Latest / Anaiya Algorithm / DORA in Practice — What Regulated Financial Services Organisations Need to Do Now
Transcript
- 0:00SPEAKER_00: I want you to picture a moment that is playing out in financial services organizations right now.
- 0:07SPEAKER_00: A regulator sends a request, not a formal enforcement action, just a request.
- 0:15SPEAKER_00: Can you show us your AI inventory?
- 0:18SPEAKER_00: Would like to understand how you're managing your AI related ICT risk.
- 0:25SPEAKER_00: And the person who receives that request has to make a call.
- 0:29SPEAKER_00: Do we have that?
- 0:30SPEAKER_00: Is it current?
- 0:31SPEAKER_00: Is it complete enough to share?
- 0:34SPEAKER_00: In some organizations, the answer is yes.
- 0:37SPEAKER_00: They can pull it out.
- 0:39SPEAKER_00: Everything is documented, monitored, evidenced.
- 0:43SPEAKER_00: The conversation with the regulator is a review.
- 0:47SPEAKER_00: In others, and I want to be honest, this is most.
- 0:52SPEAKER_00: The answer is closer to we have something, but let us have a few weeks to pull it together properly.
- 1:00SPEAKER_00: That difference, that gap between we can show you and give us time to build it.
- 1:07SPEAKER_00: That is what this episode is about.
- 1:33SPEAKER_00: And today we're talking about DORA, the Digital Operational Resilience Act, what it actually requires of you and what AI has to do with it.
- 1:46SPEAKER_00: DORA came into full force in january twenty twenty five.
- 1:51SPEAKER_00: Unlike the EU AI Act, which is phasing in gradually, DORA is not phasing in.
- 1:59SPEAKER_00: It is here.
- 2:01SPEAKER_00: If you are a regulated financial services entity operating in or serving the EU, you're already in scope.
- 2:10SPEAKER_00: Banks, insurers, investment firms, paying institutions, trading venues, and critically the ICT, third party providers, all of those entities depend on.
- 2:23SPEAKER_00: The purpose of DORA is simple in principle.
- 2:26SPEAKER_00: Ensure the financial sector can withstand disruptions to its digital infrastructure.
- 2:32SPEAKER_00: Regulators looked at a world where banks run on software, where decisions are made by algorithms, where the entire operational fabric of finance is digital.
- 2:45SPEAKER_00: And they said we need a resilience framework for that world.
- 2:50SPEAKER_00: And here is the thing that trips people up.
- 2:53SPEAKER_00: Dora doesn't specifically call out AI.
- 2:57SPEAKER_00: It talks about ICT.
- 2:59SPEAKER_00: But AI is ICT.
- 3:02SPEAKER_00: Your fraud detection model, your credit risk engine, your customer decisioning system, these are ICT infrastructure.
- 3:12SPEAKER_00: And they are in scope.
- 3:15SPEAKER_00: Not because someone added an AI clause, but because AI is the infrastructure.
- 3:23SPEAKER_00: So when Dora says you need to manage, test and evidence the resilience of your ICT systems, it means those two.
- 3:39SPEAKER_00: It asks whether you can prove you already are.
- 3:44SPEAKER_00: And that distinction matters because planning is easy, however, proving is a different discipline entirely.
- 4:14SPEAKER_00: The first is risk management.
- 4:17SPEAKER_00: You need a documented living framework for identifying and managing ICT risk, updated at least annually and updated whenever something significant changes in your ICT environment.
- 4:34SPEAKER_00: Every time you deploy a new AI system, that is a change, which means your risk framework needs to know about it, classify it and track it.
- 4:46SPEAKER_00: A lot of organizations have risk frameworks that predate their AI deployment.
- 4:52SPEAKER_00: The framework exists, the AI isn't in them.
- 4:58SPEAKER_00: The second is incident reporting, and this is the one I want you to sit with for a moment.
- 5:06SPEAKER_00: So if a major ICT related incident occurs, Dora sets very specific timelines.
- 5:14SPEAKER_00: Initial notification within four hours of classifying the incident, intermediate report within seventy two hours, final report within a month.
- 5:27SPEAKER_00: Now, what counts as a major incident?
- 5:30SPEAKER_00: Think about what that means for an AI system.
- 5:35SPEAKER_00: If your fraud model starts misclassifying a scale, if your credit decision engine produces systematically wrong outputs, if your risk monitoring platform goes dark during a critical period, these are potential Dora reportable incidents.
- 5:56SPEAKER_00: And the question is not just whether they happen.
- 6:00SPEAKER_00: The question is whether you have the monitoring to detect them fast enough to meet those timelines.
- 6:08SPEAKER_00: Four hours is not enough when you're scrambling to understand what's wrong.
- 6:14SPEAKER_00: Third, third is resilience testing.
- 6:19SPEAKER_00: Your AI systems need to be tested, not just the infrastructure they run on, but their behavior under stress.
- 6:28SPEAKER_00: For significant institutions, DORA requires threat led penetration testing every three years.
- 6:36SPEAKER_00: Your AI isn't exempt.
- 6:39SPEAKER_00: Now let's look at the fourth.
- 6:41SPEAKER_00: And this catches people by surprise.
- 6:43SPEAKER_00: Third party risk.
- 6:46SPEAKER_00: The AI vendors you depend on sit inside your DORA obligations.
- 6:51SPEAKER_00: If your primary model provider experiences an outage, what happens to you?
- 6:57SPEAKER_00: That needs to be documented.
- 6:59SPEAKER_00: Exit strategies need to be real, not theoretical.
- 7:04SPEAKER_00: The vendor you rely on most is the one that creates the most concentration risk, and Dora requires you to manage that explicitly.
- 7:15SPEAKER_00: And fifth, intelligence sharing.
- 7:19SPEAKER_00: There's a framework for sharing threat intelligence across financial entities.
- 7:24SPEAKER_00: Regulators expect significant institutions to be active in it.
- 7:31SPEAKER_00: So five areas real obligations and AI touches all of them.
- 7:38SPEAKER_00: Here is something I genuinely didn't expect when I started working on this.
- 7:44SPEAKER_00: How much DORA and the EU AI Act overlap in what they ask for?
- 7:51SPEAKER_00: So take a credit scoring AI used with EU customers.
- 7:56SPEAKER_00: Under the EU AI Act, it's high risk.
- 8:00SPEAKER_00: It needs a conformity assessment, human oversight mechanisms, technical documentation, continuous monitoring.
- 8:10SPEAKER_00: Under Dora, it sits inside your ICT infrastructure, it needs risk classification, resilience testing, incident reporting capability.
- 8:22SPEAKER_00: Under GDPR, it processes personal data with legal obligations attached, so one system, three frameworks, three sets of evidence requirements.
- 8:34SPEAKER_00: The organizations I see struggling are the ones running these as separate work streams.
- 8:42SPEAKER_00: A compliance team doing EUAI Act, a risk team doing DORA, no one joining them.
- 8:52SPEAKER_00: And the result is duplication, gaps and probably most importantly, two evidence trails that don't tell the same story.
- 9:04SPEAKER_00: Because here's what's true.
- 9:06SPEAKER_00: The technical documentation your EU AI Act conformity assessment requires overlaps heavily with the risk framework documentation DORA needs.
- 9:20SPEAKER_00: Your continuous monitoring serves both the Article fifteen performance requirements under the AI Act and the incident detection obligations under Dora.
- 9:32SPEAKER_00: These are not separate exercises.
- 9:35SPEAKER_00: They are the same exercise done once, coherently or twice, expensively and inconsistently.
- 9:46SPEAKER_00: And there's another thing.
- 9:48SPEAKER_00: The regulators are talking to each other.
- 9:51SPEAKER_00: The ECB, the EBA, the FCA, the national competent authorities responsible for AI Act enforcement.
- 10:01SPEAKER_00: These conversations are happening.
- 10:05SPEAKER_00: If you have a DORA compliance team and an EU AI Act compliance team and they're not working from the same evidence trail, you are creating gaps that will be visible to the people asking questions, even if they are not visible to you.
- 10:24SPEAKER_00: Let me get specific about what your evidence trail needs to look like.
- 10:29SPEAKER_00: Your ICT risk register, which Dorat requires, needs to include your AI systems, not the ones IT formally manages, all of them, including the ones that were deployed by business teams without a formal IT project, including the ones embedded in third party platforms, including the ones that have drifted out of the original specification since they were first deployed.
- 11:02SPEAKER_00: Your incident classification criteria need to include AI specific failure modes, model drift, output degradation, data quality deterioration.
- 11:16SPEAKER_00: These aren't the same as a server going down.
- 11:20SPEAKER_00: They are often gradual, and gradual is harder to catch.
- 11:26SPEAKER_00: If your fraud model's accuracy drops fifteen points over six weeks, that's a risk evident.
- 11:35SPEAKER_00: The question is whether your monitoring catches it at week two or whether you find out because a regulator flucks it first.
- 11:46SPEAKER_00: And this is important.
- 11:48SPEAKER_00: Your resilience testing needs to cover AI behavior, not just infrastructure.
- 11:54SPEAKER_00: It's not enough to test that the server stays up.
- 11:58SPEAKER_00: You need to know how the model performs underload under unusual data conditions, under adversarial inputs.
- 12:08SPEAKER_00: The organizations that are in a strong position on this are the ones that have built monitoring as infrastructure rather than as a project.
- 12:20SPEAKER_00: Continuous drift detection, performance alerting, automatic evidence generation.
- 12:28SPEAKER_00: Not because it's elegant, because the regulation asks you to demonstrate resilience in real time, and that requires real time infrastructure.
- 12:41SPEAKER_00: Periodic audits and manual review are not enough.
- 12:46SPEAKER_00: The timelines DORA requires for incident reporting make that clear.
- 12:52SPEAKER_00: You cannot afford to find out what your AI systems are doing by reading a quarterly report.
- 13:00SPEAKER_00: This is exactly the problem Veridian was built to solve.
- 13:04SPEAKER_00: When your organization is subject to DORA, the EU AI Act, GDPR, the Cyber Resilience Act, ISO forty two thousand and one, and you're trying to manage compliance across all of them, the question isn't which framework to prioritize.
- 13:22SPEAKER_00: The question is how to build governance infrastructure that serves all of them from a single continuous evidence trail.
- 13:32SPEAKER_00: Veridian classifies every AI system against all six frameworks simultaneously.
- 13:39SPEAKER_00: Your RAI inventory becomes your ICT risk register contribution.
- 13:46SPEAKER_00: The continuous monitoring generates the performance and availability data that supports both incident detection under DORA and the Article fifteen monitoring requirements under the EU AI Act and the immutable audit trail.
- 14:05SPEAKER_00: The evidence your regulator will ask to see is building from the moment each system is registered.
- 14:14SPEAKER_00: When a DORA reportable incident does occur, the documentation that supports your initial four hour notification, your seventy-two hour intermediate report and your final report has been accumulating continuously.
- 14:34SPEAKER_00: You are not assembling a timeline under pressure.
- 14:38SPEAKER_00: You are retrieving one.
- 14:41SPEAKER_00: Viridian is on Microsoft Azure Marketplace and the link is in the show notes.
- 14:48SPEAKER_00: Dora is live.
- 14:50SPEAKER_00: Your AI systems are inside its scope.
- 14:54SPEAKER_00: Not as a special case, but because AI is your operational infrastructure now.
- 15:02SPEAKER_00: The question isn't whether you're in scope because you are.
- 15:06SPEAKER_00: The question is whether you can demonstrate it when someone asks and someone will ask.
- 15:15SPEAKER_00: Next month, episode eight, we go into Agentic AI.
- 15:21SPEAKER_00: Because as AI moves from producing outputs to taking actions, the governance questions change completely.
- 15:30SPEAKER_00: If you are deploying co pilot agents, autonomous workflow tools or any AI that acts rather than advice, that episode is for you.
- 15:44SPEAKER_00: Until then, keep leading intentionally.
- 15:48SPEAKER_00: Thank you.