Sarah Mathews on why AI governance fails without a shared language (lessons from the Adecco group)

Most companies treat AI governance as a compliance exercise. But if your teams can't even agree on what counts as AI, your governance framework is built on sand. 

In this episode, I'm joined by Sarah Mathews - Global Responsible AI Manager at the Adecco Group - who leads AI governance, ethics, and literacy across 60 countries and millions of hiring decisions. 

We discuss: 

• Why AI governance must start with a shared definition, not an inventory of tools
• How Adecco built role-based AI literacy training that reached 44,000 employees voluntarily
• Why distributed governance ownership beats putting one team in charge
• The hidden danger of making AI too human-like and what it means for trust 

Connect with Sarah: 

LinkedIn: https://www.linkedin.com/in/sarah-mathews-87b481122/ 

Connect with Valeriya: 

LinkedIn: https://www.linkedin.com/in/valeriya-pilkevich 

YouTube: https://www.youtube.com/@aimadesimpletalks 

Podcast: https://aimadesimple.buzzsprout.com

Need help building AI capability in your organization? Book a call. 

Transcript

SPEAKER_00 0:00

Most companies think AI governance starts with mapping their AI systems. But what if the real first step is something much more basic, like getting everyone to agree on what AI actually means? Welcome to AI Made Simple, the Transformation Series. I'm Valeria Pikevich, and I talk to global leaders, innovators, and practitioners for shaping the future of work in the HFAI. In this episode, I'm joined by Sarah Matthews, Global Responsible AI manager at the ADECA Group, where she leads AI governance, ethics, and literacy across 60 countries. She's also studying at the University of Cambridge and is a thought-after speaker on AI governance and workforce transformation. We talk about why a shared AI glossary matters more than most companies realize, how a DECA-built role-based AI training that 44,000 employees took voluntarily, why no single team should own AI governance and how to distribute it across departments. And why making AI feel too human might actually be a governance risk? Sarah, it's great to have you on this show.

SPEAKER_01 1:01

Thank you so much. It's a pleasure to be with you today.

SPEAKER_00 1:04

Sarah, you started your career as a recruitment consultant and now you lead responsible AI governance for the entire Adeca group. What was the moment or experience that made you think I need to move from the people side into the AI side?

SPEAKER_01 1:18

That's a very good question. And I think for me, both are still intertwined. So it sounds a bit like it's uh from the people to the tech side, but for me, AI is a lot to center on people. So actually, it's not moving away from One Piece, but just centering technology around people in the end. And I have to say I had quite an uh actually, yeah, not uh the nicest transition factor, let's say like that. I was always very passionate about the recruitment business, the HR business, and I still am. That's why I'm still with a Deco. But we had a phase when I was a recruitment consultant and team lead where we had to let go really a larger proportion of our workforce. Just because the economic situation wasn't good. Those people were basically positioned in a project that had to end because the site was shut down. And we had over 70 people searching for a new role, and we were searching for a new role for them, but basically we couldn't manage for a large proportion of them. So we had to give them notice and let them go. And for me, doing that not only with one person, that's hard enough, but doing that with more people and especially also people who are more advanced in their career was really a point that shocked me for a while. And I was thinking, how can you change that? Because it was purely human, right? Screening through job descriptions, doing sales calls with potential clients, finding new job opportunities for these people was really a purely human task that a small team of three people couldn't manage to do. So that was for me point where I thought, okay, we need to improve here somehow. Coincidentally, at the same time, I was looking into how we can change the way how we're recruiting in the part of the business that we were responsible for, and what are new trends that we need to look into. There, I stumbled across a field that was growing in popularity, which was data science. So I started digging into what is this actually, right? What does it mean to be a data scientist? What is this field about? And I got so curious because I started to understand actually that the art of data science obviously is to use statistics and mathematics to make sense out of, for example, large volumes of data. So I started digging deeper and deeper into that and started to educate myself a lot and thought, okay, that's the solution to a certain extent, right? And that's where I actually had the pleasure to connect with many of our experts around the globe, because obviously I was seeking their advice, I was seeking for more knowledge. And I came across our global data science team and uh got the chance actually to join the team and to try actually to bring the technological side and really the human side, so the recruiters, closer together in being a so-called engagement manager. So you could say the translating function between the technical people and the business. So for me, well, all of that is still intertwined. It was a journey that I just stumbled into out of a not so nice situation, but I'm very thankful that I did.

SPEAKER_00 5:02

Adeca Group places millions of people into work across 60 countries. And when you make uh from your position governance decision about how AI can be used ethically in hiring or assessment, it does not just affect your company, it affects the whole labor market. How do you think about that responsibility? And what are your guiding principles of using AI responsibly in?

SPEAKER_01 5:25

Yes, I mean it sounds huge if you're spelling it out. I think we need to be clear about that, right? Um, but I am, and we are as a group well aware of that responsibility. And this is really for us also something where we could say it gives us the drive to continuously learn and improve. Because obviously, also here, speaking about implications of 60 countries with very different populations, with different cultures, thinking about things like global mobility, right, is actually something that one person can't do alone, right? I can't understand all cultures, all countries by heart. I can't understand all dynamics of labor markets solely on my own. So for us, this responsibility comes with the responsibility actually to really reach out to our countries to actually understand the bigger picture that we're navigating in and to build the right networks that we can tap into to basically first understand, learn, adapt, and actually also just incorporate other perspectives and Ellies. We can't do that alone. So you will see us also doing a lot of lobbying with, for example, the World Employment Confederation, with other big organizations and parties, because obviously we can have a strong voice in there because of our size of the company and because of the work that we're doing. But ultimately, it really is about connecting to the right people to basically roll that out in an ethical and responsible way in the end, given the fact that it is obviously influencing the labor market.

SPEAKER_00 7:19

Inside the company, you launched the training about responsible AI in action. I believe many companies right now are talking about responsible AI or responsible AI being important, but not every company is moving into implementing those principles. So, how um how did you make this move from having them on page to governance that actually changes how teams build and buy AI?

SPEAKER_01 7:41

That's a very, very important question. And also one where I would say there is not a single truth behind it. But our training that we have built is basically really based on different building blocks that are intersecting with each other. So what we did first was basically obviously pushing out the principles, right? So pushing out the principles to make people aware and to make people understand. But that obviously is oftentimes very abstract. So what does it mean to be ethical? Might have a baseline that everybody understands, but how can I myself apply that? If I'm a salesperson, if I'm a recruiter, if I'm working in finance, is something that we need to think through. The second step that we really did, on the one hand side, because obviously we want to increase also the overall AI governance literacy, or AI literacy overall, not only for governance, is that we rolled out a training across countries that is role-specific for our dominant groups in the workforce. Basically, one of our largest employee groups is obviously recruiters. So we were thinking through and talking through with different recruiters in the business, where do you actually apply AI? And the next step was then to think, okay, if we're looking into our principles, how do we apply those principles on what people are actually doing? To give you an example, writing a job description with the help of a large language model. What do I need to do there? First, I need to check the output. I think that's clear, but people need to be educated and reminded. And second, I can prompt it in a way that targets neutral communication. So I actually need to teach people or I need to make people aware what does actually neutral language mean and how does this look like. So we pushed that out as a training, first on a voluntary basis. And it really was a great success that we did not really expect in the beginning, because you know how it goes, right, with trainings. Everybody's a bit tired of trainings. But we reached with that voluntary uh training actually 44,000 colleagues, which is amazing for us. This year we made it actually a mandatory compliance training because people need to go through it to be reminded again. But the the uh tailoring to the different stakeholder groups really helped that people know, okay, what do I actually need to do here? The next step that we did then, and I think this is very important also for companies that acting or have business in different countries, is that we did a so-called roadshow. So we were really visiting a couple of countries to interact directly with the business because that allows us also to have a critical discussion. And the discussion is very healthy, even though not always comfortable. So sometimes a bit uncomfortable. But we need to have these discussions to really understand do they get what we actually want? Is this something that makes sense in their context? And obviously connecting in person, although online is great from a cost and environmental perspective, but it's great to see people in person and to allow them also to just come to you and to ask the respective questions. And lastly, obviously, after that, we have rolled out a bigger AI literacy program that we're also moving again more in a role-based approach, and we have pushed out for the different stakeholder groups actually guidelines for an orientation so that they can navigate it. As a last point to mention all of the use cases. So whenever somebody wants to actually buy a new system or build a new system, we have a process that starts very, very early. So already in the ideation phase, the security team, the legal team, and the ethics team are being consulted by the use case owner so that we can ensure that while we're building, while we're implementing, while we're piloting, we have already the right elements in place and the right discussions happening.

SPEAKER_00 12:10

I I really like what you mentioned. Basically, to sum it up, you said it's important if you roll out literacy trainings and if you want people to actually implement those principles and follow them, it has to be role-based. Or first you have to actually understand what these people are doing day to day. Just talk to them, and then try and tailor this training to their use cases. But also that this training is uh customized to different groups in the company or or different people in the company, depending maybe on their AI use or how high risk their use cases are, or what kind of tools they have access to. Also involving a legal department, what were the two other departments? IT, security, security, yeah, security is.

SPEAKER_01 12:50

Always very important, and then the ethical part, which is currently covered by me.

SPEAKER_00 12:55

What I observe a lot that many companies right now have challenges creating these governance systems uh around AI. And again, maybe it's something that they should have been doing already 10 years ago or 20 years ago, you know, starting with data. Once they realize there is a lot of data in the company and you like you have to somehow also control the data and where it is going, and so on. But now it became really an issue with AIU Act and actually making sure that you know what are all the systems that you're using, what risks do the systems have with classification and so on. But for many companies right now, it's a huge challenge. So, what would be your advice to get governance in order, whether it's to comply with AIU Act or being able to provide this kind of services within the company and making sure it also works?

SPEAKER_01 13:44

That's a very good point. And I have to laugh a bit about what you said, because honestly, that's for me one of the biggest points. We should have done that in many companies quite a while back. Because oftentimes we forget that generative AI produced this hype, and some of the companies actually do label only generative AI as AI. So for me, it needs to start at a very simple step that most of the companies don't do from the beginning, getting their terms clear, right? So really producing a glossary or an aligned language across the company, what AI actually is. So, what does it mean if we're setting up an AI governance? Does this include predictive algorithms? Does this include traditional machine learning? And only if you start with the very, very basic, and it sounds a bit weird, but it is what it is, right? If you if you start with the very basic, then you can, as a next step, identify where do we actually have these systems already in place? And this is something that becomes more and more critical on the one hand side, but complicated on the other. Many have a catalog of whitelisted or allowed systems. Nice. 50% at least of these whitelisted systems, and most of the companies have now an AI component. Are you aware of that? So in defining actually first the baseline and then looking into where do we have the baseline actually already in our company, that gives you a very good starting point to understand: do we have something? Do we have nothing? Where do we need to start with? And if your split is you're purchasing 70, 80, 90% of your systems, then your governance focus should maybe not be, am I now somebody who's building general purpose models or not? That's not your question to ask. Then it's more about how can I actually ensure that the systems that we purchase have the right flow and the right questions? The systems that we have already implemented, do we need contractual enhancements? Do we need to ask about different documentation from our vendors? Do we actually understand where AI really is in these systems? And what's the transparency obligation towards our vendors for those? So that's a very different starting point than if I'm a company who does 50, 60, 70% of development themselves. They want to look into how do we actually define our use case? What does this mean? How do we choose the right type of AI that we want to implement? How does a proper AI ops flow look like? Where do I need to have monitoring in in place? What's the database line, etc., etc. Right? So here I think for me the most crucial point is to set your baseline. And many companies ignore that step, but from there, it gives you actually quite practical steps, to be honest, in my opinion at least, on what do you actually need to ask for or what do you need to build in able to satisfy then, so in order, sorry, to satisfy then also what the regulation actually requires from us, or what the regulation actually does not require from us, and we're just over engineering.

SPEAKER_00 17:25

I I really like what you said because when I talk to businesses, when you start with look at what you already have in the company, like create some kind of like inventory of use cases. But I like that you started that. Okay, there's actually a pre-work, even before that stuff that you have to do. You actually have to set this baseline, or you have to even understand what AI are you talking about. Right. Predictive AI, generative AI, maybe a genetic AI if you already have agents in the company. So what this AI question related to this. So sounds like it's a lot of work, right? What I observe in the companies is that it either falls into the IT team, which quite frequently happens, but it seems like it would be a better approach if let's say companies that understand right now, okay, we have all of this responsibility, well, both by regulation, but also because we want to stay competitive. We need to create a department and you know, appoint a specific people, a business owner who is responsible for that. How do you see this? Giving an advice to other companies, like how do they should set up, who should be responsible for this question? Because it's not something you do once, because the system evolved. Maybe today your your tool that you've been using for 20 years doesn't have AI, but tomorrow they have their AI features, so suddenly you have to look into that as well. Can you give any advice or intel on setting up governance? Who should be responsible for for doing this in the company?

SPEAKER_01 18:46

I love that you're asking this question. Because, in my opinion and how we have built that in our company, as I said, it's not a one-person show, and it will never be, because there is also too much responsibility behind that. So we have actually built it, I would say in a rather uncommon way, but by purpose. We have actually a distributed split of responsibility, as I already said. And we have positioned the responsible AI principles and basically the me as the responsible AI manager underneath under a group public affairs. Very different than many others are doing this. Why? Basically, because we wanted to make sure that it's not seen just as a compliance tick box and that the legal lens somehow is independent. Then many companies put that either under legal or under IT. But IT is a department that obviously is very much driven by, first of all, costs, but second of all, also innovation, which is a bit more, I would say, a department with a more risk appetite by nature. So we also said that's actually not the point where responsibly I should sit. Nonetheless, what I'm mainly doing is I coordinate with the others. So I'm looking at these topics from the ethical, human-centric, and transparency side. And obviously, I'm coordinating the other teams here, but we have the legal department always in there with AI and legal tech from an IP standpoint and from an employment law standpoint. We have IT security always represented. We have enterprise architecture in our discussion group. We have data science, so really the experts with the meat on the bones in the discussion group. And we have the purely AI governance perspective, so more from the IT governance side, also on the table. And only because we're covering that from a holistic level, I would say we can cover all of those different angles that are important to implement. And to be honest, sometimes there are also phases where we just need to split responsibility and focus area because one person or one team could never cover all of that. So everybody is working on their angle, on a specialized topic. We come together, we discuss, we adjust, and then we're pushing it out. It is a bit more coordination, and coordination has sometimes also its bottlenecks, no question about that. But I would say, especially in a topic that is so complex as the AI landscap today, it's absolutely worth to put in the effort of coordination and alignment between the teams.

SPEAKER_00 21:48

So, what you're saying is that it cannot be just something you do on top of your day to day job, right? That's first. Second is that the dedicated business owner or or team actually are needed to do it full time, but But also you need the coordination with all other departments like legal, like IT security, data scientists. You really have to have as a company, you also have to make sure to empower these people to have time to do this, right? Or to give them some time them time to do this.

SPEAKER_01 22:14

Yes, and for me it's clear that if you're a smaller organization, you might not be able to hire five new people for every department, right? But I think here really just carving out a bit of time from the respective teams, from security, from legal, and so on and so forth is absolutely worth it.

SPEAKER_00 22:32

It's a very interesting conversation, but I know another passion topic of yours is you talk a lot in the conferences as well about um things like anthropomorphization of AI. I almost cannot pronounce it. And for instance, you're given a talk at Genai Zurich in April called The Illusion of Intent about how saying things like the model thinks or the AI decided isn't just imprecise language, it's actually a governance hazard. And can you walk us through what you mean by that?

SPEAKER_01 23:02

Yes, absolutely. And you're right, I'm very passionate about this topic, so I'm happy that you're asking. So, first of all, not everybody is familiar with the term of anthropomorphism, and it's a just something that always breaks while you're spelling it out. But in the end, it's the tendency of us humans to put human attributes on objects or animals, so non-human things. And I think we need to understand a bit why we're doing this as humans. It's actually very simple. First of all, we try to make sense of something that is for us a bit unpredictable and unknown. So we want to navigate that complexity, and the easiest thing is to give it attributes that we are very familiar with, so human-like attributes. The second point is that we are a social species, right? So we always try to create these bonds between us humans, but also with things. So, how can we do that? We are giving it again human-like attributes so that we can connect with it on a social level. And this is very important to understand as the baseline of what is happening actually. And in Cambridge, one of my professors, for example, has shaped a new term which is called anthromimesis. Basically, here or anthropoporimetic design, even harder to pronounce. But here we're speaking about the intention or in an intentional tailoring of these systems so that humans actually think that the system is human-like. And we see that a lot, or to attach human attributes. We see that a lot, for example, in ChatGPT itself, where this tendency that it always agrees with you and it greets you by your name and so on and so forth. Those are conversations flows that are very, very familiar with us by nature, because if you speak to your friend, similar behavior. So, all of that to lay the baseline what we're speaking about. Why is this now dangerous, especially if we're speaking about a field like human resources and recruitment that is in itself very human-centric and human-driven, is basically that many people in today's society do not really understand what an AI system is. For many people, this is very new still, and they don't really know how to navigate it, to trust it, not to trust it, to just take what comes out, to not do so. And we see that with all of these AI companions and boyfriends and girlfriends, we have seen the dark side of different suicides, different people who are really getting into a depressed state because of that. So we have seen that there is the real potential for harm. Now, if we as a company produce the false impression that this is a human-like thing that people absolutely need to trust over everything else, then actually all of our efforts of putting the human first, ensuring that we are behaving ethical, ensuring that we are transparent, is gone. So if you can't distinguish between Maria, your because it's so friendly and pretty and has a human-like appearance, and the actual recruiter, how shall you notice if this is really a decal group's bot or just a scammer trying to get your data? Or if this bot says you're too old for a new job, which will never happen with our bots because we're tap we're testing, obviously, for for this on a very frequent basis, although maybe I should never say never, but we're doing our best that this is never happening. Basically, it feels you might feel offended by our company, although this is something where just a system went nuts for a moment that we couldn't control. So here it's very, very important that we're making clear from the beginning. These systems are systems. Systems can make mistakes, systems are not a human, they can't feel like a human, they don't empathize with you like a human. They're just here to automate some parts that we can't cover solely with humans. If this is clear, the expectation is a different one. And that creates then also, in the end, more trust and more ethical behavior.

SPEAKER_00 27:42

That's a very interesting perspective. For me, it's even controversial in a way because uh, see, like what I'm doing personally as AI trainer, when I teach, let's say, custom GPT or even agents, it's like my AI team, you know, there is this uh podcast producer, Mark, and there is this sales assistant Anna, and you know, and you give them like a role description when you write your instructions, you tell them how they should communicate with you, and so on. And it does help the adoption. I even talked to uh one of the researchers recently who is uh researching acceptance and AI adoption and why people resist AI, who's actually said that hey, uh we could call it like smart technologies and not AI, and we could actually make the make it feel more human so people actually adopt it easier. So you're literally saying that our kind of AI adoption strategy, not so safe to use it. So, what are the alternatives? So people actually are more accepting this new technology and being willing to engage with it.

SPEAKER_01 28:38

I absolutely agree with you, and that's the purpose, right? Why do I do that? I want people to create social bonds, I want people to actually do not see that as something that is dangerous, but as your best friend that creates adoption. But exactly that is the point. As long as people are not literate enough to really understand what's behind it, that's a dangerous approach. So, what we are proposing to do, or what we also are developing, is the transparency baseline. So we're not giving it a human-like appearance, it doesn't think, feel, or has any human-like states. But we are very clear in explaining what is this system actually about? Because as soon as you remove the unknown and you remove especially the language that is abstract so that not everybody can understand it, then people will also adopt because they do suddenly understand what it is. But here we're coming back to understand your stakeholders, right? Because you can only create a transparency note or whatever, an explanation of what your system is if you understand at which level your baseline user is.

SPEAKER_00 29:54

So staying with psychological hazards and when interacting with AI systems with technology, I came across one of your LinkedIn posts about something you call anticipatory self-reduction. The idea that people are now actively stripping away their own complexity to become more machine readable. And I think your example was again with recruitment, because many AI systems obviously now kind of pre-screening candidates, and so the candidates are trying to make themselves more fitting within this description or within this box, so they are more likely to be selected by this machine or this AI system on the background. So tell us more again about what this for you means, anticipatory self-reduction, maybe where you see it manifests, and again, how do you see it is developing in the future and what can we do about it?

SPEAKER_01 30:42

Yes, very happy to. I mean, this comes or we see this tendency more and more, and that's why I was writing about it. We need to go back one step. And what is reality already for a long time, and that has nothing to do necessarily with AI, but becomes stronger through AI, is that people are reducing elements of themselves to fit into a specific type they think is required from the humans, from the companies. And to give you an example, there is a term that is called resume whitening. And that is existent from before AI, but basically it is that especially people with a darker skin color or with a name coming from an Asian or African country are actually, first of all, changing their names, removing certain words that would lead to their origin, removing images, for example, from themselves so that people can't see firsthand that they are belonging to a certain origin or ethnicity. Given that, we see that happening now more and more on a larger scale because of AI, and because so many companies are leveraging AI algorithms to filter and screen candidates in the background. And you might be familiar, but there were quite a couple examples, not so recent and recent ones, that show how discriminatory those algorithms are. The most prominent example is the one from Amazon in 2018, where the algorithm said women are not suitable for tech roles, because no women were applying traditionally to those roles, so the algorithm picked up not suitable. Um then we had hard view with face expressions, where basically your face expressions were red, and this was highly discriminatory against people with speech impairments or basically with neurodivergent patterns. We have seen that just recently with workday as well, right? Where the algorithms were shown to be biased against certain people of certain age groups, for example. So the more we see that and the more cases we hear coming up, the more people are asking themselves the question: do I fall into one of these categories? Now we need to put one level on top, unfortunately, which is that most of the job descriptions are written by LLMs, right? It's easy, it's fast, it saves time and money. But candidates see also that they have more success with their applications or at least getting interviews if they tailor their resume with an LLM so that the right buzzwords are popping up. And that became even worse through the pandemic and moving everything into online because you haven't had the chance to interact in person any longer. You were doing everything just online, right? So you're anyway reduced to this two-dimensional image almost. So what we see is that many people do not spend time or effort any longer in tailoring their resume themselves to really stand out what they think is important about themselves for this role. They just ping it into an LLM. The LLM creates some buzzwords that sound fancy or blows up some skills that it thinks for it thinks. You see, I'm doing it myself. It actually sees statistically were better in the past. And then the applicant is sending the application. Now, what most of the time is striking is that these systems obviously are just searching for buzzwords. So the more buzzwords you have in there, the higher the chances are that you're at least invited to an interview. And secondly, we see the self-preference bias of an LLM. So if you have written your CV with JetGPT and the company is using a JetGPT model in the background, the likelihood that you're getting the job is higher. So, with all of that, actually, we're not putting the emphasis here more on us. And if we not stick only to the documents, we see then candidates not answering truthfully in an interview, but actually always having their assistant on the phone next to them that gives them the answers, even though it might be not your personal opinion. It's just what ChatGPT blurs out. Or we see people actually not doing the assessments themselves any longer, but let them be done by an AI assistant. Does this reflect your true self? It's not at all, right?

SPEAKER_00 35:37

No.

SPEAKER_01 35:38

So, and this is getting more and more drastic so far that in some countries we're actually moving back to more in-person interviews because there is no other chance than removing the online element from the equation.

SPEAKER_00 35:52

Yeah, it's very interesting. I think it could be a larger society challenge in the coming years. It it has been a really amazing discussion. Is there anything else that you would want to uh tell to the audience, to the business leaders, let's say, who are contemplating all of these topics: governance, responsible AI, AI literacy trainings, AI adoption. Is there some advice you want to give or something else that you want to add that we haven't discussed so far?

SPEAKER_01 36:16

I think there's just two short messages that I would like to give to every leader, actually. The first thing is AI is really not a software system in the traditional sense. Don't try to scale it across all countries just as is because you have done that with your CRM and ATS systems in the past. It's not the same. It's dynamic and it needs to be treated like this. So please, please educate yourself what it is, and don't just stick to the old patterns. That's the first thing. And the second thing is get out of your ivory towers. And we're a big company, we have an headquarter also in Zurich, right? So oftentimes we're stuck in this mindset very naturally, as everybody is. But the healthiest thing you can do is to engage with your countries and with your employees and really to speak to them about what do they need, what understanding do they have, and what needs to slow it. And only then can be really successful with response play and response play governance.

SPEAKER_00 37:24

Thank you so much, Sarah. It was an amazing discussion, like a deep dive into so many topics that concern many organizations, many companies, and many people right now. Thank you very much. It was a pleasure. You can find Sarah Matthews on LinkedIn. The link is in the show notes. If you enjoyed this episode, follow AI Made Simple, the transformation series, for more conversations with practitioners shaping how AI is actually governed, adopted, and skilled inside organizations. Thanks for listening.