Latest / From principle to practice: AI governance for compliance teams / What AI governance actually means for on-chain risk
Transcript
- 0:09SPEAKER_01: I'm David Carlisle, LipTix Vice President of Policy and Regulatory Affairs.
- 0:14SPEAKER_01: I'm here with my colleagues Adam Grimshaw and Casey Chappelle.
- 0:17SPEAKER_01: We're going to be talking about the topic of AI governance and risk management, huge topics of conversation at the moment.
- 0:23SPEAKER_01: But broad terms that could mean different things to different people depending on who you're talking to.
- 0:28SPEAKER_01: But maybe that's a good place to start.
- 0:30SPEAKER_01: Let's begin there.
- 0:31SPEAKER_01: What do we actually mean by AI governance?
- 0:34SPEAKER_01: What are we talking about?
- 0:35SPEAKER_01: How does it relate to the world of on-chain risk and where does elliptic fit into the picture?
- 0:40SPEAKER_02: David, that's a great question.
- 0:42SPEAKER_02: But it almost prompts a first question for me.
- 0:45SPEAKER_02: And that's what do we mean when we talk about AI, right?
- 0:48SPEAKER_02: It's a term that is used really broadly across a lot of different areas.
- 0:52SPEAKER_02: I, Adam, what do you think we mean when we talk about AI?
- 0:54SPEAKER_00: There's been a fairly seismic shift in that definition in the last five years, right?
- 1:00SPEAKER_00: From you know, something that would have been a note-taker, something that was summarizing to a position where we're now seeing AI energentic making decisions and replacing traditional decision-making methods within organizations.
- 1:12SPEAKER_00: So, you know, the the the fundamental answer is the the use of technology and to to make more intelligent learners and knowledgeable decisions.
- 1:22SPEAKER_00: But in terms of the industry and the um operation within your organization, that can differ incredibly wildly.
- 1:31SPEAKER_02: Which leads us then into the question what do we mean by AI governance?
- 1:34SPEAKER_02: And in that context, I like to think about governance as the way that you earn the right to act on the output of the AI system.
- 1:43unknown: Right.
- 1:43SPEAKER_02: And I think that Adam and I, between the two of us, we bring different skills to the table that between those skills build a proper governance program.
- 1:52SPEAKER_02: So my background is in law, policy, regulation, um, uh risk management and governance frameworks in a lot of different areas of technology risk and compliance.
- 2:04SPEAKER_02: Uh, so I tend to focus in governance on the soft controls that that means.
- 2:08SPEAKER_00: I'm coming from a security background, um, which means that traditionally I've focused on those harder outcomes, those harder controls, the the actual technical implementation of security controls to protect the confidentiality, integrity, and availability of data.
- 2:26SPEAKER_00: Um, as we've seen the AI agenda and the use of AI move, many of the parallels that we see within those two areas um have developed into this new space.
- 2:39SPEAKER_00: Not thinking about AI as a new governance framework or a requirement for a completely new way of working, but developing our existing knowledge experience, those control frameworks, those ways of working and the things and the knowledge and experience that we have individually to be able to bring together something which, as a governance framework, works for our organization, works for our sector, works for the industry.
- 3:03SPEAKER_01: But you know, they're broad terms that can mean a lot to different people depending upon who you're talking to.
- 3:07SPEAKER_01: Um, so it'd be good to start by exploring, you know, what what are the issues around AI governance and risk management that relate to on-chain risk and the space we're in?
- 3:16SPEAKER_01: And um, where does elliptic fit into that picture?
- 3:18SPEAKER_02: Oh, that's a great question, David.
- 3:20SPEAKER_02: Um when Adam and I think about governance in general, we think about a combination of hard controls and soft controls.
- 3:27SPEAKER_02: And, you know, I feel like we work really well together because we bring both of those to the table.
- 3:31SPEAKER_02: I think about it in terms of those soft controls.
- 3:33SPEAKER_02: How are we building the protocols, the decision making, and the accountability that earns us the right to use the output from AI governance?
- 3:41SPEAKER_00: I tend to look at the harder controls.
- 3:44SPEAKER_00: So, how are we actually implementing uh controls which provide us with the oversight and governance at a technical level within our technology stack, within the um AI models, and also with our hosting, but also within our tools as well.
- 3:58SPEAKER_00: So taking a very deep dive into uh how those technical controls piece together and ultimately produce the metrics, the information, which allows us to be able to demonstrate as part of that governance program doing the right thing.
- 4:12SPEAKER_02: One of the interesting things about providers in this on-chain risk space is the aspect of the analytics that we perform, right?
- 4:20SPEAKER_02: So governance for us really has to start with the idea of data integrity.
- 4:24SPEAKER_02: And then on top of that, AI adds very specific risks.
- 4:28SPEAKER_02: It adds risks around uh the potential for bias, right?
- 4:32SPEAKER_02: The fact that the outcomes aren't the outcomes that you're expecting, the uh decisions that are made that may potentially degrade the data quality.
- 4:40SPEAKER_02: And so a lot of what we build around uh AI governance needs to take that data integrity as a first principle.
- 4:47SPEAKER_01: I guess in the in the world of AI, which is sort of the general term um that gets used, there is this sort of shift now to agentic, uh, as we all know, where you're going from AI that you know used to just maybe be limited more to summarizing or um doing certain more basic tasks to more, much more autonomous AI, essentially that can uh take more independent action, as it were, um, to put it simply.
- 5:10SPEAKER_01: Um, what are the implications of that when we talk about issues of risk management and governance?
- 5:15SPEAKER_00: Sure.
- 5:15SPEAKER_00: So I I think there's an interesting point there in what we actually mean by governance, right?
- 5:20SPEAKER_00: Typically, and the way that we've looked at this historically is around the the systematized procedures, but also the the responsibility and accountability um with any particular management system, be that information security or resilience.
- 5:34SPEAKER_00: But in the AI space, we're now seeing that blending of uh accountability and responsibility across different teams, different uh roles within organizations.
- 5:43SPEAKER_00: And as an industry, it's not as easy anymore to be able to identify those individual accountable people within that decision-making chain.
- 5:52SPEAKER_02: That's absolutely right.
- 5:53SPEAKER_02: As we move from um AI that historically would have just summarized existing information to AI that can now take action on that information, I I would say that the four principles of governance remain the same, right?
- 6:06SPEAKER_02: Have we asked who is accountable for this decision?
- 6:11SPEAKER_02: Have we determined what is allowed within the context of that decision?
- 6:16SPEAKER_02: Uh, have we identified how that that decision is tested so that we know that it continues to be valid?
- 6:22SPEAKER_02: And have we laid out the protocols for what happens when it fails?
- 6:26SPEAKER_01: And I know um you mentioned data integrity, but uh at the start of the start of the session.
- 6:31SPEAKER_01: I mean, that seems to be a real key piece to me because if you have uh agents uh executing actions for you and the underlying data they're drawing upon is of a poor quality or is uh lacking integrity or has been compromised in some in some way, um certainly that would seem to uh jeopardize the the outputs of the agents themselves.
- 6:53SPEAKER_01: Um, where does that fit into the picture in terms of ensuring that data quality is really sufficient to guarantee that you're getting the results you need or intend to from agents or other AI capabilities you might utilize?
- 7:09SPEAKER_02: I think one of the really interesting things about this topic is the fact that AI governance doesn't operate in a vacuum, right?
- 7:15SPEAKER_02: And it pulls on principles that have been in place for a long time, right?
- 7:19SPEAKER_02: It pulls on data governance, it pulls on model risk management, it pulls on security and technical controls, the way that we can ensure that that data stays accurate, that stays, stays valid, stays used in the way that it's supposed to be used, is it's not a new concept, right?
- 7:37SPEAKER_02: And it it does lend itself to the hard controls.
- 7:40SPEAKER_00: Yeah.
- 7:40SPEAKER_00: And that that's a really interesting point, specifically if we're talking about model risk within certainly the financial uh sector.
- 7:47SPEAKER_00: We've seen for a long time that model risk has been um used as a method for um assuring that process, the inputs, the outputs, and then the testing and the validation of those processes.
- 7:59SPEAKER_00: So we we can draw upon a lot of the parallels there with regards to AI and the way that we implement those governance frameworks.
- 8:07SPEAKER_00: But back to that previous point from Casey, it's it's really important that we have those foundational uh controls in place as well.
- 8:14SPEAKER_00: Just because we're now working in a uh an AI uh native environment no longer means that any of those uh foundational requirements are no longer required.
- 8:22SPEAKER_00: They are as important as ever, and in many cases, more important in order to be able to gain that assurance and gain that uh trust in those uh ways of working and those processes, where those control points from a human are uh less involved or less in the loop.
- 8:37SPEAKER_02: I think what we're trying to say here is that there are other forms of governance that are table stakes, right?
- 8:41SPEAKER_02: What you do is you layer on top of those other forms of governance that additional layer of AI governance, right?
- 8:48SPEAKER_02: Um one of the interesting projects that we've been working on is how do we bring together some of these different um, I'm gonna use ISO terminology here, right?
- 8:57SPEAKER_02: The different management systems.
- 8:59SPEAKER_02: How do you consolidate them so that when you're building your risk management protocols, when you're building your uh governance frameworks, you're not duplicating across, say, data governance, security governance, data protection.
- 9:13SPEAKER_02: We do things like privacy impact assessments.
- 9:16SPEAKER_02: Should we also be doing AI impact assessments in the same context?
- 9:20SPEAKER_02: There's um a lot of uh commonalities you can find across all of those different governance frameworks.
- 9:26SPEAKER_02: Um, and there's a powerful result that you get when you start to think about those as different aspects of the same set of controls.
- 9:34SPEAKER_01: Yeah, and I suppose you know, thinking about this in in the context of uh the regulated financial services sector, uh, you know, I think from the perspective of, say, uh a MLRO or compliance officer introducing all this new fairly sophisticated technology into your business can seem quite essentially overwhelming, um, maybe scary in a way.
- 9:54SPEAKER_01: Um it sounds though like what you're saying is uh from a governance perspective, it's not all about completely reinventing the wheel or starting from scratch.
- 10:03SPEAKER_01: It's about drawing on some pre-existing principles, concepts, existing frameworks, things you already do, but adding something to it in many cases.
- 10:11SPEAKER_01: Is that is that a fair way to frame it?
- 10:14SPEAKER_00: I think that's fair.
- 10:15SPEAKER_00: New technology will always introduce new risks, and this is a new way of working.
- 10:20SPEAKER_00: We are seeing significant changes in the way that processes and operations work, but that doesn't necessarily mean that it increases the risk profile or decreases the risk profile.
- 10:30SPEAKER_00: It simply means that organizations such as EllipTech need to tackle those differently.
- 10:34SPEAKER_00: How do we gain the assurance?
- 10:36SPEAKER_00: How do we build the control framework to be able to scale securely and be able to speed up, be able to move more quickly with the use of uh AI tools whilst also having those same assurances where a human previously would have carried out a task?
- 10:52SPEAKER_00: Now maybe an AI, uh maybe an agent carries out that task.
- 10:57SPEAKER_00: Does it mean that the the quality of the output is better or worse?
- 11:01SPEAKER_00: That's how you measure your framework.
- 11:03SPEAKER_00: And it will be very, very different depending on the organization you are and the risk profile that your organization presents.
- 11:09SPEAKER_00: In Elliptic, we have a commitment to ensure that uh data quality of the services and the information that we provide is is high quality.
- 11:19SPEAKER_00: It's used to make incredibly important decisions.
- 11:21SPEAKER_00: We have to be able to assure that data and present without showing um exactly how those processes work, an assurance to our customers, which is where we've drawn upon the principles of model risk management uh in our governance framework to be able to uh demonstrate that end-to-end practice and working again from a governance and the uh the regulatory requirements and how they flow down into the technical controls within the organization that give that level of assurance and insight into the way that we operate and work so that organizations can gain trust in the quality of the information and the intel that we provide.