The next generation of credit risk management will be conversational.

There is a lot of discussion about how Artificial Intelligence will change banking, but perhaps we are asking the wrong question. Instead of asking how we can add AI to another banking application, we should ask what happens when AI fundamentally changes the way people interact with risk technology.

For decades, the relationship between a risk manager and a risk system has been remarkably one-directional. We load data, configure models, execute calculations, and then we wait. The system produces results and reports then somebody analyses those results and inevitably asks another question – Why did ECL increase? Which sectors caused the deterioration? What happens to RWA if the economy gets worse? What does this mean for our economic capital? What if we stop originating certain types of loans?

And the process starts again. New assumptions. New extracts. New calculations. New reports. In the age of AI, that suddenly feels rather old-fashioned.

What if You Could Talk to Your Balance Sheet

Imagine a CRO asking:

“What happens to our capital and provisions if GDP falls 2%, interest rates remain elevated and commercial real-estate values decline by 15%?

The bank’s AI understands the question, but understanding the question is the easy part. The difficult part is actually answering it.

Why conversational AI in banking depends on connected credit risk data

To answer the CRO’s question, the AI needs access to governed exposure data. It needs to understand the bank’s portfolio. It needs approved credit risk models, PD, LGD and EAD, IFRS 9 provisioning, Basel RWA and regulatory capital calculations. It also needs forecasting and stress-testing capabilities with sufficient computational power to run all of this quickly enough for the answer to remain relevant to the conversation.

Because these calculations cannot simply be performed independently. Credit risk is a chain.

Credit risk modelling → PiT PD/LGD/EAD → IFRS 9 ECL & provisions → Basel RWA & regulatory capital → Credit Stress Test, ECAP, ICAAP…

The outputs of one process become the inputs of another. That is where the real opportunity lies. The future isn’t simply AI + Credit Risk. It is AI connected to an integrated credit risk calculation, modelling, forecasting and simulation environment.

Now the conversation can continue: Which sectors are driving the increase? Assume we reduce new commercial real-estate lending by 20%. What happens to ECL? What is the resulting impact on Basel RWA and regulatory capital? What does that do to our economic capital? Compare all three scenarios over the next three years.

Suddenly, stress testing stops feeling like a quarterly project. It becomes a conversation.

Stress Testing Starts with Today

Every credible credit risk stress test or forecast needs to start with the bank’s actual credit risk position at T0. That means reproducing the regulatory RWA and IFRS 9 ECL provisions of the production environment before projecting them into the future.

This is necessary from a regulatory perspective, but it is also fundamental to understanding the results. If the starting position of the stress test does not reconcile with production, how can management distinguish the impact of the scenario from differences in data, models, methodology or calculation logic?

Only when T0 is consistent can we properly isolate what happens at T1, T2, T3 and beyond.

A deterioration in the economy can change credit states. Credit state determination and forecasting, whether performed top-down or bottom-up, affects PiT PDs and LGDs. Those changes flow into ECL and provisions. The resulting provisioning position flows through into regulatory calculations, including regulatory PDs, LGDs, RWA and capital. Ultimately, the impact propagates through the bank’s entire credit risk profile. A stress test isn’t another calculation. It is a scenario propagated through the entire credit-risk chain.

So, if stress testing starts with the outputs of the production credit-risk environment and uses many of the same data, models and calculations, there is an obvious question: Why does it so often live in a different system?

Why Credit Risk and Stress Testing Ended Up on Separate Systems

There are historical reasons. Stress testing was traditionally regarded as a specialist, periodic and often non-strategic capability. Dedicated teams developed their own tools, methodologies, spreadsheets and processes to satisfy specific regulatory exercises.

Meanwhile, production credit risk systems were built for a different purpose. They were designed to produce compliant regulatory numbers, every month, reliably. The question they were built to answer was: “What is the number?” They were generally not designed for credit risk management and simulation. They were not built to answer: “What if?”.

A production calculation may be executed once during a reporting cycle. A simulation environment may need to execute that calculation repeatedly across multiple scenarios, assumptions, portfolios and future time periods. The traditional technologies underpinning many credit-risk platforms were simply not designed for the processing speed and scalability that this requires.

Monthly reporting can tolerate batch processing. Interactive decision-making cannot.

AI Needs Something Intelligent to Talk to

A Large Language Model (LLM) can be extraordinarily good at understanding questions, reasoning about information and communicating answers, but AI cannot magically fix the architecture underneath it.

Agentic AI in banking needs an integrated architecture, not just an LLM

If a bank’s credit-risk environment consists of multiple data stores, separate IFRS 9 and Basel engines, independent model environments, stand-alone stress-testing tools, overnight batch processes, spreadsheets and manual reconciliations, connecting an AI to that landscape does not suddenly create an integrated risk platform. It creates an intelligent interface sitting on top of fragmented infrastructure.

Your AI can talk. Now give it something intelligent to talk to.

For AI-powered risk management to become genuinely useful, the infrastructure underneath the conversation needs to be fast, integrated, governed and inherently capable of simulation.

One Credit Risk Becomes Even More Important

This reinforces something we have believed at ElysianNxt for a long time. Credit risk should not be fragmented into technology silos simply because regulations, organisations and technology architectures evolved that way.

IFRS 9, Basel RWA, and credit risk modelling on one governed platform

These are different perspectives. But they are not independent. Same loans, same customers, same collateral, same models with same underlying credit risk. More importantly, their calculations depend on one another.

If the calculations depend on each other, why should the systems be independent? That is why integration is not simply an IT preference. It is how credit risk actually works.

Turning a Production Credit Risk Engine into a Simulation Engine

Once production and simulation operate on the same underlying architecture, an important boundary disappears. Today’s regulatory production position (today’s actual exposures, ECL, provisions, RWA, and capital) becomes tomorrow’s starting point.

Then change the assumptions: economic growth, interest rates, property prices, portfolio growth, credit strategy, sector concentration or climate pathways, and calculate again. See T1, T2, T3 – how credit states migrate, how PD and LGD evolve, how provisions change, how those changes affect RWA and regulatory capital, and ultimately, what they mean for the bank’s economic risk.

The credit risk system stops being purely a production engine. It becomes a simulation engine.

Stop Moving Data. Start Moving Assumptions.

Dynamic scenario modelling without moving data between systems

If all five perspectives operate on one governed data set, with models hosted and executed in the same ecosystem and calculations performed by native high-performance engines, something fundamental changes.

You no longer need to move enormous amounts of data between systems every time you want to ask another question. You change the assumptions instead. The underlying portfolio remains governed and consistent. The models remain controlled. The calculation chain remains intact. Only the scenario changes.

And once assumptions can be changed dynamically and calculations can be executed fast enough, AI becomes extraordinarily interesting, because natural language can become the interface to simulation.

The Bank Should Choose The AI

Banks should not have to adopt yet another AI environment simply because they adopt a risk platform. Financial institutions are already making strategic decisions about which AI technologies they trust, how those environments are governed and where their data may be processed. The risk platform should respect that choice.

Merlin and MCP: Connecting the bank’s own AI to governed risk data

This is the philosophy behind Merlin, the ElysianNxt AI agent. Through MCP connectivity, Merlin can communicate with the AI environment a bank has chosen and trusts, while ElysianNxt provides the governed credit risk intelligence underneath that conversation.

The bank’s AI provides the conversation. Merlin provides the connection and agency. The .NXT Platform provides the governed data, models, calculations, forecasting and simulations.

That means the AI doesn’t need to invent the answer. It asks the risk platform to calculate it.

For regulated financial institutions, that difference between an AI-generated answer and an AI-mediated, model-driven calculation is fundamental.

Moving Risk Teams from Static Reporting to Interactive Exploration

Perhaps the biggest change will not be technological. It will be behavioural. Risk technology has historically been designed around reporting: What happened? What is today’s ECL? What are today’s RWA? What is today’s capital position?

AI combined with native simulation changes the emphasis from reporting to exploration: What could happen? Why? What if we changed this? What if the economy changed that? What if this climate pathway materialised? What could management do about it? What would happen to ECL, RWA and economic capital?

The risk manager moves from being a consumer of calculated results to an active explorer of possible futures. That is a fundamentally different relationship with technology.

Credit Risk Becomes a Conversation

The ultimate promise of AI in risk management isn’t that it writes better reports. It isn’t even that it makes existing processes slightly more efficient. It is that the barrier between having a question and testing that question could largely disappear.

Ask, calculate, understand, change an assumption, simulate, compare, then decide. That requires AI. But it also requires something much less fashionable and arguably more important: a properly integrated credit risk architecture underneath it.

What conversational, agentic AI for credit risk actually requires

One governed data set. One model ecosystem – IFRS 9, Basel RWA, Credit Risk Modelling, Credit Risk Stress Testing & Forecasting, Climate Risk Stress Testing. Native calculations, native simulation, and an interactive stress-testing sandbox with enough speed and scalability to keep up with the conversation.

At ElysianNxt, that is where we believe credit risk technology is heading. Not another dashboard. Not another report. Not another AI chatbot sitting on top of disconnected systems. Something much more interesting. A conversation with your credit risk.