What Canadian SMEs Should Check Before Signing an AI Vendor Contract

Most AI vendor contracts that Canadian SMEs sign were written by the vendor's legal team. They are optimized for the vendor. And because most organizations treat AI tool procurement like any other SaaS purchase — a few hours of evaluation, a credit card or a purchase order, and a click-through agreement — the gap between what those contracts say and what PIPEDA requires is rarely examined before the agreement is signed.

That gap has real consequences. Under PIPEDA's accountability principle, the organization that collects personal information from Canadians remains responsible for it even after handing it to a third-party processor. The Office of the Privacy Commissioner of Canada made clear through two landmark enforcement actions in 2026 that this principle applies to AI vendors. And in September 2026, the OPC published draft guidance on assessing third-party service providers — including AI tool providers — that for the first time gives organizations a formal due-diligence framework grounded in existing Canadian law.

For Canadian SMEs heading into 2027 vendor renewals and new AI deployments, the guidance and the enforcement record together define a minimum standard that the OPC will apply if something goes wrong.

Why AI Vendor Contracts Aren't Like Other Software Agreements

A standard SaaS agreement governs access to software hosted by the vendor. The vendor processes your data to deliver the service. AI tool agreements do that too — but they typically go further in ways that standard software contracts were not designed to address.

AI foundation model vendors commonly use customer data to improve their models. Their systems route prompts through multiple subprocessors — the foundation model provider, cloud infrastructure vendors, vector databases, logging services — each governed by its own terms. The underlying model may change materially between the time you sign and the time you renew. And AI system outputs carry liability implications that static software outputs do not. As the BC Civil Resolution Tribunal established in *Moffatt v. Air Canada* (2024 BCCRT 149), organizations cannot disclaim responsibility for incorrect information their AI systems provide to customers — the AI is treated as the organization's agent, not as a third party whose errors are its own problem.

Each of these characteristics creates a risk category that standard SaaS contract terms were not designed to address.

The PIPEDA Accountability Principle — and the OPC's New Guidance

PIPEDA's Principle 4.1.3 states that organizations are responsible for personal information transferred to a third party for processing, and they must use contractual or other means to provide a comparable level of protection. This is not a new rule. What is new is how the OPC has started operationalizing it specifically for AI vendors.

On September 10, 2026, the OPC published its draft guidance on assessing third-party service providers under PIPEDA. The comment period closes December 4, 2026. The guidance covers any PIPEDA-covered organization that uses a third party's products or services to handle personal information — explicitly including AI tool providers. Bennett Jones' summary of the draft highlights three AI-specific requirements that go beyond general due diligence:

• Organizations must carry out due diligence on where the vendor's AI training data came from, including whether it was lawfully collected under applicable privacy law.
• Where a vendor claims personal information has been anonymized, organizations must assess whether re-identification is realistic given the nature of the dataset. A vendor's assertion alone does not satisfy the accountability standard.
• Due-diligence obligations extend to subprocessors throughout the AI pipeline, not only to the primary vendor named in the agreement.

The guidance matters because it converts the general accountability principle into a checkable due-diligence process — one the OPC can benchmark any organization against when investigating a complaint. For businesses that have never formally reviewed their AI vendor contracts, this draft defines the standard they will be measured against once it is finalized.

It is also worth noting the legislative context. AIDA and the AI provisions of Bill C-27 died when Parliament was prorogued in 2025, and successor legislation remains in progress. For now, PIPEDA — interpreted through this guidance and enforced through investigations — carries most of the regulatory load on AI vendor accountability for federally regulated businesses.

Seven Contract Clauses Canadian SMEs Should Check

MLT Aikins and other Canadian legal commentators identify contract review as one of the most consistently underdone steps in AI procurement. These are the seven clauses that matter most under PIPEDA and the September 2026 guidance.

1. Training data rights. Does the contract permit the vendor to use your data — inputs, uploaded documents, and generated outputs — to train or improve the model? This is described consistently as the single most consequential clause in any AI vendor agreement. Acceptable terms are increasingly available: Claude for Enterprise and Microsoft 365 Copilot, for example, explicitly do not use customer data for model training by default. Standard consumer-tier agreements often do not make this commitment. If your team is using the consumer tier of an AI tool with client, employee, or patient data, check the terms before the next renewal.

2. Subprocessor disclosure. Who does the vendor route your data through? AI systems commonly use multiple subprocessors for inference, vector storage, logging, and infrastructure. The contract should identify these subprocessors, state which jurisdictions they process data in, and require the vendor to provide advance notice before adding new subprocessors. Without this, you cannot run a meaningful cross-border transfer assessment under PIPEDA.

3. Data residency and cross-border transfer. If your AI vendor hosts infrastructure in the United States or elsewhere outside Canada, PIPEDA allows the transfer — but your organization remains responsible for ensuring a comparable level of protection exists, documented in the contract. For businesses operating in Quebec or handling Quebec residents' personal information, Quebec Law 25 adds a formal Privacy Impact Assessment (PIA) requirement before personal information leaves the province. This requirement applies to any US-hosted AI service and must be completed before deployment, not after.

4. Data retention and deletion. How long does the vendor keep your data, in what form, and what happens to it at contract end? Standard deletion clauses in SaaS agreements often do not map correctly to how AI vendors actually store data. Prompt histories, cached outputs, and any fine-tuning data are frequently retained separately from the core service data. The contract should specify each data category, the retention period for each, and the deletion process — including confirmation that deletion requests propagate to subprocessors.

5. Anonymization and re-identification. If the vendor claims your data is anonymized before any secondary use, the OPC's September 2026 draft guidance says organizations must assess whether re-identification is realistic given the nature and richness of the dataset. A vendor's assertion that data is anonymized is not sufficient on its own. Where the data is sensitive — health information, financial records, HR data — the bar for satisfying this requirement is higher.

6. Audit rights. The contract should give you the right to request information about the vendor's current security and privacy practices, and to pass downstream any information requests that arise in a privacy breach investigation. Without audit rights, demonstrating your own accountability under PIPEDA becomes difficult, and verifying that the vendor's practices match their representations is not possible.

7. Model change and exit rights. Foundation models change. API versions deprecate. Vendors are acquired. The contract should specify advance notice periods for material changes to the underlying model, and exit terms that include portable deletion of your data within a defined timeframe. As Canadian legal commentary on AI procurement consistently notes, exiting an AI vendor relationship can be considerably harder than ending a conventional software licence, particularly where the vendor has ingested your data into systems not easily separated.

What Two 2026 Enforcement Actions Show

Both major OPC enforcement actions in 2026 involved AI vendors and found violations rooted in accountability and consent failures that contracts could have partially addressed.

The joint OPC, CAI, OIPC-BC, and OIPC-AB investigation of OpenAI's ChatGPT (PIPEDA Findings #2026-002, May 6, 2026) found that OpenAI violated federal and provincial privacy laws in how it collected and used personal information to develop ChatGPT. The regulators rejected OpenAI's claim that it relied on valid implied consent. The practical guidance issued alongside the findings is specific: organizations using third-party AI services that may train on personal information have accountability obligations they cannot delegate to the vendor.

The June 2026 OPC finding on Grok (PIPEDA Findings #2026-004) found that X Corp. and X.AI violated PIPEDA's consent and appropriate-purposes requirements in the deployment of the Grok image generation tool — including launching the product without conducting a privacy impact assessment. This matters for Canadian SMEs not because they are likely to be the subject of an enforcement action but because both cases define what the OPC looks for when investigating organizations that relied on third-party AI tools without adequate accountability measures in place.

Starting the Review

The September 2026 OPC draft guidance gives Canadian organizations a de facto checklist before the December 4 comment period closes and the guidance is finalized. The practical step is straightforward: for each AI tool your business currently uses with personal information, review the vendor's terms of service and data processing agreement against the seven clauses above. Flag the gaps. For enterprise-tier agreements, engage legal counsel to negotiate improved terms. For consumer-tier tools being used with client, employee, or patient data, the question is whether an enterprise agreement or a different tool is warranted.

The organizations that face difficulty under PIPEDA and Quebec Law 25 are, with few exceptions, the ones that never asked these questions before they signed.

Sources

• Office of the Privacy Commissioner of Canada. *OPC Draft Guidance on Assessing Third-Party Service Providers.* News release, September 10, 2026. priv.gc.ca
• Bennett Jones LLP. *OPC Issues New PIPEDA Guidance on Assessing Third Party Service Providers.* bennettjones.com (September 2026)
• Office of the Privacy Commissioner of Canada. *PIPEDA Findings #2026-002: Joint investigation of OpenAI.* priv.gc.ca (May 6, 2026)
• MLT Aikins LLP. *Practical takeaways from the landmark joint investigation of OpenAI's ChatGPT.* mltaikins.com (May 2026)
• Office of the Privacy Commissioner of Canada. *PIPEDA Findings #2026-004: Grok (X Corp. and X.AI).* News release, June 11, 2026. priv.gc.ca
• MLT Aikins LLP. *Negotiating technology agreements in the age of AI: Key considerations for Canadian organizations.* mltaikins.com (2026)
• Commission d'accès à l'information du Québec. *Act respecting the protection of personal information in the private sector (Law 25).* cai.gouv.qc.ca

Cloud Forces' AI Advisory service includes an AI governance and security review designed for Canadian SMEs and mid-market companies that want to know where their AI vendor relationships stand before an enforcement action makes that question urgent. We review your AI tool portfolio against the OPC's September 2026 draft guidance, map your data flows to identify subprocessor and cross-border transfer risks, and produce a remediation plan your legal and IT teams can execute. As an AWS Consulting Partner since 2019 and Microsoft Solutions Partner since 2022, we bring the technical depth to verify vendor claims against actual architecture alongside the advisory expertise to strengthen your contracts and governance posture. Contact us to discuss your AI vendor review.