Why RDARR is the foundation for safe agentic AI in risk management

Agentic AI is rapidly becoming one of the most discussed topics in banking technology. Unlike traditional dashboards or analytics tools, AI agents can interpret a question, decide which data to retrieve, perform multi-step analysis, generate explanations, and recommend next actions. For risk functions, the promise is compelling: faster investigations, more responsive reporting, and better support for senior decision-makers during periods of uncertainty.

But in risk management, intelligence is not enough. AI agents cannot be trusted unless the data they use is accurate, complete, timely, traceable and governed. That is why RDARR (Risk Data Aggregation and Risk Reporting), underpinned by BCBS 239 principles, is becoming more than a regulatory obligation. It is becoming the control layer for safe, scalable AI in risk.

Why does agentic AI raise the stakes for RDARR?

Traditional risk reporting is often slow, manual, and fragmented. A senior manager asks why market risk increased, why a limit was breached, or what the bank’s exposure is to a specific counterparty, sector, country, or event. Teams then extract data from multiple systems, reconcile numbers, apply adjustments, validate outputs, and prepare commentary. Depending on the complexity, this process can take hours, days, or even weeks.

Agentic AI is changing this operating model. A risk analyst could ask:

• What drove the increase in VaR today?
• What is our exposure to this counterparty group across all legal entities?
• What would happen to liquidity if wholesale funding markets closed for five days?
• Which data quality issue has the largest impact on the board risk pack?
• Can we explain the difference between finance and risk exposure numbers?

An AI agent could, in principle, investigate the question, identify the relevant data, run aggregations, compare results against previous periods, detect anomalies, prepare drill-downs, and draft a management explanation. This is powerful, but also risky if data is inaccurate.

If the agent queries incomplete data, applies the wrong hierarchy, uses stale market data, ignores manual adjustments, or cannot explain how a figure was produced, it may generate a confident but misleading answer. In risk management, that is unacceptable. The problem is not hallucinations. The bigger issue is reasoning over uncontrolled data.

How does RDARR become the control framework for AI in risk?

BCBS 239 was not written for generative AI or autonomous agents. It was introduced to strengthen banks’ ability to aggregate risk data and produce reliable risk reports, particularly under stress. But its core principles map directly onto the conditions required for trustworthy AI-enabled risk management.

Accuracy and integrity matter because AI agents must work from reliable numbers. If exposure data is inconsistent, duplicated, or incorrectly mapped, the agent’s output will be wrong regardless of how sophisticated the model is.

Completeness matters because AI agents must understand the full perimeter of risk. A partial view of trades, counterparties, legal entities, portfolios, collateral or market data can lead to underestimation of exposure.

Timeliness matters because many risk decisions are time-sensitive. During a market shock, counterparty event, liquidity stress, or supervisory request, stale data can be as dangerous as incorrect data.

Adaptability matters because risk questions are rarely predictable. Senior management and supervisors often ask new questions during stress. Banks need the ability to produce ad hoc analysis quickly and with evidence.

Traceability matters because every AI-generated answer must be explainable. A risk manager must be able to ask: where did this number come from, which adjustments were applied, what assumptions were used, and who approved the result?

Governance matters because AI does not remove accountability. The board, senior management, risk owners, data owners, and control functions remain responsible for the use of risk data and the decisions made from it.

In this sense, RDARR is no longer only a compliance topic. It is becoming the trust layer for agentic AI in banking risk functions.

From regulatory reporting to risk command centers

Many banks still approach RDARR through the lens of compliance: policies, controls, data dictionaries, lineage documentation, remediation plans, and reporting attestations. These remain necessary but are not sufficient.

Banks need risk data environments where users can move from a high-level indicator to granular evidence in seconds. They need to aggregate exposures across dimensions, drill down from group-level metrics to individual trades, compare versions, test scenarios, identify data quality issues, and reproduce the same result later. This is the idea of a risk command center.

A risk command center is not just a dashboard. A dashboard shows predefined metrics. A command center enables controlled investigation. It lets risk teams ask new questions, explore emerging events, identify concentrations, explain movements, and prepare evidence for management, audit, or supervisors.

Agentic AI makes this concept even more important. The future risk user experience will not be limited to static reports. It will become conversational, investigative, and interactive. Users will ask questions in natural language. AI agents will propose analyses, surface anomalies, and guide drill-downs. But behind that conversational layer, the bank still needs a controlled data foundation. Without that foundation, agentic AI becomes an attractive interface over an unreliable core.

The biggest blocker is not the model

Many AI discussions in banking focus on model selection, prompt engineering, LLM performance, and deployment architecture. These are important topics, but are not the main blocker for risk functions. The main blocker is trusted risk data.

Risk data is complex. It is granular, high-volume, multi-dimensional, and constantly changing. It spans trades, positions, market data, reference data, collateral, limits, sensitivities, cash flows, scenarios, accounting views, legal entities, and reporting hierarchies. It is produced by multiple systems and consumed by multiple teams. It often requires reconciliation between front office, risk, finance, treasury, and regulatory reporting.

An AI agent cannot solve this complexity by itself. In fact, agentic AI will expose weaknesses that may previously have been hidden by manual processes. If different systems produce different answers, the agent will surface the inconsistency. If lineage is incomplete, the agent will struggle to justify the output. If data quality issues are resolved manually in spreadsheets, the agent may not know which version of the number is official. If hierarchies differ across risk and finance, the agent may produce conflicting views.

This is why banks should avoid treating agentic AI as a shortcut around RDARR challenges. Agentic AI does not replace RDARR. It increases the need for it.

What is BCBS 239 Principle 6, and why does it matter for agentic AI?

Among the BCBS 239 principles, adaptability is especially relevant to agentic AI. BCBS 239 Principle 6 requires banks to generate aggregate risk data to meet a broad range of on-demand and ad hoc reporting requests. This includes requests during stress, crisis situations, changing internal needs, and supervisory queries.

This is precisely where AI agents could create major value. During a crisis, the questions are not neatly predefined. A geopolitical event, bank failure, market dislocation, cyber incident, sanctions announcement, or counterparty default may trigger urgent questions from the CRO, board, regulator, treasury, or front office.

The bank may need to know:

• Which exposures are affected?
• Which legal entities are involved?
• What is the impact by desk, country, sector, counterparty, product, and maturity?
• Are limits breached?
• What collateral is available?
• What is the liquidity impact?
• Which numbers are final, estimated or subject to data quality limitations?
• What assumptions were used?
• Can the analysis be refreshed in two hours?

This is the natural environment for agentic AI. But it is also the natural test of RDARR maturity. A well-controlled AI agent could help interpret the request, retrieve relevant data, run aggregations, identify concentrations, check data quality, compare results with prior versions, and draft an incident report. Yet the answer must remain traceable, reproducible, permissioned, and governed.

In a supervisory context, “the AI said so” will never be an acceptable explanation.

The new operating model: human-in-the-loop, evidence-by-design

The realistic future is not a fully autonomous CRO. The near-term opportunity is a human-in-the-loop model where AI agents accelerate risk work while humans retain accountability.

In this model, agents perform investigation, preparation, comparison, summarisation, and first-level explanation. Human experts validate the analysis, challenge assumptions, approve outputs, and make decisions. For this to work, evidence must be built into the workflow.

Every AI-generated risk answer should be linked to source data, calculation logic, timestamps, assumptions, adjustments, data quality indicators, and approval status. Users should be able to drill down from narrative to number, from number to aggregation, and from aggregation to underlying records.

This is where the connection between agentic AI and RDARR becomes concrete. BCBS 239 gives banks the principles. Agentic AI gives banks a new interface and operating model. The missing link is the controlled risk data platform that connects the two.

What banks should do now

Banks exploring agentic AI in risk should start with a simple question:

Can we trust the data that the agent will use?

If the answer is uncertain, the priority should not be to deploy more AI pilots. The priority should be to strengthen the risk data foundation.

That means creating governed, granular, and reusable risk data products. It means improving lineage, data quality controls, reconciliation, versioning, access management, and auditability. It means reducing dependence on manual extracts and spreadsheets. It means making risk numbers explainable by design.

The banks that succeed with agentic AI will not simply be those with the best models. They will be those with the most trusted and accessible risk data.

Why Opensee and Agensee are the solution

Making RDARR operational in the agentic AI era requires two complementary capabilities: a trusted risk data foundation and a controlled agentic layer on top of it.

Opensee provides the risk data foundation. It helps financial institutions bring large, complex, and fragmented risk datasets into a high-performance analytics layer where data can be aggregated, drilled into, reconciled, and explained. For RDARR, this matters because risk teams need more than static reports. They need to understand where numbers come from, how they were calculated, what changed, and whether the underlying data is complete and reliable.

Agensee provides the agentic AI layer. It can sit on top of governed risk data to help users investigate questions, explore drivers, identify anomalies, prepare explanations, and support ad hoc analysis. In this model, the agent does not replace the risk analyst or bypass governance. It helps accelerate investigation while keeping humans in control of interpretation, validation, and decision-making.

Together, Opensee and Agensee illustrate how banks can move from manual, report-centric RDARR processes toward a more interactive and evidence-based operating model. Opensee makes the risk data accessible, granular, and traceable. Agensee makes it easier for users to question, investigate, and explain that data through an agentic AI interface.

This distinction is important. Agentic AI in risk should not be treated as a standalone capability. It needs to be grounded in governed data, controlled workflows, and transparent evidence. Otherwise, it risks becoming another layer of complexity over already fragmented systems.

In summary, RDARR defines the operating discipline for trusted risk data. BCBS 239 provides the underlying principles. Opensee supports the data foundation. Agensee brings the agentic AI interaction layer. Together, they help make risk aggregation, investigation, and explanation more operational, controlled, and scalable.