the industry.
00:09:23.519 --> 00:09:28.159
Then I suppose you know that bubbles its way up to up to you as like can we please have this feature?
00:09:28.480 --> 00:09:36.159
This idea of us measuring lines of code or measuring token usage is sort of us uh grasping at scrolls as a way of measuring the productivity.
00:09:36.399 --> 00:09:44.399
So I I think it's just interesting to see how you you perhaps need to bring those features in, and then in a year's time you might find that actually they're not so important for teams.
00:09:44.720 --> 00:09:45.200
Yeah, for sure.
00:09:45.360 --> 00:09:56.639
Like we also we brought this uh feature in of the AI at work survey, where we want to understand like how people actually using AI and what's what's blocking them, what are they concerned about, and so on, to understand where you're at.
00:09:56.879 --> 00:10:09.759
What's interesting is when you start bringing that together with the usage data, because say six months ago everybody was talking about token maxing and how like token usage uh is or should be an indicator of performance, and like we all learned that didn't work out.
00:10:10.159 --> 00:10:13.519
Um but what we see much more now is that we like we treat it as a signal.
00:10:13.840 --> 00:10:29.519
Like if there is like really low usage somewhere where like outside in you'd expect high usage, you can put the survey against it and we treat it as a signal to go in and ask folks questions, and like you'll get a lot of sometimes surprising results, like also in the engineering space.
00:10:29.600 --> 00:10:37.200
Like we've had things like people wanting to use it but operating on a different part of a code base that was hosted elsewhere and therefore not connected.
00:10:37.360 --> 00:10:37.840
Okay, cool.
00:10:37.919 --> 00:10:39.200
That's something we can fix.
00:10:39.360 --> 00:10:51.039
Um, I had a lot of conversations with my team where people said, Hey, I want to use it, but I'm unsure what I can use it for, and
I'm concerned about uh sensitive data that I'm putting in, and I don't want to be liable.
00:10:51.360 --> 00:10:53.440
Cool, that's something that we can work on.
00:10:53.679 --> 00:11:11.679
And so for me right now, it's much more about understanding usage, blockers to it, and then what we can do, and also understand like where it really works and makes a difference, um, which I think is much more nuanced than the reality is much more nuanced than just the blanket approach, it's great.
00:11:11.759 --> 00:11:16.080
Or how many people sent their agent to fill the survey instead of doing it themselves?
00:11:17.600 --> 00:11:18.799
That's a good question.
00:11:18.960 --> 00:11:20.159
Uh I I don't know.
00:11:20.480 --> 00:11:21.840
Probably should check.
00:11:23.039 --> 00:11:26.240
Which brings us back to the the chief people officer thing.
00:11:26.399 --> 00:11:32.879
Like uh we live in a world right now where we can emulate a lot of humans with the with AI, and this is always a big headline.
00:11:33.120 --> 00:11:45.759
And to me, it's always a worrying thing because uh, I mean, it was like 78% of people fell for a new a new avatar that looked very human and is actually not a person at all.
00:11:46.000 --> 00:11:52.960
Um, and in the HR space, I mean, I'm just fringe HR, of course, but I'm a I'm a uh I'm a director here.
00:11:53.120 --> 00:11:54.799
So basically I get all these offers as well.
00:11:54.879 --> 00:12:03.360
Like, oh, we can have these training courses where avatars talk to your people instead of hiring uh the trainers to actually to actually train your people.
00:12:03.600 --> 00:12:07.600
And I feel like all of these things seem like automation to the wrong end.
00:12:07.759 --> 00:12:12.320
It feels to me like not human and therefore just almost insulting.
00:12:12.399 --> 00:12:19.039
If I were to get go to a job interview and you show me an avatar instead of coming to the meeting yourself, I'd be out.
00:12:19.200 --> 00:12:24.639
I'd I'd be not caring, but uh I mean I'm on a I'm in a good position where I actually can do that.
00:12:24.799 --> 00:12:33.679
But as a junior developer nowadays, getting through these systems and having automated HR systems with like avatar interviews, that seems to be on the rise.
00:12:34.000 --> 00:12:39.759
I'm definitely curious how that'll play out for uh a lot of companies that are starting this right now or using it.
00:12:39.840 --> 00:12:46.480
We we don't, and I very much agree with you that we the biggest question should always be like, what's what are we trying to achieve?
00:12:46.639 --> 00:12:47.840
Like, what's the goal here?
00:12:48.080 --> 00:12:58.320
And of course, as you look, if you look at recruiting like predominantly as a function of cost, then like it's tempting to say, okay, can we get an AI recruiter?
00:12:58.639 --> 00:13:01.919
And that'll make our hiring process cheaper.
00:13:02.320 --> 00:13:07.120
I and we very much look at our recruiting process as a
function of like quality of hire.
00:13:07.200 --> 00:13:13.200
Like, what's the process that will get us the best people for the role and the company?
00:13:13.440 --> 00:13:24.639
And if that's my function, an AI interview right now, at least that doesn't cut it because people don't like people want to talk to human in their interview process, they feel cheated.
00:13:25.039 --> 00:13:34.480
Um like you have similar similar debates in other areas uh around AI slop or uh fully AI generated content posted somewhere, people feel cheated.
00:13:35.120 --> 00:13:39.360
And I think that's like right now not the path that we will pursue.
00:13:39.600 --> 00:13:44.799
What we are doing, however, is like our recruiting processes are changing or our assessments are changing.
00:13:45.039 --> 00:14:06.000
So in the in the engineering space, like we're mostly gotten rid of the uh the previous uh code interview and replaced it with uh an AI-assisted coding interview, because like right now we want to know are people effective in producing high quality code and output together with AI rather than just by themselves.
00:14:06.159 --> 00:14:07.840
But that's it's still a human interview.
00:14:08.159 --> 00:14:09.279
You know, what is it you measure them on?
00:14:09.440 --> 00:14:12.320
The efficiency of their output relative to the tokens spent?
00:14:12.399 --> 00:14:22.159
Or I'm just curious about that because I think we've talked about how it is so hard now, isn't it, to measure in in the same way as it is for I was a teacher beforehand, very hard to measure um what students know.
00:14:22.240 --> 00:14:26.639
And things like essays are just were were overnight, you know, less less useful than ever before.
00:14:26.720 --> 00:14:33.200
So I'm just curious about kind of whether that's kind of impacted, you know, uh or what what it is you're measuring when you're doing these kind of interview questions.
00:14:33.519 --> 00:14:38.000
It's essentially live live interview where you co-work on a problem.
00:14:38.240 --> 00:14:47.759
Um interviewers of us, uh you get uh into the platform, you get a task that you're solving, and you have an agent that you can use to work on it.
00:14:48.000 --> 00:14:55.519
And like we're asking questions along the way and are observing just like how efficiently do you work with an agent?
00:14:55.600 --> 00:14:58.240
Like, do you know how to prompt it effectively?
00:14:58.399 --> 00:15:00.480
Like, do you know what the major pitfalls are?
00:15:00.639 --> 00:15:03.440
But then also do you understand what's happening?
00:15:03.600 --> 00:15:08.080
Like, can you of course correct like if you get additional information, do you know what to do next?
00:15:08.159 --> 00:15:11.679
And then like what does the uh what does the output look like?
00:15:11.840 --> 00:15:14.080
Like is that uh well structured?
00:15:14.320 --> 00:15:15.600
Is it functional?
00:15:15.919 --> 00:15:19.120
Does it uh conform to some of the principles that you were given?
00:15:19.679 --> 00:15:21.919
And it's like it's it's super interesting.
00:15:22.240 --> 00:15:35.679
The like the a couple of years ago, you know this better than I do, but the uh some companies you have to go through these things where you like come up with sort sorting algorithms, and it's just not what's relevant for us, at least anymore.
00:15:35.840 --> 00:15:40.080
But we want to know like with the stack that we have, are you effective?
00:15:40.159 --> 00:15:44.720
And it's slightly it's like a slightly different toolkit and mindset now than it used to be.
00:15:45.039 --> 00:15:47.840
It's it's fascinating to see how many of these go in the round.
00:15:47.919 --> 00:15:57.039
I mean, there was also this round of like where Google, especially, and other big um kabaker companies did all these clever questions, like why are manhole cover rounds?
00:15:57.200 --> 00:15:59.919
How many piano tuners are there in New York?
00:16:00.000 --> 00:16:01.360
And these kind of things.
00:16:01.759 --> 00:16:09.600
It seems like every year or so there is a new trend about how to do an interview and what questions to ask to trip people up.
00:16:09.840 --> 00:16:18.720
Um, have you seen people trying to do like prompt injection in like hidden text in in uh uh in like TVs and things like that already?
00:16:18.879 --> 00:16:20.639
Or is this more of an urban myth?
00:16:20.960 --> 00:16:22.720
Um not yet.
00:16:22.879 --> 00:16:27.039
Um, at least not at scale and it hasn't been an issue for us.
00:16:27.200 --> 00:16:39.840
But what we definitely like what you can definitely see is just like volumes going up, like things that used to be a thing a decade ago, like how your CV is structured and so on, like what it looks like visually, like that's all gone away.
00:16:39.919 --> 00:16:42.720
But like that shouldn't be shouldn't be a secret to folks.
00:16:42.960 --> 00:16:45.919
What is your advice for somebody who's looking for a job right now?
00:16:46.080 --> 00:16:56.639
Is it to have an MCP server and some markdown file somewhere with your CV on it, or is it just having a nicely designed PDF that you print out and fax to somebody if it's in Germany?
00:16:58.240 --> 00:17:00.480
I mean, faxing wouldn't work with us.
00:17:00.799 --> 00:17:04.480
I guess know what company you're applying to and what what they value.
00:17:04.720 --> 00:17:16.079
And like for us, that depends on like the the experience level that you apply to, but we want to know like what what's your track record of like producing great things, building great things.
00:17:16.319 --> 00:17:21.279
And like if you're early in your career, there are a lot of different ways to showcase that.
00:17:21.440 --> 00:17:34.559
And it can be that you have uh a pretty impressive uh public repo somewhere that we can see, it can be just references of like previous work engagements, internships, whatsoever.
00:17:34.640 --> 00:17:44.400
And then like we want to see live just can you
solve problems that we're throwing at you, which we found is a pretty good predictor for for our interview process.
00:17:44.960 --> 00:17:58.720
And so many different things can work, but I also I mean, if you are people are in very the companies are in very different parts of the adoption curve right now, so depending on where you go, that might look might look different.
00:17:58.799 --> 00:18:00.640
But that that would be my advice.
00:18:00.720 --> 00:18:11.839
But in any case, I think it's moving towards it's moving away from formal credentials more towards how can you showcase what you've built and what you're building and how good you're at that.
00:18:12.240 --> 00:18:25.039
Do you do you see this being sort of uh something that's increasingly used in teams whereby you know where they may not have a kind of people department or an HR department that's extensive or even you know particularly well kind of established?
00:18:25.119 --> 00:18:36.160
It might be is your software sort of designed to almost help teams where they may not have that HR team to share the load and non-HR people to kind of take on those responsibilities?
00:18:36.319 --> 00:18:45.759
Um, and if so, do you see that being the kind of direction of travel that perhaps HR becomes intertwined into other areas of the business rather than a standalone thing?
00:18:45.920 --> 00:18:50.559
Because the software allows us to perhaps kind of take on some of those um tasks.
00:18:50.720 --> 00:18:51.440
What do you think?
00:18:51.680 --> 00:18:53.519
Yeah, yeah, really interesting question.
00:18:53.680 --> 00:18:58.559
So to the first part of it, yes, like that's exactly part of our mission and what I believe in.
00:18:58.720 --> 00:19:03.759
Like I'm in a lucky position of leading an HR team in an HR tech company.
00:19:03.920 --> 00:19:12.319
So I'm afforded the luxury of being able to have a couple of roles and just great talent that I wouldn't be able to have in a different kind of company.
00:19:12.559 --> 00:19:44.480
And what we're doing is like what I'm what I very much believe in is that the things that we're building for our own HR team for the company that work and the approach that we're taking is to package up the things that work and like start integrating it into the product so that we can almost give these things as a service to companies who wouldn't be able to hire or afford a people data scientist working on their team, working on questions like which part of our assessment process actually move the needle on quality of hire.
00:19:44.960 --> 00:20:00.799
But if I can bring these things into the product as a standard feature of Personio, I'm allowing, like I'm giving the power to other HR teams to like work in the same rigorous scientific and like evidence-based way.
00:20:01.200 --> 00:20:13.359
And that's definitely a journey that and a trend that I'm excited about because I think it can only lead to better outcomes for people and companies, like if we can if we can strengthen that.
00:20:15.279 --> 00:20:17.119
I want to talk about bias a bit.
00:20:17.200 --> 00:20:30.559
Uh, I think there is a big issue that in the early days of AI, we had like models that were trained on a certain environment, and uh certain people only fell through the fell through the track the cracks when they actually tried to apply for a job.
00:20:30.880 --> 00:20:32.240
And uh I had that as well.
00:20:32.400 --> 00:20:39.279
I mean, I had a job offer from a company, they asked me to apply for the job, I applied for the job and immediately got rejected because I don't have a degree.
00:20:39.440 --> 00:20:44.720
So then I emailed them again and I'm like, so thanks for that, uh putting me that into that kurde sack.
00:20:44.799 --> 00:20:48.720
And he's like, no, uh sorry, we turned that off again, and now I can try applying again.
00:20:48.880 --> 00:21:00.480
But do you think there is a uh there is a problem that people rely on technology without looking into the uh social impact and the also the cultural impact for European companies?
00:21:00.960 --> 00:21:02.720
Yeah, super interesting question.
00:21:02.880 --> 00:21:11.039
I mean, there are a lot of like there are some ongoing lawsuits uh right now, some pretty prominent ones around AI in recruiting and assessment.
00:21:11.119 --> 00:21:14.720
And I think it'll be interesting to see where all of these lands.
00:21:15.039 --> 00:21:27.200
What I see is the I think it's all a function of like how like effectively you can gear automation or also agents towards the right outcomes.
00:21:27.279 --> 00:21:34.160
And I think the foundation is needs to be like you need to have a good understanding of your biases in the first place and how you measure and track it.
00:21:34.319 --> 00:21:41.119
And that's not necessarily been the dominant conversation in the HR space in the past uh 10 years.
00:21:41.440 --> 00:21:54.000
And so from my point of view, uh it's a really positive thing that this conversation is surfacing again because uh it's not the like the reality is not that recruiters or hiring managers are unbiased.
00:21:54.079 --> 00:21:59.599
Like we all have our biases, and I guess an HR team, you always do your absolute best to try.
00:22:00.240 --> 00:22:09.599
reduce them but a human going through 200 CBs like will fall into their own biases and it's like very measurable.
00:22:09.839 --> 00:22:24.000
And so what I'll be interesting I'll be interesting to see is what how can we use agents and technology to like yes help us with some of these assessments but actually get to a less biased outcome.
00:22:24.160 --> 00:22:27.440
And like right now the conversation we have is like what bias do you want to have?
00:22:27.599 --> 00:22:39.039
Do you want the bias of an LLM that's trained on biased data and so will reproduce biased outcomes or do you want the bias of a recruiter that's just very hard to measure and quantify?
00:22:39.200 --> 00:23:03.200
And I believe
we're probably going to end up in a middle scenario where we all have to learn a bit more about these biases, optimize the the optimization functions of these algorithms and put in place just more rigorous um evaluation both for humans doing work that's prone to biases, but also for any sort of technology.
00:23:03.359 --> 00:23:10.160
And then I think we'll get to a better place than where we were five years ago and also where we are today.
00:23:10.480 --> 00:23:25.599
I think so often we talk about AI first, you know, as the um and we've heard this term used a lot now haven't we like an AI first company and and for a while it seemed like there was a kind of the hope was that you know everything can be completely automated by AI or as much as possible.
00:23:25.759 --> 00:23:36.000
And that now there feels to be there's this slight rolling back where um as you said Philip you know we kind of we we put the features in place and then we figure out the bugs or the problems with them a few you know months later or whatever.
00:23:36.480 --> 00:23:49.200
But it feels like something you just said there about you know the the meeting in the middle is kind of what I'm hearing a lot from developers about their hope for whether it be the state of uh you know development and actually do we write all code by hand?
00:23:49.359 --> 00:23:49.599
No.
00:23:49.839 --> 00:23:52.400
Do we generate everything by AI and not look at it?
00:23:52.559 --> 00:23:52.720
No.
00:23:52.880 --> 00:24:24.559
So the truth is sort of somewhere in the middle and I suppose it'll be interesting just to sort of maybe see from your side are there any horror stories of things that have happened when maybe people have put trust too much in AI, where maybe they have kind of tried to outsource the entire hiring process um at least you know these are ones maybe you've come across is there anything are there any kind of lessons you can take from that or that you maybe have learned from it about putting too much trust in it can result in very expensive mistakes, whether it be token wastage or hiring somebody who's just not fit for the role.
00:24:24.880 --> 00:24:33.440
I'd say no no horror stories at Presonian like we've certainly made our learnings and we've uh certainly wasted tokens and so on but no no horror stories.
00:24:33.680 --> 00:24:47.359
But the I think wherever like and you see this in the industry you see it everywhere Chris you mentioned like the the interviews with AI avatars that then backfire like we've seen companies lay off entire teams only to rehire them a year later.
00:24:47.440 --> 00:25:07.279
So there are the horror stories and I think they often come from situations where companies or people have sometimes a justified belief that AI will like take over a certain job to be done and fully commit to it before having tested it or knowing that it actually works.
00:25:07.519 --> 00:25:27.519
And that's that's always dangerous and that can work but that can also just backfire massively and so what I and we always try to do is say okay let's let's say there is a there is a workflow there is a job to be done there is a task where we say based on everything we know it should be perfect for AI.
00:25:27.759 --> 00:25:52.960
Okay let's let's actually build it let's try it let's run it in parallel evaluate it let's run it on a like test it on a subset do a proper a b test and one we see once we see it works then commit to it and say okay let's take out the manual process let's assign the people to other work and so on but I think the the pre-committing to an unknown is a bit a bit tricky.
00:25:53.599 --> 00:26:13.519
One thing that we always get promised by uh by AI is to uh to democratize the work to make sure that anybody can do the job but that we always used to be a developer used to call a developer for and we were the stalwarts we were the people that people said like okay we need to build something and we're like okay our our backlog is full we don't have any time to do that.
00:26:13.680 --> 00:26:41.599
So there is a whole feature of people of companies doing like white coding of little helper apps inside the company itself we see it in our company where people in the sales team and people in the uh in the partner team are actually building little tools and then using them and they just see the light of day outside the company but it empowers people to do a task that easier than they didn't have to before and they didn't have a chance to actually ask the developer because they were full up on it.
00:26:41.759 --> 00:27:24.079
So do you see this as a uh as a as a powerful thing for AI to do to allow people to build throwaway apps to for for tasks that they wouldn't be bothering the the tech department for we're exploring something right now inside the product uh which we're calling the AI Studio something that we've built for us in testing for Personew only so far but what we hope to bring to customers pretty soon and the the insight there was Personew is a SaaS platform it's pretty comprehensive by now it's pretty customizable and pretty flexible yet every company has their quirks and has their processes that you would never be able to put into a platform like that.
00:27:24.480 --> 00:27:34.559
And yet with AI I should in theory be pretty uh straightforward now to say like I need a slightly different interface on the same existing data model.
00:27:34.640 --> 00:27:58.000
I want slightly different business logic on top of it can we not put this into a custom interface, a custom UI for you to use that works perfectly for your own company and so we're exploring that right now it's um it's like super interesting because there are definitely like our team came up with like 20 use cases I think within a week where we all say this is something that we'd absolutely love to have inside the product.
00:27:58.240 --> 00:28:09.279
But you also run into like limitations and so on and you realize there are
situations where it just doesn't make sense or where it would break in the future or where we'd have to maintain things.
00:28:09.599 --> 00:28:13.440
But I definitely believe this is going to like it's a huge huge help.
00:28:13.599 --> 00:28:17.200
But again the the important thing is to be deliberate about it.
00:28:17.359 --> 00:28:52.640
And it's also the same when we do it internally like we're we're we're starting to get very structured about like what are the things that we're very happy for people to build for themselves like personal productivity stuff things to build within the team that they shouldn't just build themselves maintain themselves but if the person leaves the company it's like fine we can continue what are things that are like ongoing processes where something is like built into our operating model like that needs to transfer over to central stack there needs to be proper maintenance on it there needs to be security checks on it and so on.
00:28:53.119 --> 00:29:00.720
And then there are things that I I don't want anyone to be able to work on or vibe code.
00:29:00.799 --> 00:29:10.799
Like I don't want anyone to be able to say cool let me vibe code a solution that screens CVs for me and I just rely on the results.
00:29:10.960 --> 00:29:30.960
Like there are a couple of things that we also just outright block where we still encourage folks if they have good ideas to approach us and the central team that builds these things for an assessment and then we can uh we can think about if and how we build this in a safe and compliant and sensible way.
00:29:31.200 --> 00:29:42.079
And I think that's where what I see where a lot of companies are landing also I do think the impact is real the potential is real but you have to be like we have to guide it into the right direction.
00:29:42.559 --> 00:29:49.599
We're currently going through an assessment phase of like what people are using inside the company and with AI and models and tools and things.
00:29:49.759 --> 00:29:57.599
And I mean I had meetings where people invited like random note takers that they that they thought thought were cool and I basically always blocked these things immediately.
00:29:57.839 --> 00:30:13.839
But uh that is a thing that kind of happened do you think the uh do you see that there's a lot of companies that have been falling for like a siren song of a free tool that they put into their company and made uh and didn't look at the security implications or the privacy implications?
00:30:14.000 --> 00:30:16.880
Is this something that you encountered yourself or from your customers?
00:30:17.200 --> 00:30:40.160
Yeah we we definitely see it we see like on the extremes of the spectrum we see two things right now one is the the companies that are just very concerned and by default block out pretty much all of it and like don't allow their people access which we believe is a shame and often also doesn't land well with all the employees because they want to
use it.
00:30:40.319 --> 00:30:54.960
They see the potential and then there is the other extreme which is free for all most of the time there is a policy in place but if it's not technically enforced people will just sign up for other software download something whatsoever and use it.
00:30:55.119 --> 00:31:01.200
And we see that in the results from the AI at work survey like we're producing benchmarks also on top of it.
00:31:01.279 --> 00:31:07.119
There is a question around the the tools the I tools that you're using whether they were company provided or ones that you found yourself.
00:31:07.200 --> 00:31:14.880
And we definitely see that there is a good chunk of people that use tools that they found themselves so that are not company sanctioned.
00:31:15.119 --> 00:31:32.960
Now that doesn't necessarily have to mean that that is actually an issue like someone asking an unsanctioned AI tool like a generic question about their role or use it for sparring on a technical problem can be totally fine.
00:31:33.039 --> 00:31:36.079
Like there's no sensitive data transmitted whatsoever.
00:31:36.400 --> 00:31:38.480
But the issue that we see is like you don't know.
00:31:38.799 --> 00:31:48.000
And I think that's where like we see a lot of companies now also thinking about based on these insights also what do they have to technically block?
00:31:48.480 --> 00:31:49.920
What do they allow?
00:31:50.319 --> 00:31:51.759
Where can they apply policies?
00:31:51.920 --> 00:31:56.720
But that's where we aim to shed some light for for companies so that they know where they stand.
00:31:57.039 --> 00:32:57.839
I know you sort of said you use it internally and I mean that was the big question that I had in my mind it was like do you kind of dog food you know do you do you uh use your own product basically and I think um but as as you've answered that I suppose the next part was where does the sort of feedback come from then you know is it sort of an internal thing where you know by you using your own product you come up with these ideas and then try them you know release them to a small uh you know bet a beta group or is it uh something that that bubbles up from um you know just conversations with with customers is that like a process that you go through in order to kind of do that every every quarter what's your what's your kind of process in terms of that feedback loop comes from comes from different angles at the same time so there's the the core product development loop which is very customer centric we have um a customer advisory board we have focus groups that we co-develop with who use like early versions of the features who will also help us in the in the design and give feedback and like we also get a ton of input and asks and feedback from our customers like every other company.
00:32:57.920 --> 00:33:15.279
So there is there is this like starting point and then what I always want to do and what I always ask for is that we get the alphas already into our own internal account so that we can test it with our own people team um like give any additional feedback learnings and so on.
00:33:15.519 --> 00:33:30.960
And then what's what's growing now is things actually starting in the Personio people team and there are just things about HR work that are just hard to grasp from the outside like you have to be in the middle of it.
00:33:31.039 --> 00:33:42.319
And so like we constantly ask ourselves what are the things the core parts of HR work that we're doing outside of Personio and why is that like shouldn't we be able to do that inside the system?
00:33:42.559 --> 00:34:29.519
And so there's very close collaboration with like all the different product teams like every product team that we have um has like a an embedded member of the Personio people team virtually in their team to like provide insights feedback test early versions and like that's something that I'm excited about that we're increasingly able to push learnings from Personio things that we built in the HR team or for the HR team into the product test it with other uh customers and then roll it out one thing that a lot of companies do is now in uh uh introduce people to a new AI agent coworker out of a sudden every Slack channel has a random name in it with a random avatar picture that's supposed to be a very normally a very diverse person that has nothing to do with your company.
00:34:30.239 --> 00:34:31.920
Do you see that working?
00:34:32.079 --> 00:34:40.400
Do you see people embracing these um let's say synthetic coworkers or is this something that feels like enforced?
00:34:40.800 --> 00:34:46.400
I think people still struggle with the the picture of a synthetic coworker.
00:34:46.639 --> 00:34:57.360
Like I I see like that bringing up a lot of questions and pushback and personally also get why like I haven't seen a truly convincing synthetic coworker yet.
00:34:57.679 --> 00:35:33.599
What people do embrace is the idea of just like having agents whatever you want to call it that serve a specific purpose in the places where you already work like people are getting very much used to meetings being transcribed and summaries being sent around and appreciating that people are getting used to being able to access Claude in Slack to pull in data or summarize a thread and that's very much happening is these like purpose bound agents being included in the work.
00:35:33.840 --> 00:35:52.320
And so like we're already also experimenting on that path and including agents into the work chart in Personio um including supervisors and um functions to measure spend so their salary and uh and their effectiveness.
00:35:53.840 --> 00:35:57.039
This is early but like we see we see that it's happening.
00:35:57.360 --> 00:36:03.119
We have these agents now at Presonio which are like I wouldn't call them synthetic coworkers.
00:36:03.199 --> 00:36:10.239
I think that like for me that like draws up the wrong image of what this technology is and
does.
00:36:10.719 --> 00:36:20.320
But the fact remains that there is technology now that does work that is not deterministic and that costs real money depending on the model that you put behind it.
00:36:20.480 --> 00:36:28.960
And so we're experimenting with reflecting that also in Personio in the org chart and make sure Skip gets properly managed also.
00:36:29.679 --> 00:36:59.760
It does feel like something that we always used to have I mean we had like Clippy and then we had like the the uh i care bot in like 1998 and there was always this like cutesy little thing that helps us on the side and I always found this weird because as a professional person I don't trust it when it's too cutesy or when it's trying to be too convincing like the amount of companies that have like I mean I see with Amazon for example the Amazon hub has this Rufus bot I never clicked that thing on purpose.
00:36:59.840 --> 00:37:24.960
I always thought it was a search and actually out of a sudden I had a conversation with a thing that I didn't want to have a conversation with do you see that a lot of people are doing that like there's a lot that there seems to be like a lot of about pages where it's like you can now chat with our bot from Steven from accounting uh which is then a bot that do you see that something that customers want to have as a feature and how do we tell them that it doesn't work or is this my bias?
00:37:25.599 --> 00:37:34.400
Yeah I I think there we have to like I I try to separate the the use cases or what this actually does.
00:37:34.559 --> 00:38:30.159
So the where it's actually helpful is say like the example I mentioned earlier um the the HR service desk like you need something from HR like your insurance has changed you want to go on paternity leave uh you need a certificate for your visa whatsoever like in most companies this either means like if it's a small company you slack someone or pop by someone's desk if it's a larger company you file a ticket often like repeating all the information the company already has that is not a workflow where you want to speak to a human all you're interested is like getting this thing done that you need and like getting it done quickly and being able to provide you an interface right where you work already that like helps you get this done or gets it done for you I think is amazing and it's like a much deaf much better employee experience than what we had before.
00:38:30.320 --> 00:39:30.639
So say you need this certificate you work in Slack you type help I need certificate for my visa you get two follow-up questions the chat ends with awesome document is created it's with a UN for review it'll be deposited in your personal profile within the next hour awesome like that's cool if you have a like serious issue with your manager if I sent you to a bot right now I don't think that would land well like you want to talk to someone who has like more context than fits in a one million uh token context window and have a human-to-human confidential conversation and so I think separating these things is important and then we can actually like use it where I believe it's not only more efficient but will make for a much better employee in the human experience but also separate from how jobs used to be cut.
00:39:30.800 --> 00:39:39.280
Like the the the goal for me is not to replace my people operations person with a synthetic version of the same person.
00:39:39.440 --> 00:39:59.199
My goal is to take some tasks that used to sit with people operations which is not be a task anymore in the future which we can do in a different way but it make more space and retain the things that are truly needed from a human in this role right now and which I also don't see going away.
00:39:59.519 --> 00:40:11.760
What are tasks that you see is a high redundancy that actually are not necessary for humans to do that we still waste our time on what could be really automated that basically we do by hand now.
00:40:12.320 --> 00:40:43.920
Yeah so there is a lot of use cases around the the general question use case and whether that's in customer support whether that's in HR whether that's to finance like I just want to know something and I don't know that this is actually documented somewhere or I wouldn't know where to look or I don't I can't be bothered to search something you just ask your question somewhere like you get an answer informed like on a knowledge base or or some rag search awesome that seems to work pretty well.
00:40:44.800 --> 00:41:00.639
Then there is this entire um like documents related use case repository where it's around like creating documents, comparing documents, looking through documents, like reviewing them whatsoever that seems to work really well.
00:41:01.519 --> 00:41:14.800
And then there's everything around processes that are pretty much standardized but where like the number of unique cases exceeds what you could just do with deterministic automations.
00:41:14.960 --> 00:41:26.079
And right now you see these cases having a lot of humans in the loop still to check the final results but where an AI is perfectly capable of like giving you a perfect first draft.
00:41:26.480 --> 00:41:38.800
So say this uh you need a certificate for like a specific type of visa like that's not something that I like where we would call it a deterministic automation for because it happens like once a year.
00:41:39.280 --> 00:41:51.119
But uh with the right model AI is perfectly able to research the requirements get the data from the system put together a draft and then I let someone check it who's actually uh specialized in this in this kind of stuff.
00:41:51.360 --> 00:41:55.039
I think all of that works um works really well.
00:41:56.079 --> 00:42:13.440
How well it works in coding you're you're the better you're the better experts and I think what we'll what we'll yet have to see is the like the truly taking over for this more fuzzy strategic
long-term type of knowledge work.
00:42:13.679 --> 00:42:26.639
I think there it's it's a really helpful sparring partner really helpful for analysis but uh I think like letting that be done by some sort of agents um I don't see right now.
00:42:29.039 --> 00:43:04.559
In terms of uh I mean we're about to wrap up because we had like almost uh almost 45 minutes and people are aren't not having any uh uh any like attention span anymore um in terms of things that you find uh companies struggling with in uh embracing HR and AI uh tooling right now what is your advice to actually say like uh uh um what's the pace of of doing that do companies feel left behind force their people to use AI that doesn't work at the same time they block AI from it where do you find the middle ground and what what worked for you?
00:43:04.880 --> 00:43:25.360
Yeah yeah it's so it's so interesting it's exactly as you say like do you you have this like weird mix of people having like an exaggerated belief in what AI is already able to do today and at the same time uh we see many who are scared just because they don't know better on actually actually using it.
00:43:25.519 --> 00:43:27.119
It's uh it's a very interesting mix.
00:43:27.280 --> 00:43:33.760
So what I always say to everyone I think it is really important to get moving.
00:43:33.840 --> 00:43:56.320
Like if you haven't started today like you need to get started you need to understand the technology you need to start to understand yourself what the type of problems are that you will run into once you start using it, what's working, what's not working and you need to start to let your people work with it in whatever way so that they can also develop and understand it.
00:43:56.480 --> 00:44:07.119
And it's the the riskiest thing to do right now is to just say Do nothing because that'll be very hard to catch up in a year from now and two years from now if you don't don't get started now.
00:44:08.079 --> 00:44:15.840
What we also see is like companies have different approaches that like fit their specific situation, their industry, whatsoever.
00:44:16.000 --> 00:44:25.280
Like we did the way we took is like give the technology to everyone first within some constraints, then see what breaks and learn and iterate from there.
00:44:26.000 --> 00:44:31.360
What I see with other c with customers of ours is like also approaches where you are much more targeted.
00:44:31.440 --> 00:44:35.760
You take the use case way, like you do an assessment, say what are the three things where AI could help the most.
00:44:35.920 --> 00:44:38.639
You build those, you try it, and then you expand.
00:44:39.679 --> 00:44:56.400
I I think it doesn't matter so much, like as long as you get started and then just like continuously monitor where you are, with the type of problems you have, how your people respond to it, what they tell you, and just constantly react and get better to that.
00:44:56.559 --> 00:45:02.719
I think there is not so much that um that you can do that can go wrong.
00:45:02.800 --> 00:45:05.440
Like the biggest mistake is to stand still.
00:45:05.679 --> 00:45:21.679
And then on the other hand, of course, if you're if you're in the camp where you're a bit overly bullish on what it can do right now, and like you risk um leaving your people behind or not believing in what you're saying anymore, like that's the other extreme, but the same applies.
00:45:21.840 --> 00:45:34.480
Like, listen, like critically evaluate how it's going, what the value is of those investments, and then ask ask your people, and like they will tell you we see this ourselves.
00:45:34.639 --> 00:45:45.440
There are parts of the company where we have rather have the problem that we're not moving fast enough and we're not enabling and allowing people to use AI as much as they would want to.
00:45:45.599 --> 00:45:58.079
And we have parts of the company where we have people telling us that they feel that what we're expecting, that the impact that we're expecting AI to have is not what they see on the ground.
00:45:58.239 --> 00:46:01.039
And that is like completely normal.
00:46:01.280 --> 00:46:06.559
But I think as long as you have an open conversation about that and say, hey, looks like uh something's off.
00:46:06.639 --> 00:46:08.400
Like what's the what's the reason here?
00:46:08.559 --> 00:46:10.239
I think that cuts both ways.
00:46:10.639 --> 00:46:11.440
Well, thanks very much.
00:46:11.519 --> 00:46:17.440
This was Philippe Ricta in one of the first work uh focused uh coffee with developers, and uh you fed well.
00:46:17.519 --> 00:46:26.480
You're you're half developer, half HR person, it seems, and it's great to see how companies like you are embracing this technology and putting it into the system as well.
00:46:26.719 --> 00:46:33.840
I love the whole concept of democratizing the software development process because software is running our world.
00:46:33.920 --> 00:46:35.440
It's you cannot avoid it any longer.
00:46:35.519 --> 00:46:42.480
So it actually makes sense that more people are involved in it rather than just like having to call the expert and hope for the best.
00:46:42.719 --> 00:46:48.719
So with that in uh with that in mind, uh go to HR people, talk to HR people.
00:46:49.039 --> 00:46:55.360
I think uh the main thing we can do about not being redundant in the AI world is be more human.
00:46:55.599 --> 00:47:00.480
And uh HR people are the ones that actually are professionally human, so that's a good thing to have.
00:47:00.800 --> 00:47:02.559
Couldn't couldn't have said it any better.
00:47:02.719 --> 00:47:04.000
Thanks so much for having me.
00:47:04.079 --> 00:47:06.400
This was a super interesting conversation.
00:47:06.559 --> 00:47:18.159
And yeah, if anyone out there is interested in like applying engineering to a company where the product is actually the one that's influencing how your work is shaped up, uh, hit us up.
00:47:18.239 --> 00:47:20.559
Like we're always looking for great people to join us.
00:47:20.880 --> 00:47:21.440
Thank you very much.
00:47:21.599 --> 00:47:25.280
This was Philip Richter from Posonio, and this was Coffee with Developers.
00:47:25.440 --> 00:47:26.639
See you next time.