What counts as a model in model risk management?
By Nick Jang, founder of MINERVA Studio ·
The short answer
Under OSFI’s Guideline E-23, a model is the application of theoretical, empirical or judgmental assumptions or statistical techniques, AI and machine learning included, that processes input data to produce results. In practice: if something takes data, applies a method that involves assumptions or estimation, and produces an output people rely on to decide, it is a model, whether it lives in Python, a vendor product or a spreadsheet.
The definition, in parts
Strip the wording down and you get three pieces. Data goes in. Something turns it into an estimate. The estimate comes out in a form a person or a system acts on. American banks have worked with the same three-piece idea since SR 11-7 in 2011, so if your team learned model risk in the US, nothing here will surprise you.
The middle piece is where the argument usually happens. Ask yourself whether the thing could be wrong because somebody guessed. If it relies on a fitted parameter, a chosen method or a view of how next year will look like last year, the answer is yes, and it’s a model. If it just adds up balances or converts dollars at the Bank of Canada’s published rate, there’s nothing to guess; that’s arithmetic. It may still need end-user computing controls, but it doesn’t need a validator.
OSFI didn’t leave AI to interpretation. The guideline says AI and machine learning methods are inside the definition. Your gradient-boosted fraud score is a model. So is the neural network that reads pay stubs in the mortgage file, and so is the large language model behind the chat window on your website.
A practical test
Most arguments end once someone asks these four questions out loud:
- Does it take input data and produce a result (a score, an estimate, a classification, a forecast, generated text)?
- Does the result depend on assumptions, estimation, statistics or a learned method, rather than a fixed, verifiable rule?
- Does someone rely on the result to make or support a business decision, or does it act on its own?
- Could the result be wrong in a way that matters: money, customers, capital, compliance, reputation?
Two yeses to the first pair and you have a model. The second pair doesn’t change that; it tells you how much attention the model deserves, which is the job of the risk rating. People often conflate the two. A model can be squarely in scope and still be low risk. It goes in the inventory either way.
Examples: model or not?
| The thing | Usually a model? | Why |
|---|---|---|
| Probability-of-default or loss-given-default model | Yes | Statistical estimates that drive capital and credit decisions. |
| IFRS 9 expected credit loss model, and its management overlay | Yes | Forward-looking estimates; the overlay is judgement applied to model output and belongs in the record. |
| AML transaction monitoring or sanctions screening | Yes, if it scores or ranks | Thresholds and scenarios are calibrated assumptions; regulatory risk is part of its rating. |
| Credit bureau score bought from a vendor | Yes | A third-party model. E-23 asks for vendor models to be inventoried and rated on their own. |
| Fraud detection with machine learning | Yes | A learned method making real-time decisions. |
| Customer-facing chatbot on a large language model | Yes | AI producing outputs customers rely on; E-23 asks for controls on explainability, bias and black-box behaviour. |
| Spreadsheet that forecasts staffing from past volumes | Often yes, low risk | It estimates; whether it matters decides its rating. |
| Spreadsheet that sums balances | Usually no | No estimation; an end-user computing control instead. |
| Deterministic rules engine (if X then decline) | Depends | If the rules and cut-offs were set by analysing data, many institutions treat it as a model. |
Edge cases that trip teams up
- Models hiding in vendor software. Plenty of loan origination systems ship with a scorecard baked in. The contract calls it a “decision engine” or doesn’t mention it at all. It’s still a model, and both E-23 and B-10 expect you to go looking for it.
- Overlays and judgement. When a credit committee bumps the IFRS 9 number up by 8% because it doesn’t trust the macro scenarios, that adjustment isn’t usually treated as a model of its own. E-23 does want the reason written down, though, next to the model it adjusts.
- Agents. An AI agent is a model with hands. The model underneath belongs in the E-23 inventory. What the agent is allowed to touch, and who approves its riskier actions, is covered by OSFI’s July 2026 AI bulletin. We wrote more on that in agentic AI governance.
- Models you just switched off. E-23 counts recently decommissioned models too. The scorecard you retired in August keeps its row, with the date and the reason.
If you’re still unsure, list it. Over-listing costs you a row in a spreadsheet and a light review. Under-listing costs you the conversation where an examiner finds a scoring model in a vendor contract that nobody on the risk team had heard of.
Related questions
Is a large language model a model under E-23?
Yes. E-23’s definition includes AI and machine-learning methods, and an LLM that processes inputs to generate results someone relies on meets it.
Is an Excel spreadsheet a model?
When it applies assumptions or statistical techniques to data to produce results people rely on, yes. A spreadsheet that only adds or looks up values is usually an end-user computing tool rather than a model.
Are vendor models in scope of E-23?
Yes. E-23 asks for third-party models to be identified, inventoried and rated on a standalone basis, with reviews extending to their data and components, in line with B-10.
Who decides whether something is a model?
The institution’s MRM policy sets the definition and usually the second-line model risk team rules on unclear cases, with the reasoning recorded.
Sources
- OSFI — Guideline E-23, Model Risk Management (2027)
- Federal Reserve — SR 11-7, Guidance on Model Risk Management (2011, superseded)
- Federal Reserve — SR 26-2, Revised Guidance on Model Risk Management
New answers are posted on MINERVA on LinkedIn. Follow along, or send the question you’d like answered next to minerva@minerva.studio.
A plain-language explanation, not legal advice. The regulator’s own text is the source; read it, and take advice on how it applies to your institution.