Latest / The Road to Accountable AI / Caroline Louveaux: Trust is Mission Critical

Transcript

Caroline Louveaux: Trust is Mission Critical · The Road to Accountable AI

• 0:00
  Caroline, welcome.
• 0:01
  Thank you so much for joining me.
• 0:04
  Thank you.
• 0:04
  Hi, Kevin.
• 0:05
  Nice to meet you.
• 0:07
  You too.
• 0:07
  So MasterCard is a data company really at heart.
• 0:13
  That's, that's really the core of what the company offers.
• 0:17
  So I'm curious how that heritage influences the way that you think about first privacy and data protection, which I know is one of the areas that, you have focused on for a long
• 0:29
  time, but, but then also AI.
• 0:32
  You've done your homework, Kevin.
• 0:35
  Very few people know that MasterCard is a data company, but you're right, we are.
• 0:40
  And if you think about it, we have access to a lot of data, know, transaction data across a multitude of banks and financial institutions.
• 0:49
  And so we have decades of experience leveraging that data and leveraging AI, not only to process these transactions to make sure that you can make a payment, but also to enhance
• 1:00
  fraud protection.
• 1:02
  and to combat crime and cyber attacks.
• 1:06
  so all different shader, if you think about it, is about data or technology, but also our global scale and our expertise.
• 1:16
  But of course, nobody would even rely on our network if they would not trust it.
• 1:22
  So trust is mission critical.
• 1:24
  And this is a reason why back in 2019, we launched what we call a data.
• 1:29
  and technology responsibility principles.
• 1:33
  These are a set of consumer centric principles that guide everything that we do with data and AI across the company.
• 1:40
  And look, they are very easy to understand from experts.
• 1:44
  It's your data, you control it, you benefit from it, and we protect it.
• 1:49
  As simple as that.
• 1:51
  And then coming back to AI, uh yes, our expertise in AI has also been foundational for us.
• 1:57
  ah
• 1:59
  Our main use case is really cyber security and fraud protection.
• 2:03
  And so we've been investing about 11 billion in cyber over the last couple of years.
• 2:09
  And last year you may have seen that we've acquired a new company called We Go To The Future.
• 2:14
  And they are the largest threat intelligence company in the world.
• 2:18
  And what they do is they really um enhance our ability to anticipate and prevent cyber threats uh before they happen.
• 2:26
  So this is just amazing.
• 2:29
  No, that's great.
• 2:30
  So what's different though about AI, especially today, um where things have evolved tremendously, there's generative AI as well as machine learning.
• 2:40
  What's different about thinking about these issues in the AI context versus purely data protection and privacy?
• 2:48
  Thank you, Kevin.
• 2:48
  Look, there are a lot of similarities, but I would mention definitely some differences.
• 2:52
  Let me mention three of them.
• 2:54
  One, and you mentioned it, speed.
• 2:58
  The pace of change in this area is unprecedented.
• 3:02
  A world where AI agents can make decisions on your behalf was just science fiction a couple of years ago, and right now it's the reality.
• 3:10
  Agent commerce is here.
• 3:12
  You can ask actually your agents to make transactions for you, Kevin.
• 3:16
  It's super convenient, but of course it also creates some risks, cyber risk, privacy risks, and that's the reason why AI governance has to constantly evolve and adjust to
• 3:27
  address these new risks.
• 3:29
  And then I would like to mention two of the differences.
• 3:32
  One, AI governance is definitely a collective sport.
• 3:36
  It is a shared responsibility across multiple teams.
• 3:39
  No one can do this on their own.
• 3:41
  Because if you want to have effective AI governance, you have to bring
• 3:45
  experts from across different skill sets, backgrounds, disciplines.
• 3:49
  So of course you need technology experts, engineers, but you may also want to bring to the table lawyers, policy experts, risk officers, data scientists, human science experts.
• 4:03
  So the more you are, better in this space.
• 4:07
  And then the last, maybe let me mention one last point is AI literacy that is extremely important.
• 4:14
  At end of the day, everyone uses AI today ah in the workforce, but also your parents, your kids.
• 4:21
  But do they even know how to protect themselves from AI risks?
• 4:24
  Of course not.
• 4:25
  So if you want to have effective AI governance, it's really critical to bring everyone along the way on the journey.
• 4:32
  What exactly does that mean, AI literacy?
• 4:35
  I hear people talk about that and it's in, for example, the European Union's AI Act and it makes a lot of sense what you're describing.
• 4:43
  But how do you actually implement that idea that people need to understand what the technology is doing?
• 4:50
  That's a great question, Kevin.
• 4:51
  AI literacy doesn't mean that you need to learn how to code.
• 4:56
  That's not what it means.
• 4:57
  You don't need to be a technical expert in order to be literate.
• 5:02
  I think what we mean by that is understanding the potential risks and be cautious so that you use it, you experiment with it, but in a way that is responsible and that protects you
• 5:11
  from some risks.
• 5:13
  For example, when you prompt into an AI tool,
• 5:16
  you may need to adjust the way you prompt.
• 5:18
  You don't disclose personal information there.
• 5:22
  You understand that what comes back may not be the reality.
• 5:25
  You may want to validate your sources or double check the responses, understand that they may be biased in the response.
• 5:33
  So it's more like about understanding the limitations of AI and how to make the most of it so that it benefits you without actually creating some risks for you.
• 5:43
  Does that make sense?
• 5:44
  Absolutely.
• 5:45
  ah Absolutely.
• 5:46
  Now, and it's, it's important to understand all these pieces come together.
• 5:49
  There, there's a number of things in, that you already said that I want to follow up on, but let me just push you a little bit more on, on some of the, the practicalities.
• 5:56
  So, there is this notion in, privacy, of privacy by design, that, that you have to implement privacy protection all the way throughout the, life cycle of data.
• 6:06
  And that seems pretty comprehensive.
• 6:08
  So, as we move now into an AI era.
• 6:12
  Where do you need a new playbook or where is it just a matter of extending those existing mechanisms that a company like yours has in place to just new kinds of analytics tools?
• 6:25
  It is a little bit of both, So you're right.
• 6:29
  Privacy by design provides a solid foundation for AI governance.
• 6:34
  And for example, we've been able to leverage some of the principles of the GDPR, for example, transparency and control, accountability.
• 6:42
  We've also been expanding some of the processes that we had established for GDPR compliance.
• 6:48
  You talk about privacy by design, you can think about data subject access requests.
• 6:53
  incident response, impact assessment.
• 6:57
  So many, many processes can be leveraged and expanded for AI.
• 7:03
  These being said, in the context of AI, you have to go above and beyond on certain areas.
• 7:09
  And remember, the GDPR and privacy laws only apply to personal data.
• 7:15
  But AI governance has to apply to all systems and applications, irrespective of the nature of the data.
• 7:21
  So let me mention a couple of things that are really specific to AI.
• 7:25
  So one, you have to have um a comprehensive inventory of all your systems and applications leveraging AI.
• 7:34
  That may seem very basic, but believe it or not, it's complex.
• 7:37
  In a global organization where everyone's trying to do something with AI, herding the cats is not that easy.
• 7:45
  Second, some principles like transparency,
• 7:50
  they have a different color of perception or different uh connotation in the context of AI.
• 7:57
  Because what you need to do is to come to explainability.
• 8:00
  So you need to explain to people how the AI works, how decisions are being made, and what are the limitations coming back to literacy.
• 8:10
  And so we have started to establish model cards or explainability protocol to explain to our customers, which are often the banks.
• 8:19
  You know how our AI works.
• 8:21
  And of course we speak a lot about oversight, you know, having a human in the loop or human oversight is really important, at least for these applications of AI that make
• 8:30
  decisions for people.
• 8:32
  um Let me mention that.
• 8:36
  I'm going to pause here.
• 8:37
  This being said, I want to come back maybe later about the fact that there's sometimes some tension still, some tension between privacy and AI governance.
• 8:46
  So, you know, while most is aligned,
• 8:49
  Sometimes you have some tensions.
• 8:50
  Okay, no, can't let that go.
• 8:51
  So why don't we talk about that now?
• 8:52
  What are the tensions?
• 8:54
  You know what, that's interesting.
• 8:56
  For example, in our financial industry, um as you know, we have been leveraging data and AI for decades to fight against fraudsters.
• 9:06
  But you know what, actually, that requires a lot of data, a lot of training data, because that provides better, more accurate fraud outcomes.
• 9:14
  But in how do we reconcile that with the principle of data minimization?
• 9:20
  So data minimization and
• 9:23
  accuracy of AI models and then data minimization from a privacy perspective.
• 9:28
  You have two very important principles that are not always fully aligned.
• 9:35
  And so that's the reason why we worked very closely with policymakers and regulators to really advance the thinking there and to make sure that we come up with a practical
• 9:44
  framework that provides clarity for the ecosystem.
• 9:49
  Okay, let me ask you one more thing about MasterCard in particular.
• 9:54
  uh You mentioned that your direct customers are typically banks and credit card issuers and so forth, but MasterCard is also a consumer brand.
• 10:03
  Most people are familiar with the priceless ads and so forth.
• 10:06
  So how does that structure in terms of having a broad public consumer presence, but directly interacting with institutions?
• 10:16
  influence the way you think about governing AI.
• 10:22
  You know what, actually, for AI to work, trust is everything.
• 10:27
  So I often say that AI governance is also an ecosystem play.
• 10:32
  You need to make sure that all participants in the value chain, in the payment ecosystem, uh do the right thing for trust to be there.
• 10:42
  So what we're doing is we are providing a lot of tools and solutions to our customers to help them on their journey.
• 10:49
  and to help mature the program.
• 10:51
  for example, we provide, as I mentioned earlier, we provide visualization tools and model cards to explain in simple terms how AI solutions work so that they can then explain to
• 11:03
  the consumers.
• 11:05
  And we also provide them with tools to quantify some risks in their training data.
• 11:08
  So for example, privacy risks, bias risks, to help them as a solution.
• 11:15
  And then for the smaller customers, what we also do is we provide a capacity building workshops to educate them because not everybody has the resources and the expertise in
• 11:24
  house, right?
• 11:25
  And the space is moving so fast.
• 11:28
  So having actually um these conversations where we say, hey, these are the latest trends.
• 11:33
  These are some tools, some resources that you could use.
• 11:36
  And these are some best practices that we have identified is extremely useful for the smaller customers.
• 11:45
  You mentioned earlier that AI governance is a team sport and it requires these different areas of expertise and parts of the company to be involved.
• 11:52
  Can you talk specifically on how you structured that function within MasterCard?
• 11:59
  Yes, so our AI governance team.
• 12:03
  So first of all, we have an AI governance team, right?
• 12:05
  And it sits already across two different teams, across my team.
• 12:09
  And we can discuss about my team in a second.
• 12:11
  So I'm part of the law department.
• 12:14
  And across our data strategy team, that is part of our AI office.
• 12:18
  So it's the business.
• 12:19
  uh
• 12:21
  And so that team is already comprised of many different skill sets.
• 12:24
  So we have lawyers, we have data scientists, have engineers, communication experts, we have a little bit of a broad representation of skill sets in that team.
• 12:34
  They deal with the day to day.
• 12:37
  And when they have like some major developments or some, they identify some risks, they elevate these to an AI governance council that I chair together with our chief AI officer.
• 12:51
  And that console has uh representation across all different functions at the executive level.
• 12:57
  So it provides executive oversight over our novel use of AI, if you want.
• 13:03
  uh And this is how we bring the whole company together so that we have an overview of everything going on and we are sure that we stay true to our principles.
• 13:14
  What triggers uh issue going up to that council?
• 13:21
  You know what we do?
• 13:22
  We have an intake process.
• 13:25
  And the intake process assigns a priority level to use cases with risk assessment, scorecard, et cetera.
• 13:33
  And in terms of taxonomy, our approach is that any use case that has the potential to impact people or to impact society needs actually a deep dive review.
• 13:45
  And so as you can see, this is actually more inclusive than, for example, the EU AI Act.
• 13:51
  which has a very narrow definition of what is considered to be high risk.
• 13:54
  We want to take a safe approach there.
• 13:58
  Then the teams are doing the analysis.
• 14:00
  And if they view that a particular activity has the potential to raise some risks for people in society and cannot be easily mitigated, that's where it's being escalated.
• 14:10
  And to be honest, if the council considers their use case is worth going on, because for example, uh it's a very important tool to fight against fraud.
• 14:21
  Then actually we use our enterprise wide risk management processes because we have at Mastercard, you know, as in any big company, an enterprise risk management function that
• 14:33
  is owned by our chief risk officer who we put directly into our CEO.
• 14:38
  And the benefit of that is that we then document and track these risks in a similar way for everything.
• 14:47
  That's a lot of different parts of a big organization, right?
• 14:50
  You've already got uh legal compliance, privacy, AI business and risks.
• 14:57
  So how do you effectively encourage those different groups to work together?
• 15:03
  That's a great question.
• 15:04
  You know what?
• 15:05
  I think the answer is very simple.
• 15:08
  The tone from the top, Kevin.
• 15:11
  Once you have a very clear message from the top, um everyone wants to play around and wants to play with you, right?
• 15:20
  And wants to help out and wants to be part of the success.
• 15:23
  You know, data and tech responsibility is part of our corporate objectives every year.
• 15:29
  And we just had last week, we had our leadership offsite with our CEO in Boston.
• 15:36
  And these topics have been again front and center, including responsible AI, not only AI innovation, but in a way that is responsible.
• 15:44
  So we think that the companies who are going to win in the next decade are the ones who can innovate fast, at scale, but also do so responsibly.
• 15:53
  And we think that...
• 15:54
  there, isn't there, between those two things?
• 15:58
  I'm sorry to cut you off, but maybe it goes back to something you alluded to before.
• 16:02
  People think fast and responsible are inconsistent sometimes.
• 16:08
  No, I think that's the easy route.
• 16:09
  I think it's definitely possible.
• 16:11
  And our company is convinced about that, that you can do both.
• 16:15
  And if you want to be sustainable in the long run, you have to choose a path of responsibility and trust.
• 16:24
  No, that's great.
• 16:25
  Obviously, every company is struggling with these issues.
• 16:28
  Is there anything else that you can talk about within either the culture or the message that comes down from senior leadership at MasterCard that gets people across the
• 16:39
  organization to accept that framing?
• 16:43
  I think it's also about repetition, right?
• 16:45
  It's good to have it in your corporate objectives, but at every single town hall, this is brought back to the scene.
• 16:54
  so people, all employees hear that on a very regular basis from the top of the house, and that makes a huge difference.
• 17:02
  And you know what?
• 17:03
  think it also, you want to make sure also that your legal team has the right mindset.
• 17:08
  It's not about saying no, it's about if, yes, if, right?
• 17:13
  and finding solutions together.
• 17:14
  Right now, are so many exciting, for example, technologies, Kevin, like privacy enhancing technologies that help data to be collected, used, and shared in a way that you can
• 17:26
  extract the value from the data and from AI while fully protecting privacy security and confidentiality.
• 17:33
  I'm thinking, for example, about synthetic data that is fake data, right?
• 17:37
  You can use that, actually, to train your models.
• 17:42
  And so you can extract all the values from your AI models, but it doesn't, it's fully compliant with privacy because it's not about real people's data.
• 17:50
  And there many, many new technologies there.
• 17:52
  So I think it's also about being creative about striking the right balance between fast and responsibility.
• 18:03
  You mentioned before that, uh for example, in some areas, your internal process goes beyond what's required by the EU AI Act in terms of what triggers a high risk review.
• 18:16
  Can you talk more about, there's a variety of external standards and frameworks like the NISQ AI risk management framework in the US and ISO standards and so forth.
• 18:27
  What do you look to or benchmark against uh in
• 18:31
  thinking about how you apply these principles.
• 18:35
  So we've decided to establish a single global framework that applies globally, one.
• 18:42
  And then we incorporate all these standards and legal requirements into our framework.
• 18:48
  And so the EOEI Act has been one of them.
• 18:51
  The NIST uh Risk Management Framework has been another.
• 18:54
  We've been looking at the Singapore frameworks as well.
• 18:57
  And so we update our framework on a regular basis to make sure that it's...
• 19:01
  um
• 19:02
  It reflects all these requirements and standards.
• 19:05
  And of course it's grounded in our data and tech principles as well.
• 19:10
  By doing so, we provide a business with a one-stop shop.
• 19:14
  And they just love it.
• 19:15
  You sometimes it's much easier to deal with one standard, even if it's a strict one, rather than a multitude of different standards around the globe.
• 19:23
  It's just not manageable.
• 19:25
  If you want innovations to scale.
• 19:28
  We try also to automate some of the risk assessments um as much as we can to make it easy for our AI innovations to go fast so that we automate for low risk and we have a deep-dark
• 19:39
  review for what we consider to be higher risks.
• 19:44
  And then the last thing that we do there, we are heavily engaged with standard setting bodies and with policymakers to really promote for and to advocate for convergence of
• 19:54
  standards.
• 19:56
  It doesn't serve anyone, Kevin, to have fragmentation of standards.
• 20:00
  The more we can align on key principles, the better for everyone.
• 20:05
  How do you see the evolution of global standard setting in this area going?
• 20:13
  We are not on the journey there.
• 20:14
  We're on the journey.
• 20:16
  And the thing is, it's not only about AI standards, right?
• 20:21
  I think if you want to make it work for AI, you also need to have some convergence on data standards.
• 20:29
  And to be honest, today we face really a patchwork of complex and fragmented standards when it comes to data.
• 20:35
  And data is being regulated by privacy laws, but also payments laws in many different regulations.
• 20:42
  I often talk about the soup of digital regulations.
• 20:48
  And unfortunately, these laws, they overlap and they sometimes diverge or even conflict with each other, which is very challenging to navigate.
• 20:55
  So yes, I think we have to continue uh to advocate for this convergence, which is really, important.
• 21:03
  Yeah, I hear this a lot from professionals in this area, from major companies, that there's an assumption that people have that industry wants no regulation and complete
• 21:16
  freedom to do everything.
• 21:17
  And no, I think what you're making is an important point that if there's clear standards and it's aligned with the business objectives, then that's uh advantageous to a company
• 21:30
  like
• 21:31
  To be honest, think that Enmaltekat has a history and track record of innovating, taking actually the regulations to innovate.
• 21:40
  So back in, for example, back in 2018, when the GDPR came into force, everyone at the time was still struggling with just basic compliance.
• 21:50
  We have used GDPR as a catalyst for innovation.
• 21:53
  At the time, we created TRUATA, which is a combination of trust and data.
• 21:59
  And it's a company offering privacy enhancing technologies so that companies can continue to innovate with data and conduct analytics in full compliance with the GDPR.
• 22:10
  And so I think there are many ways to innovate with regulations.
• 22:13
  I think again, the worst case scenario is fragmentation.
• 22:19
  Yeah, and related to what you just described, there now is a whole industry of AI governance and data protection service providers.
• 22:29
  All of the big consulting firms have major practice areas in this field.
• 22:35
  How much are you able to build in-house versus needing to or wanting to use that external expertise?
• 22:43
  You know, we've built our AI governance program back in 2018.
• 22:48
  It was way before the AI hype.
• 22:50
  At the time, no one was talking about AI, you know?
• 22:54
  And so there was none of these providers were there at the time.
• 22:57
  So we've built everything by ourselves from scratch.
• 23:01
  This being said, of course, we are constantly investing in our program and we are on the lookout for solutions and technologies out there that could complement what we've built
• 23:11
  in-house.
• 23:12
  And so we are looking at different partners.
• 23:16
  We also working with a number of academic institutions and tech providers to see actually how we can further enhance our program.
• 23:26
  And then we also joining, we've joined as a member of the Data and Trust Alliance.
• 23:31
  I don't know what you, yeah, you are familiar with them.
• 23:33
  So one of the pillars and they goal is to publish best practices.
• 23:37
  Because that you don't have that many.
• 23:39
  mean, you have this.
• 23:40
  the typical standard setting organizations where it takes a long time to publish standards.
• 23:46
  But actually people and companies are eager to understand actually, for example, what bias testing mean in practice, what explainability means in practice.
• 23:56
  And so I think having some, the data and trust alliance is a not for profit coalition of organizations.
• 24:05
  And for example, one of the standards and best practices that we've published is about data provenance.
• 24:10
  understanding where your data comes from, and being able to trace it back throughout their life cycle, which is extremely important for AI governance.
• 24:18
  um So yeah, that's what we do.
• 24:22
  What would you say is the biggest challenge that you face right now in AI governance?
• 24:29
  That's a great question.
• 24:31
  That's a great question.
• 24:36
  you know, I think I'm going to go back to what I said, speed.
• 24:41
  Everything changes so fast.
• 24:43
  It's hard to keep up.
• 24:45
  And the technology is changing so fast.
• 24:47
  If you want to really, really understand the potential risks, you need to go to deep dive into these technologies.
• 24:54
  And then regulations are changing so fast all around the globe.
• 24:58
  I'm based in Brussels.
• 25:00
  And we often call Brussels the rules factory of the world.
• 25:06
  But this is about Europe.
• 25:07
  And then every single state in Europe has their own stuff.
• 25:11
  And they look around.
• 25:12
  Singapore, Brazil, Chile, Peru, the US, Canada.
• 25:18
  So which says speed of change is the biggest problem.
• 25:24
  can you address that speed of change given all the things you've talked about in a big organization?
• 25:33
  I think it's having um an extensive network of allies being well connected so that actually you have people feeding you with information that is relevant to you.
• 25:49
  I want to go back to something you mentioned very early on, is uh agentic commerce.
• 25:55
  And obviously, there's a tremendous amount of excitement broadly in the AI space about agents.
• 26:00
  But in particular, all the payments companies, including MasterCard, have in recent months announced frameworks for agentic payment.
• 26:10
  talk about how these issues of trust and governance relate to that initiative.
• 26:17
  well, you know, that's a good question.
• 26:18
  For Genticama to scale and to be successful, trust is going to be everything, uh And that's where MasterCard has a key role to play, to leverage our technology.
• 26:30
  And we can come back to that.
• 26:31
  have tokenization technology, fraud protection, to leverage um our rules, our standards setting, because we published a lot of standards, and our expertise.
• 26:43
  And yes, that's actually what, as you know, we have
• 26:46
  published some announcement where we focus on a couple of areas, four areas.
• 26:53
  The first area is we want to make sure that we register and certify agents.
• 26:58
  You want to make sure that you have a legitimate agent and not a fraudster uh making a transaction.
• 27:04
  So we're be there and register and certify these agents.
• 27:08
  The second pillar is about making sure that it's the right consumer who is using the agents.
• 27:15
  And so you need to authenticate the consumer and have a lot of authentication methodologies there that we can leverage.
• 27:22
  But it's also to make sure that we capture the consumer's intent.
• 27:27
  What actually has a consumer, what did they want to buy?
• 27:29
  It's very important.
• 27:32
  Third, in order to make sure that transactions are safe and secure, we're going to leverage our tokenization methodologies.
• 27:39
  And then fourth, we want to create an OVICH trail.
• 27:42
  and to enable dispute management just in case so that everyone is accountable in the ecosystem.
• 27:49
  And I think everyone is going to have to feel safe for this to work.
• 27:54
  consumers are going to have to feel safe.
• 27:56
  Merchants are going to have to be reassured that they're going to be paid.
• 28:00
  So it's going to be important for every stakeholders across the whole value chain to feel confident that the space is a safe space.
• 28:13
  And back to what we talked about before, MasterCard is generally not directly the organization touching the consumer.
• 28:19
  So let's say there's a scenario where someone deploys an agentic payment technology and doesn't take those protective steps you talked about and consumers have a bad experience.
• 28:32
  There are incidents that get a lot of attention about fraud or misuse and it leads to a backlash saying, is risky, we shouldn't use it.
• 28:42
  How can you think about, or the industry think about avoiding those outcomes?
• 28:47
  think actually that's where we come into play where we partner and we publish some standards.
• 28:51
  So yes, not everybody is going to be part of our standards, but I think we're going to have, we partner with a lot of tech providers, the banks, the merchants, so that everyone
• 29:00
  is agentic on us ready.
• 29:04
  And I think everyone is going to have an incentive to be part of that network because they know that actually it's the way that safe transactions take place.
• 29:13
  Now, if there are some...
• 29:16
  Other providers are there who do not want to abide by these rules.
• 29:20
  But consumers are going to know because we're to make sure that it's very clear to consumers if they take place in that safe environment where everybody is bound by the same
• 29:30
  rules or whether they want to do this separately.
• 29:33
  So I think we're going to have to be very transparent to consumers about what they agree to.
• 33:15
  uh One final thing, uh as you talked about, uh you've been at this for a long time at MasterCard on both data protection and AI governance.
• 33:33
  Many companies now are just realizing that they need to put these kinds of mechanisms at the place given the explosion of AI activity.
• 33:41
  So what would be your advice be?
• 33:43
  What would your advice be to a company in terms of where they should start in this process?
• 33:50
  That's a great question, Kevin.
• 33:52
  And I would definitely say appoint someone as responsible for AI governance and give that person the mandate actually to make it happen and make sure that the tone from the top is
• 34:04
  very clear that a mandate is about AI innovation together with AI responsibility because that's the magic formula to have success.
• 34:15
  Fantastic.
• 34:16
  That is a perfect place to stop.
• 34:17
  Caroline, thank you very much for being with me.
• 34:20
  Thank you, Kevin.