What is third-party AI, and what is an AI vendor?
By Nick Jang, founder of MINERVA Studio ·
The short answer
Third-party AI is any AI system or model an organization uses but did not build itself: a foundation model reached through an API, AI features inside software it licenses, or an AI service run by a vendor. An AI vendor is the company that supplies it. Third-party AI brings the usual supplier risks plus some of its own, mainly that the model can change under the contract and that data may travel on to the vendor’s own model provider.
The kinds of AI vendor
| Kind | Examples | What you control |
|---|---|---|
| Foundation model providers | Large language and vision models reached through an API or a cloud platform. | Your prompts, data and how you use the output; not the model. |
| AI built into software you buy | A CRM that drafts emails, a fraud tool with machine learning, document extraction in a lending system. | Settings and the contract; often little about the model. |
| Specialist AI products | Credit decisioning, AML screening, underwriting or chat assistants sold as AI products. | Configuration, data feeds and acceptance testing. |
| AI services and agents | Vendors that run AI agents or automated workflows on your behalf. | The scope you grant, the actions you allow, the monitoring you require. |
The second kind is the one that catches people out. A product you bought three years ago as plain software can pick up AI features in a quarterly update. There’s no new contract and no new review, and often nobody in risk hears about it until something goes wrong.
What makes third-party AI risky
- The model can change. The vendor may retrain it, swap the provider behind it or change its settings. What you assessed in March may not be what runs in September.
- Data can travel. Prompts, documents and customer data may be stored by the vendor, used to improve its model or passed to its own model provider, which is your fourth party.
- You see less. Vendors rarely share training data, methods or full test results, which makes independent review harder.
- Many products, few providers. A large share of AI products run on a handful of model providers and clouds, so one outage or policy change can hit many suppliers at once.
- The outputs are yours. If a vendor’s model makes an unfair credit decision or gives a customer wrong advice, the organization that used it answers for it.
How banks govern third-party AI in Canada
Three OSFI documents meet here. B-10 covers the supplier relationship: risk and criticality, due diligence, contracts with audit rights and incident notice, subcontracting, monitoring and exit. E-23 treats the vendor’s model as a model: it must be inventoried, rated on its own and reviewed down to its data and components. And OSFI’s July 2026 AI bulletin asks institutions to extend third-party risk management to vendors’ use of AI, to have vendors disclose it, and to judge how replaceable each AI service is.
In practice that means a few extra questions at every stage:
- Before buying: what the AI does, which model and provider sit behind it, what data it uses and keeps, and what evidence of performance, bias and security the vendor will share.
- In the contract: notice of material model changes, limits on using your data for training, disclosure of subcontracted model providers, incident notification, audit and information rights, and an exit plan.
- After go-live: monitoring of outputs you can see, a reassessment when the vendor changes the model, and a record of which of your models, agents and processes depend on the supplier.
Our AI procurement guide lists twelve due-diligence questions, and the vendor checklist is free.
Where MINERVA fits
MINERVA keeps every AI supplier on the same record as the models and agents that depend on it, walks each through B-10’s stages from due diligence to exit plan, keeps vendor models under E-23, and limits what a vendor’s AI may do with a passport every dependent agent inherits. A declared vendor incident opens a finding on every dependant for a person to decide. See AI vendor risk software.
Related questions
Is ChatGPT third-party AI?
Yes, for any organization that uses it or a product built on it. It is a model run by another company, reached through an app or an API.
Are third-party AI models covered by OSFI E-23?
Yes. E-23 expects vendor and third-party models to be identified, inventoried and risk-rated on a standalone basis.
What should an AI vendor contract include?
Notice of material model changes, limits on training with your data, disclosure of model providers and other subcontractors, incident notification, audit and information rights, data location and retention, and exit terms.
How do you find AI hidden in software you already use?
Ask vendors to disclose AI features and the providers behind them, review release notes, and check the third-party register for products that score, rank, classify or generate content.
Sources
- OSFI — Third-Party Risk Management Guideline (B-10)
- OSFI — Guideline E-23, Model Risk Management (2027)
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.