A plain-language guide Updated 26 October 2026
Third-party AI vendor risk, explained.
How a Canadian bank or insurer governs the AI it buys, under OSFI’s Guideline B‑10 on third-party risk and E‑23 on model risk.
01 At a glance
- What B‑10 is
- OSFI’s principles-based guideline on third-party risk management, covering every third-party arrangement, not only outsourcing.
- Published
- 24 April 2023, as the final guideline.
- In effect
- 1 May 2024.
- Applies to
- Every federally regulated financial institution, foreign bank and insurance branches included.
- How hard
- In proportion to each arrangement’s risk and criticality.
02 The eleven principles
| Principle | Section | What it asks, in plain words |
|---|---|---|
| 1 | 1.1 | The institution is accountable for third-party risk, whatever it hands to a supplier. |
| 2 | 1.2 | A third-party risk management framework sets how arrangements are found, assessed, managed and reported. |
| 3 | 2.2.1 | Every arrangement is assessed for its risk and criticality, before and during the relationship. |
| 4 | 2.2.2 | Due diligence is proportionate to risk and criticality, before contracting, at renewal and after material change. |
| 5 | 2.2.4 | Subcontracting risk is managed, including the supplier's own suppliers. |
| 6 | 2.3.1 | Arrangements rest on written agreements, reviewed by counsel. |
| 7 | 2.3.2 | Data and records are kept secure. |
| 8 | 2.3.3 | The institution and OSFI have rights to information and audit. |
| 9 | 2.3.4 | Business continuity is planned for. |
| 10 | 2.4.1 | Arrangements are monitored, individually and in aggregate, against metrics and thresholds. |
| 11 | 2.4.2 | Incidents are managed and reported. |
03 Why an AI supplier is different
- The model changes
- A supplier can retrain or replace the model behind a product without a new contract, so what was assessed is not what runs.
- AI arrives unannounced
- An update can add AI features to a product that had none, which moves it into model risk without anyone asking.
- Data goes further
- Prompts, documents and customer data may be used, stored or passed to the supplier's own model provider.
- Fourth parties
- Many AI products run on a handful of model providers and clouds: a subcontracting and concentration question B-10 already asks.
- It is also a model
- OSFI E-23 asks for vendor models to be inventoried, rated on their own and reviewed down to their data and libraries, in line with B-10.
04 What to have on file for each AI supplier
- An inventory of third parties by risk and criticality (B-10, 1.2.2), with every AI product and the models inside it marked.
- For each AI supplier: the risk and criticality assessment, the due diligence, and when it is owed again.
- The written agreement, with audit and information rights, subcontracting terms, data use and notice of material change.
- Which of your models, agents and processes depend on each supplier, and how far a failure would reach.
- Monitoring results, incidents and what was done about them.
- An exit plan for every critical AI supplier, reviewed regularly.
05 Software for third-party AI vendor risk
A vendor register in a spreadsheet cannot tell you which models and agents stop when a supplier does.
Whatever tool you choose, it should do the eight things below. Beside each is how MINERVA Studio does it. See also agentic AI governance, explained.
| What the software should do | How MINERVA does it |
|---|---|
| Keep the inventory of third parties by risk and criticality (1.2.2) | Every supplier is on the same record as your models and agents, with its band, owner, assessment and reassessment dates, and the B-10 fields still owed. |
| Walk each supplier through due diligence to exit | Each supplier stands at a station read from B-10's own groups: identified, diligenced, contracted, carried, monitored, exit planned. |
| Show what depends on each supplier | Traces every model, agent and process that uses or depends on a supplier, hop by hop, with the mitigations beside it. |
| Keep vendor models in model risk (E-23) | A vendor's model is inventoried, rated and reviewed like your own, with the supplier strung to it. |
| Limit what a vendor's AI may do | A vendor passport states which tools and data a supplier's AI may use; every agent that depends on the supplier is narrowed by it. |
| Handle a vendor incident | One declared incident opens a finding on every dependant, proposes isolating dependent models in production for a person to decide, and is closed by someone other than who declared it. |
| Remind you before a reassessment is due | Computes when each supplier is owed a reassessment from its band and its dependants, and says so when nothing was ever recorded. |
| Prove it to an examiner | Every assessment, decision and incident is on a time-stamped, hash-chained record exported as an evidence packet an examiner can verify offline. |
What it does not do: MINERVA does not assess a supplier for you or make an institution compliant. A person assesses and decides; MINERVA keeps the record, says what is owed and refuses steps that skip a decision. It runs on MINERVA’s cloud or in your own data centre, in English and French.
06 Questions people ask
What is OSFI Guideline B-10?
OSFI's guideline on third-party risk management for federally regulated financial institutions. It replaced the 2009 outsourcing guideline with a risk-based approach to every third-party arrangement, built on eleven principles covering accountability, a framework, risk and criticality assessment, due diligence, subcontracting, written agreements, data security, audit rights, business continuity, monitoring and incident management.
When did OSFI B-10 take effect?
On 1 May 2024. OSFI published the final guideline on 24 April 2023. Arrangements entered on or after 1 May 2024 are expected to follow it, and existing ones were to be brought in line by then or as soon as possible after.
Does B-10 cover AI vendors?
B-10 covers every third-party arrangement, so an AI supplier is covered like any other; the guideline's text does not single out AI. AI vendors are also reached through OSFI E-23, which asks institutions to inventory, rate and review vendor models down to their data and libraries, in line with B-10.
What makes third-party AI risk different?
The supplier can change the model behind a product without a new contract, updates can add AI to products that had none, data can travel to the supplier's own model provider, many products depend on the same few model providers, and the vendor's model is itself a model under E-23.
What should an AI vendor assessment include?
What the AI does and which of your data it uses; the models behind it and who provides them; how changes to the model are announced; performance, bias and security evidence; audit and information rights; subcontractors; data residency and retention; incident notification; and an exit plan, scaled to the supplier's risk and criticality.
How do you monitor a third-party AI vendor?
Against stated metrics and thresholds, individually and across all suppliers: service delivery, model changes, incidents, financial soundness and concentration, with reassessment on a cycle set by risk and after any material change, as B-10 section 2.4.1 asks.
What software helps banks manage third-party AI vendor risk?
Look for software that keeps suppliers in a register by risk and criticality, tracks each one from due diligence to exit, shows which models and agents depend on each, keeps vendor models in model risk, limits what a vendor's AI may do, handles incidents across dependants, and keeps an audit trail an examiner can verify. MINERVA Studio is built for that record, for Canadian federally regulated financial institutions.
What does MINERVA do for third-party AI vendor risk?
MINERVA keeps every supplier on the same record as the models and agents that depend on it, walks each through B-10's stations from identified to exit planned, shows the exposure hop by hop, keeps vendor models under E-23, limits each vendor's AI with a passport, turns one vendor incident into findings and isolation proposals on every dependant for a person to decide, and keeps all of it on a time-stamped record an examiner can verify offline.
This page summarises OSFI’s published guidelines in plain words. It is not legal advice; for any decision, read the guideline itself.