Agentic AI set to drive end-to-end automation of MRM workflows

Agentic artificial intelligence is opening the door to end-to-end (E2E) automation of model validation processes, according to industry experts. At a fireside chat at the Risk Live Europe event on July 1, the view was that, while traditional automation may have already streamlined certain elements of validation workflows, the addition of agentic AI enables banks to reimagine the whole lifecycle

Several factors are converging to drive interest in automation and AI for model risk management (MRM) processes.

“The demands placed on internal validation functions have increased significantly,” explained Carlos Diaz, associate partner at McKinsey & Co. “The scope of the models covered by MRM has expanded – it’s not just about traditional risk models, but also business applications and non-model [analytical tools]. Regulatory expectations have also increased, requiring more in-depth analysis of models. Meanwhile, banks are under cost pressure, increasing the drive to do more with the same resources. Finally, AI itself has introduced new risks, such as hallucinations and bias, which need to be factored into processes.”

Against these growing demands, many banks have invested heavily in traditional automation to assist validators with their workloads. Dorothee Ancira, head of internal validation, Santander UK, detailed the bank’s transformation journey during the past 18–24 months.

“We started by looking to automate the more repeatable tasks of the process, such as gathering the information required for validation, to create more capacity for value outputs,” she said. “We now have pipelines to the different model taxonomies, which enable us to create a first report draft detailing the data to be examined and the tests to run. This has provided space for the validator to step back and think more about the business and regulatory context and challenge assumptions more effectively.”

Driving efficiencies

Adopting traditional automation within workflows delivers undoubted efficiencies. Diaz noted that many banks that have taken this path report 20–30% faster validation turnarounds. However, the benefits extend far beyond efficiencies realised. “Automation increases consistency and exhaustivity in testing, which improves the quality and standardisation of output,” he added.

For repeatable processes, standard deterministic automation will continue to play a crucial role. However, there are areas where AI can add considerable value on top of this, which many banks are exploring in depth.

“We’re starting to think about how we could embed AI into the validation process,” said Ancira. “Where could AI support the validator in the process? We’re also questioning how we are going to make ourselves comfortable as a validation function that AI really is adding value and that we can confidently rely on the outputs.”

Rewiring validation for efficiency

The introduction of AI has considerable implications for how validation functions operate. Although banks may start by introducing AI for specific tasks, such as helping validators navigate code libraries and other tools, ultimately its power lies in combining it with traditional automation approaches to enable complete E2E workflow automation with validator oversight and expert judgement.

“We see [rewiring] this as a five-step reimagined process,” said Diaz. “This starts with the configuration and planning phases where agents use standardised taxonomies to come up with an initial plan for the particular requirements of a given model.

“Once the validator has approved the plan, it moves to the pipeline phase, using agents to retrieve and configure testing components from libraries and access additional information from business reports to enrich understanding or propose additional tests. It can then move to the execution phase, which may include standard automation, but where AI can add value by performing tasks such as generating first drafts of documents, inserting figures and creating text.

“Finally, there is the document review phase, where the validator will review and approve the report, but agents can also assist with quality checks and compliance with standards.”

Throughout this reimagined workflow, the role of the validator remains vital. Humans will move from executing workflows to supervising intelligent, multi-agent systems. Diaz noted that “Agents will not replace validators, but free up their time to focus on where it matters. Sometimes, when we talk about efficiencies, people think this means reducing team sizes. But I don’t think this will be the case, given the increased demands on validation functions.”

Technological transformation is not, however, sustainable without transforming the underlying processes and investing in people, so they have the necessary skills to be proficient in this new working world.

“Validation teams need to be upskilled, so they are fluent in all the new technologies. ‘Keep on learning’ is also a key mantra, because the pace of change is unprecedented. Ultimately, a human will be accountable for the end outputs, so you need to overlay agent activity with human expertise,” commented Ancira.

Additionally, validation teams need to have confidence in the AI tools used to support processes. “Before you start using new tools, they need to undergo specific predevelopment testing. Continuous monitoring also needs to be in place, so there is live feedback on the accuracy of the process and details on where it might be failing,” stated Diaz.

A virtuous circle

Ancira, meanwhile, highlighted the importance of using different AI tools to avoid getting caught in the loop of validation tools validating their own model creations. “We have to be careful that we’re not using exactly the same tools everywhere. We need to ensure that, when you have all these agents, you configure them so you still have an effective second line of defence. In the same way that you wouldn’t have the same analyst doing first- and second-line work.”

While there are undoubtedly challenges to work through, embedding AI into MRM processes also introduces new possibilities for model development. “Agents will learn from the team’s interactions, as well as interactions between the model owners, users, developers, validators and perhaps even with the regulator. They could then actually become advisers to the development process and bring a different lens, which is really quite exciting,” concluded Ancira.