A plain-language guide By Nick Jang, founder · updated 28 October 2026
Third-party AI procurement, step by step.
How a Canadian bank or insurer buys AI from a vendor: the eight steps from naming what you buy to planning the exit, the questions to ask, and the clauses to ask counsel for, read against OSFI’s B‑10, E‑23 and its 2026 AI bulletin.
01 The eight steps
- 1. Name what you are buying
- A model, an AI system, an AI agent, or ordinary software with AI inside it. Ask the vendor to say whether and how it uses AI; OSFI's 2026 AI bulletin suggests requiring that disclosure.
- 2. Rate risk and criticality first (B‑10 2.2.1)
- Before the shortlist: what it will decide or do, for which customers or operations, and what breaks if it fails. The rating sets how deep the rest goes.
- 3. Due diligence, in proportion (2.2.2)
- Ask the questions below, before contract, at renewal and after a material change such as a new model behind the product.
- 4. Look behind the vendor (2.2.4)
- Name the vendor's own suppliers: the model provider, the cloud, any data broker. Their failure is yours.
- 5. Write it down (2.3.1–2.3.3)
- A written agreement reviewed by counsel, with the clauses below, data kept secure and rights to information and audit for you and OSFI.
- 6. Put its models in your inventory (E‑23)
- A vendor's model is a model: an owner, a risk rating, a validation approach that fits what you can see (outcome testing, benchmarks, monitoring) and an approval like any other.
- 7. Monitor and manage incidents (2.4.1, 2.4.2)
- Watch its performance against agreed lines, reassess on a cadence set by its rating, and know which of your models and agents depend on it when it fails.
- 8. Plan the exit
- Assess how replaceable the service is and keep a manual fallback for critical uses, as OSFI's 2026 AI bulletin suggests; agree how your data comes back.
02 Twelve due-diligence questions for any AI vendor
Ask them before the contract, at renewal and after a material change. Keep the answers on the supplier’s file.
| The question | Why it matters |
|---|---|
| Which models does the product use, whose, and which version? | What you assess is only what runs if you know what runs. |
| How and when will you tell us the model changed? | A supplier can retrain or replace a model without a new contract. |
| Is our data used to train or improve any model? | Customer data in someone else's model is a privacy and confidentiality question. |
| Where are our data, prompts and outputs processed and stored? | Residency and outsourcing rules may decide the deal before features do. |
| What evidence of performance, bias testing and limits can you share, on data like ours? | Your validation of a model you cannot see inside starts from theirs. |
| Which subcontractors sit behind the service? | The model provider and the cloud are your fourth parties. |
| Can the product act on our systems, and with which permissions? | An agent with broad access is a different risk from a model that returns a score. |
| What logs do we receive, and for how long? | Without logs you cannot show an examiner what the AI did. |
| How do you protect against prompt injection and data leakage? | Generative and agentic AI bring their own attacks. |
| How fast do you notify incidents? | Your incident clock starts when theirs does. |
| Will you give us and OSFI information and audit access? | B‑10 expects those rights in the agreement. |
| What happens to our data and the service at the end? | An exit you cannot execute is not an exit. |
03 Clauses to ask counsel for
- AI use and model changes
- The vendor discloses the AI it uses and gives notice before changing a model that serves you.
- Use of your data
- No training or improvement on your data without your written consent; deletion on request and at the end.
- Performance
- Agreed measures and lines, the monitoring data you receive, and what happens on a breach.
- Incidents
- Notification within a set time, with the information you need to act and report.
- Information and audit
- Rights for the institution and OSFI, including over relevant subcontractors.
- Subcontracting
- Notice of, or consent to, new or changed model and cloud providers.
- Location and security
- Where data is processed and stored, and the controls that protect it.
- Exit
- Transition help, return of your data and logs, and a period long enough to move.
04 Keeping it on one record
A spreadsheet can hold the questions; it cannot show which of your models and agents stop when a supplier fails. MINERVA keeps each AI supplier at its procurement station from due diligence to exit plan, declares the models inside its products in your E‑23 inventory, and sends one supplier incident to every model and agent that depends on it. See the AI vendor risk platform.
05 Questions people ask
What is third-party AI procurement?
Buying an AI model, AI system or AI agent, or software with AI inside it, from an outside vendor, and governing it from the first assessment to the exit: risk and criticality, due diligence, the contract, onboarding into the model inventory, monitoring and incidents.
Does OSFI B-10 apply when a bank buys AI?
B-10, OSFI's third-party risk management guideline, applies to third-party arrangements, and an AI vendor is a third party. Its expectations on risk and criticality, due diligence, agreements, information and audit rights, monitoring and incidents apply to the purchase.
Is a vendor's AI model inside OSFI E-23?
Yes. From 1 May 2027 E-23 applies to models whatever their source, so a vendor model needs an owner, a risk rating, validation suited to what the institution can see, approval and monitoring, like a model built in-house.
What does OSFI's 2026 AI bulletin add for vendors?
Its July 2026 technology risk bulletin on generative and agentic AI suggests extending third-party risk management to vendors' AI use, requiring vendors to disclose it, mapping AI dependencies to critical operations, testing outages with manual fallbacks, assessing how replaceable AI services are and reporting material concentration to the board.
How do you validate a vendor model you cannot see inside?
From the outside: the vendor's own evidence, outcome testing and benchmarks on your data, sensitivity checks where inputs can be varied, and ongoing monitoring against agreed lines. Record what could not be tested and the limits it puts on use.
What should a contract with an AI vendor include?
Disclosure of AI use and notice of model changes, limits on use of your data, performance measures, incident notification, information and audit rights for you and OSFI, subcontracting notice, data location and security, and exit help. Have counsel review it.
Is there a free AI vendor due-diligence checklist?
Yes: a checklist under OSFI B-10 as a spreadsheet, free with no sign-up, at minerva.studio/templates.
Read more: Third-party AI vendor risk under B‑10 · OSFI’s AI bulletin · OSFI E‑23 · Agentic AI governance. A plain-language guide, not legal advice.