The Market Risk Explainability Gap: Why AI and Solvency II Are on a Collision Course
The Specific Problem With Market Risk
Not all AI compliance questions are equal under Solvency II.
The EU AI Act's high-risk classifications — AI in life and health insurance pricing, credit scoring, emotional recognition — get most of the press. But for teams running market risk functions inside insurance companies and pension funds, the more immediate compliance tension comes from somewhere else entirely: the structural mismatch between how Solvency II's market risk framework expects models to behave, and how AI-assisted analysis actually works.
This piece is specifically about that mismatch — and what it means for the SCR market risk modules that RemitRix and similar platforms operate within.
How the Standard Formula Market Risk Framework Works
To understand the gap, you need to start with how Solvency II's market risk SCR is designed.
The standard formula breaks market risk into six sub-modules: interest rate risk, equity risk, spread risk, currency risk, concentration risk, and property risk. Each module applies defined stress scenarios to the balance sheet — parallel yield curve shifts, equity shocks, credit spread widening by rating bucket and duration, FX moves — and calculates the resulting change in net asset value. The modules are then aggregated using a prescribed correlation matrix to produce the total market risk SCR.
The critical feature of this architecture is determinism. Given the same portfolio and the same market parameters, the standard formula always produces the same output. The calculation path is fully traceable. Every input, every stress parameter, every aggregation step can be documented and audited. The current reporting framework relies on XBRL taxonomy with clearly defined Quantitative Reporting Templates — QRTs — that are structured, auditable, and directly tied to the capital calculation logic.
This determinism is not just a technical property — it's a regulatory expectation. Solvency II's governance requirements around the risk management system, ORSA, and model documentation all assume that outputs can be explained by tracing back through defined inputs and logic.
Where AI Breaks the Assumption
AI-assisted market risk analysis doesn't work this way, and the gap is real, not theoretical.
Consider three concrete use cases that market risk teams at insurance companies are actively deploying or evaluating:
1. AI-generated economic scenario narration
An AI system ingests current market conditions and generates a narrative explanation of the most material stress scenarios for a given portfolio — "your primary risk is duration mismatch in the liability-driven portfolio, amplified by current spread levels in BBB-rated corporates." The insight may be accurate. But what is the model? What are the parameters? How is consistency across time periods validated? Under Solvency II's ORSA requirements, the scenario assumptions and their derivation need to be documented and defensible. An LLM narrative is neither.
2. Natural language SCR attribution
A risk manager asks: "Explain the drivers of our market risk SCR increase quarter-on-quarter." An AI system reasons over position data, market movements, and module calculations to produce a three-driver explanation. The answer may perfectly match the underlying standard formula output, but the reasoning path is probabilistic, not deterministic. The same question asked twice may produce a differently worded explanation. Under what governance framework is this output used? Who validates it?
3. Concentration risk detection
AI identifies clustering patterns in credit exposure that the standard concentration risk module doesn't capture — cross-issuer concentration by sector, geographic cluster, or correlated names below the individual issuer threshold. This is genuinely valuable analysis. But it operates outside the standard formula boundary. How does it feed into the ORSA? How is it documented in the risk management system? What's the model governance process?
In each case, the AI output is potentially useful — and potentially ungoverned.
The Regulatory Timeline Bearing Down
For insurers, 2026 represents the most operationally intensive phase of the Solvency II review — firms are expected to progress balance-sheet modelling, recalibrate capital requirements and implement governance and risk-management changes ahead of application in 2027.
Directive 2025/2 updates the Solvency II framework to enhance proportionality and integrate new macro-prudential tools, with Member States required to transpose by 30 January 2027. Among the substantive changes: tougher spread-risk stress tests, new extrapolation formulas for long-term liabilities, revised risk margin methods, and updated catastrophe risk calibration.
The recalibration of spread risk stress tests is particularly relevant. As market risk teams rebuild their SCR models to incorporate updated parameters, they face a simultaneous pressure: do they embed AI tooling in that rebuild? If yes, what governance framework governs the AI layer?
EIOPA updated its supervisory review guidelines in February 2026 to incorporate IT and cyber risks and SupTech into supervisory processes — the first explicit signal that AI-adjacent tooling is entering the supervisory lens. This isn't the full AI governance framework yet. But it's a directional statement that national competent authorities are now expected to consider technology risks, including model risks from AI systems, as part of standard supervisory review.
What a Governed AI Market Risk Layer Looks Like
Closing the explainability gap doesn't mean avoiding AI in market risk. It means building the governance architecture that makes AI-assisted outputs supervisable.
The practical requirements:
Anchored outputs. Every AI-generated market risk insight needs to trace back to the underlying data query and calculation. When the system says "spread risk is your top SCR driver," it should be able to show: which positions, which spread levels, which duration buckets, what stress parameters — in a form that maps directly to the standard formula module structure.
Consistency validation. AI outputs used in risk decisions need to be tested for consistency across time periods and parameter sets. This is model validation, applied to an LLM-assisted layer. It requires logging inputs and outputs, tracking variance, and defining acceptable ranges.
Governance boundaries. Not every AI output is a model output in the Solvency II sense. But some are — particularly where AI-generated scenarios feed into ORSA documentation or management decisions. The governance framework needs to define clearly where the AI layer sits: is it a decision-support tool (lower governance burden) or an input to a regulated calculation (higher burden)?
Audit trails. For every AI-assisted analysis that influences a capital decision, there needs to be a record: what was asked, what data was used, what output was produced. Not a summary. A trace.
At RemitRix, these four requirements shape the architecture of every feature we build. When a risk manager asks the platform to explain SCR concentration risk, the response is anchored to the position-level data, the module calculation, and the regulatory parameters — not generated from general reasoning. The AI layer explains; the calculation layer proves.
The Competitive Advantage Is the Same as the Compliance Requirement
Here's what makes this interesting from a product perspective: the governance requirements for AI in Solvency II market risk and the product requirements for a useful AI market risk tool are almost identical.
A risk manager who can trust an AI-generated SCR attribution — because they can see the underlying data and verify the calculation — uses it more. Asks harder questions. Makes faster decisions. The explainability isn't just compliance overhead; it's the feature that makes the tool valuable.
The teams that build this layer first — governed, explainable, supervisable AI for market risk — won't just satisfy their next EIOPA review. They'll have a meaningfully faster risk intelligence cycle than teams still navigating between spreadsheets and static dashboards.
That's the gap worth closing. And the direction the market is moving.
Effi Mor is the founder of RemitRix, a scenario-based risk intelligence platform focused on Solvency II market risk, covering SCR modules, economic scenario generation, and AI-assisted portfolio stress testing for insurance companies and pension funds. Risk Intelligence Weekly publishes every Wednesday.