Risk · R-087 · Model Performance & Reliability
Model interdependency and aggregation
Models feed other models, so an error or change in one propagates, and the combined uncertainty across the chain is never assessed.
- Severity
- Medium
- Audience
- Heads-of · Practitioner
- Frameworks
- MAP 4 · Principle 1 · Principle 5
- Responsible AI domain
- Lifecycle stage
- Reviewed
- 2026-11-10
What it is
Models are rarely used alone. One model's score becomes another's input; several share a data source or a base model; a decision combines outputs from many. Each may be validated individually and found sound, while the chain as a whole has never been examined. Aggregate model risk is the exposure across all of them together, including the chance that they fail for the same reason at the same time.
How it happens
- A model consumes the output of another without its owner knowing.
- An upstream model is changed and downstream models are not retested.
- Several models are built on the same base model or the same flawed data.
- Uncertainty in each step is ignored when outputs are combined.
Why it matters
An error upstream is multiplied downstream. Common dependencies mean that models assumed to be independent checks on each other share the same blind spot. The organisation's total reliance on models is larger than any single validation suggests.
Warning signs
- The inventory does not record which models feed which.
- Changes are assessed for one model at a time.
- Many models depend on one supplier or one dataset.
- No report gives an aggregate view of model risk.
Mitigating controls
Primary controls
- AI inventory and registration Detective
- Records which models feed which. All AI systems and models, whether built, bought or embedded in other products, are recorded with owner, purpose, tier, data, supplier and status, and the record is reconciled periodically.
- Independent model validation Detective
- Validates the chain as well as each model. A function independent of the developers tests each model's conceptual soundness, data, performance and limitations before approval and periodically afterwards, to a depth set by the model's tier.
Supporting controls
- AI change management Preventive
- Assesses downstream effects before an upstream model changes. Changes to models, prompts, data sources, tools and configuration are assessed, tested, approved and versioned before they reach production.
Framework and clause references
| Framework | Clause | Title |
|---|---|---|
| NIST AI Risk Management Framework (AI RMF 1.0) | MAP 4 | Risks and benefits are mapped for all components, including third-party Risks mapped across components. |
| SS1/23: Model Risk Management Principles for Banks | Principle 1 | Model identification and model risk classification The inventory records interdependencies. |
| SS1/23: Model Risk Management Principles for Banks | Principle 5 | Model risk mitigants Aggregate model risk. |
Sources
- SS1/23: Model Risk Management Principles for Banks Bank of England PRA · official text
- NIST AI Risk Management Framework (AI RMF 1.0) NIST · framework
Related terms
How to cite
Axiom Verity (2026). Model interdependency and aggregation. Reference Library, R-087. https://axiom-verity.com/risks/model-interdependency-and-aggregation. Reviewed 2026-11-10.
Found an error or a better source? Report a correction.