Moving from periodic, sample-based assurance to continuous, evidence-led challenge — with human judgement and accountability at the center.
On 31 July 2026, the EBA, EIOPA and ESMA issued a joint statement that many people in the industry probably missed. It flagged the ICT risks emerging from frontier AI models, including concentration risk behind a handful of providers, and called for “robust governance and risk management frameworks” to keep it in check.
It’s worth sitting with the question underneath that: how do you credibly assure a risk that keeps developing when most of your assurance still runs once a quarter?
Line 1 isn’t waiting around. More than 85% of large banks under European supervision already use AI, and generative and agentic systems are becoming routine, as ECB Banking Supervision’s Pedro Machado noted in February. In the UK, 75% of financial firms report using AI, and over half of those use cases involve some form of automated decision-making.
Here’s what that gap looks like in practice. A quarterly review signs off privileged access as “satisfactory.” Over the next twelve weeks, the organization processes thousands of access changes: joiners, leavers, role moves and temporary escalations. A sample of forty tells you almost nothing about the other several thousand. That assurance gap won't close on its own.
If anything, it’s getting wider on three fronts at once. Controls change by the day while assurance still runs by the quarter. Financial crime compliance alone cost EMEA institutions an estimated $85 billion in 2023, and 81% of firms say they would rather spend less on it. At the same time, the regulatory bar keeps rising: the EU AI Act’s high-risk obligations land in December 2027, DORA already requires independent Line 2 and Line 3 views of ICT risk, and the ECB has made AI governance a supervisory priority through 2028.
None of this means stripping out human oversight. It means using technology to give people better coverage, better evidence and more time to apply judgement. Here’s how that plays out in practice.
1. Move from sample testing to continuous control assurance
Sampling tells you what was true in the population you tested, not what’s true right now. The shift is to stop relying on small samples and start monitoring the whole population, using evidence already sitting in identity, cloud, SIEM, GRC and transaction systems to surface exceptions as they happen.
It helps to start narrow.
- Pick two or three high-volume controls, such as privileged access, change management or payments, where evidence already exists digitally.
- Connect the relevant IAM, SIEM, cloud or GRC data and define exactly which exceptions Line 2 wants surfaced.
- Set thresholds and escalation rules so people investigate the exceptions that matter instead of re-testing thousands of clean transactions by hand.
The same logic applies to conduct risk. Instead of reviewing a monthly sample of calls, natural language processing can assess all of them against defined standards. What changes is where people spend their time: less effort proving routine controls worked, more effort investigating the exceptions and patterns that actually need judgement.
2. Build a human–AI assurance model — and test the AI itself
AI is good at processing volume, spotting anomalies and flagging inconsistencies. People are still better at context, judgement and accountability. Used well, that combination extends what Line 2 can reach rather than replacing it — but only if Line 2 can challenge the tools it relies on.
That challenge is becoming harder to avoid as agentic systems spread: the Financial Stability Board has already pointed to architectures where AI systems monitor other AI systems.
A few things need to be in place before that works:
- Decide which assurance activities AI can perform and which decisions must stay with an accountable risk professional.
- Set minimum validation standards covering accuracy, drift, bias, explainability and false-positive rates before an assurance model is relied on.
- Run the same challenge process on Line 2’s own AI tools that Line 2 runs on Line 1’s.
A model does not need to be simple. It needs to be understood well enough that people know when to trust it and when to pull the plug — broadly, the direction of the AI Act’s requirements for human oversight and logging.
3. Turn regulatory change into a continuous process
Most organizations still treat regulatory change as a project. A rule appears, someone runs a gap assessment, a spreadsheet gets updated, and everyone moves on until the next one lands. AI can make that process continuous, mapping new and amended requirements across DORA, EBA guidance and the AI Act against controls, owners and evidence as they change rather than once a year.
- Keep a live map between obligations, controls, evidence and accountable owners rather than rebuilding it during annual reviews.
- Use AI to flag regulatory change and identify exactly which controls, policies and processes it touches.
- Require a named risk or compliance professional to sign off material mappings, and keep a record of what changed, why and who approved it.
EBA’s 2025 review reinforces the point. When it assessed the AI Act’s high-risk requirements against existing banking legislation, it found much of the ground was already covered. The model proposes, the professional decides, and that decision trail is exactly what a supervisor will want to see.
4. Bring third-party and ICT resilience assurance into the same model
The annual vendor questionnaire is getting harder to defend. AI can read contracts, assurance reports, certifications, penetration tests and incident records far faster than any team, and flag a missing obligation or shift in risk long before the next scheduled review.
- Continuously ingest supplier evidence: assurance reports, certifications, incidents, contract changes and sub-processor updates.
- Prioritize suppliers by criticality and changing risk signals rather than putting every vendor on the same calendar.
- Map the fourth- and fifth-party dependencies behind critical services and identify where one provider’s failure could cascade.
The ECB’s cloud outsourcing guidance already expects visibility into sub-outsourcing chains, and generative AI is deepening how much organizations rely on a small number of providers behind them. You cannot properly assure a dependency you cannot see.
5. Create an enterprise risk intelligence layer
The first four capabilities compound once they are connected. Right now, incidents, KRIs, control failures, audit findings, risk acceptances and near misses often sit in different systems, owned by different teams and reviewed separately. The patterns across them stay invisible as a result.
- Bring incidents, KRIs, control failures, audit findings, risk acceptances and near misses into one analytical view.
- Watch for a small number of patterns that actually matter: repeated overrides, deteriorating controls, supplier concentration and recurring exceptions across business units.
- Give executives insight they can act on: what is changing, why it matters, where risk is concentrating and what to do about it now.
The value here is not more dashboards. It is better context. That is what starts turning Line 2 from a function that reports what already happened into one that helps leadership see what is beginning to form.
The executive takeaway
AI can expand what Line 2 sees. Human judgement determines what it means.
None of this works if Line 2 adopts AI without holding it to the same standard it expects of Line 1: documented ownership, validation, explainability, monitoring and accountability. AI does not reduce accountability; it raises the bar on governance. And two things underpin all of it that you cannot simply buy off the shelf: well-governed data, and people who understand the technology well enough to challenge it.
