AI Due Diligence in Canadian M&A
AI due diligence tests whether a Canadian target owns and may lawfully use the models, data and code behind its valuation. With no federal AI statute in force, the review turns on privacy law, copyright, contracts and sector rules. This guide sets out what to examine, the red flags, and how to draft around what diligence cannot resolve.

Key takeaways
- Canada has no federal AI statute in force; Bill C-27 died in January 2025 and Bill C-36 is not law.
- Trace training data sources and licences model by model, including customer data rights.
- Privacy laws, including Quebec's Law 25, apply to personal information used in training and inference.
- Customer contracts may contain uncapped output indemnities or no-training promises.
- Use AI-specific representations plus special indemnities for known issues.
What AI due diligence covers in a Canadian deal
AI due diligence is the part of an acquisition review that tests whether a target actually owns, controls and may lawfully use the artificial intelligence on which its value depends. In Canadian M&A it typically covers six areas: the provenance and licensing of training data, ownership of models and code, open-source components, privacy compliance for training and inference, customer contract exposure, and the representations and indemnities that allocate what diligence cannot resolve. It also includes a view on how the buyer's own use of AI tools in the diligence process should be controlled.

The legal backdrop in Canada is unusual. There is still no federal statute specific to artificial intelligence: the proposed Artificial Intelligence and Data Act died with Bill C-27 when Parliament was prorogued on 6 January 2025. A successor privacy bill, Bill C-36, received first reading on 15 June 2026 and, according to LEGISinfo, had not received royal assent at the time of writing. AI risk in a Canadian target is therefore assessed mainly under existing privacy, intellectual property, consumer protection, employment and sector-specific rules, which makes careful AI due diligence more important, not less.
Training data provenance and licences
The first question is where the data came from and on what terms. A model trained on data the target had no right to use is a liability that can follow the model into the buyer's hands. Diligence should establish, for each significant model:
- Sources. Proprietary data, customer data, licensed datasets, public web data, synthetic data and data generated by users of the target's products.
- Licence terms. Whether licences permit commercial model training, whether they survive termination, and whether they prohibit transfer to an acquirer or require consent on a change of control.
- Customer data rights. Whether customer agreements allow the target to use customer data to train models that serve other customers, or only to deliver services to that customer.
- Scraping. Whether web data was collected in breach of website terms or technical restrictions, and whether it includes personal information.
- Records. Data inventories, dataset cards and lineage documentation that let the target prove what went into each model.
Copyright risk is real but unsettled. Canada's Copyright Act has no specific exception for text and data mining, and the federal government has consulted on whether to create one. Whether training on copyrighted works is permitted under fair dealing has not been resolved by an appellate court in Canada. Buyers should treat large-scale use of unlicensed copyrighted content as a contingent risk to be priced or indemnified; see our note on AI training data and copyright risk.
Model ownership, code and open-source components
Ownership of AI assets is often less clear than in conventional software deals. Questions include:
- Who owns the model weights, fine-tuned versions and prompts? Were they developed by employees, contractors or a partner under a joint development agreement?
- Were foundation models licensed from third parties, and do those licences restrict commercial use, user numbers, competing products, or use by an acquirer?
- Does the code base include open-source components under copyleft licences that could require disclosure of proprietary code if distributed?
- Have IP assignments and moral rights waivers been obtained from every contributor? Under s. 13(3) of the Copyright Act, employers own copyright in works made in the course of employment, but contractors and founders keep ownership unless they assign it.
Many targets build products on third-party model APIs rather than their own models. Then the key diligence points are the API provider's terms: data use by the provider, service levels, pricing changes, termination rights, and any prohibition on using outputs to develop competing models. Concentration on a single provider is a commercial risk worth flagging to the investment committee.
Export controls and data residency can also bite. Some advanced computing items and related technology are subject to export controls, and some public-sector and financial customers require Canadian data residency or restrict cross-border processing. A target that relies on foreign hosting for model inference may need to change its architecture to keep those customers after a change of control.
Privacy compliance for training and inference
Personal information is the most heavily regulated input to most AI systems in Canada. The federal Personal Information Protection and Electronic Documents Act (PIPEDA) applies to private-sector organizations in the course of commercial activities, except where substantially similar provincial laws apply, as in Quebec, Alberta and British Columbia. Under these laws, organizations generally need meaningful consent for collection, use and disclosure, must limit use to purposes a reasonable person would consider appropriate, and must protect the information with appropriate safeguards.
For AI due diligence, that translates into specific questions:
- Was personal information used to train models for purposes beyond those identified at collection?
- Was "publicly available" personal information scraped on the assumption that it could be used freely? Canadian privacy regulators have taken a narrow view of what counts as publicly available.
- Are privacy impact assessments in place, especially for Quebec, where Law 25 requires them in specified situations, including before personal information is communicated outside Quebec?
- Do inference outputs or logs contain personal information, and how long are they kept?
- Has the target given notice where decisions about individuals are made exclusively by automated processing, as Quebec requires?
Exposure is not theoretical. The Commission d'accès à l'information can impose administrative monetary penalties under Law 25 of up to CAD 10 million or 2% of worldwide turnover, according to the CAI's summary of the main changes. Our note on automated decisions under Law 25 covers the Quebec rules.
Customer contract exposure
The target's customer contracts reveal how much AI risk it has already promised to absorb. Diligence should review:
- warranties about accuracy, non-infringement and fitness of AI outputs;
- indemnities for third-party IP claims arising from outputs, and any caps or carve-outs;
- commitments not to train on customer data, and whether the target's practices match them;
- audit rights, transparency commitments and obligations to explain decisions;
- change-of-control clauses that let key customers terminate on an acquisition.
Customers in regulated sectors increasingly flow down their own obligations. For example, federally regulated financial institutions will need to apply OSFI Guideline E-23 on model risk management, effective 1 May 2027, to AI and machine learning models, including vendor models. A target selling AI to banks or insurers should be ready for that scrutiny; see our note on OSFI Guideline E-23.
An AI due diligence checklist
| Area | Key documents | Typical red flags |
|---|---|---|
| Data | Data inventory, dataset licences, customer DPAs | No lineage records; training on customer data without contractual right |
| Models | Model register, foundation model licences, development agreements | Licence bars commercial use or transfer; unclear ownership of fine-tuned weights |
| Code | Open-source scan, contributor IP assignments | Copyleft components in distributed products; missing assignments |
| Privacy | Privacy policies, PIAs, breach records, consent flows | Scraped personal information; no Quebec PIA for cross-border transfers |
| Contracts | Customer and vendor agreements | Uncapped IP indemnities for outputs; "no training" promises breached |
| Governance | AI policy, incident logs, testing and bias assessments | No inventory of AI use; no human review of high-impact outputs |
| Marketing | Product claims, investor materials | Overstated AI capabilities ("AI washing") |
The last row matters more than it seems. Overstating what a product's AI can do can raise misleading advertising issues under the Competition Act and, for reporting issuers, disclosure issues under securities law.
AI-specific representations and indemnities
Because AI due diligence rarely produces complete answers, purchase agreements increasingly include AI-specific representations. Common examples include representations that the target has a complete inventory of AI systems it develops or uses in its products; that it has the rights necessary to use all training data as used; that it has complied with privacy laws and its own policies in collecting and using data for training; that no customer data has been used in breach of contract; that it has not received claims or regulatory inquiries about its AI; and that it maintains governance policies proportionate to its use of AI.
Sellers push back on representations about "all" training data or about outputs, which can be impossible to verify. Knowledge qualifiers, materiality thresholds and disclosure schedules are the usual compromise. Where diligence reveals a specific problem, such as a dataset licence that prohibits commercial training, a special indemnity, escrow or a pre-closing remediation covenant (retraining without the dataset) is more effective than a general representation. Representations and warranties insurers may exclude known AI issues or broad AI risks, so buyers should check the policy wording early.
Ask the target for a model-by-model register before the data room opens: what each model does, what data trained it, who owns it, and which customers use it. If the target cannot produce one, that is itself a diligence finding, and it predicts how hard post-closing integration will be.
Limits of AI-assisted diligence tools
Buyers and their advisers increasingly use AI tools to review data rooms. These tools are useful for triage, clause extraction and summarizing large volumes of contracts, but they have limits. They can miss non-standard clauses, misread scanned documents and produce confident errors. Confidentiality agreements may restrict uploading target documents to third-party platforms, and documents may contain personal information or competitively sensitive data subject to clean-team rules. A sensible approach uses AI tools under written protocols, with human review of all material findings and a record of what was reviewed by whom. Courts have also criticized filings that relied on unverified AI output; see our note on AI in litigation.
From AI due diligence to integration
Findings from AI due diligence should feed directly into the integration plan. Typical first-100-day actions include bringing the target's AI systems into the buyer's inventory and governance framework, aligning privacy notices and consent flows, replacing or relicensing problematic datasets, renegotiating customer terms that over-promise on outputs, and putting monitoring in place for models used in high-impact decisions. Employment practices also need review: Ontario employers with 25 or more employees must disclose in publicly advertised job postings whether AI is used to screen, assess or select applicants, a rule in force since 1 January 2026.
Boards of acquiring companies are increasingly expected to oversee AI risk as part of enterprise risk management. A concise report to the board on what the diligence found, what was fixed before closing and what remains open helps directors discharge that oversight and creates a record if issues later surface. Our note on board oversight of AI risk discusses the governance expectations.
How TRW & Co helps with AI due diligence
We conduct AI due diligence for buyers and investors in Canadian technology, data and AI-enabled businesses, and we help sellers prepare. Our work covers data provenance and licensing, model and code ownership, privacy compliance under PIPEDA and Quebec's Law 25, customer contract exposure, AI-specific representations, indemnities and remediation covenants, and protocols for using AI tools in diligence. See our artificial intelligence practice page and our M&A page. This note is general information, not legal advice. To discuss a transaction, contact us.
Frequently asked questions
Is there a Canadian AI law that applies to M&A targets?
Not a general federal one. The Artificial Intelligence and Data Act died with Bill C-27 on prorogation in January 2025, and the successor privacy bill, C-36, had not received royal assent as of the latest LEGISinfo entry. AI risk is assessed under privacy, copyright, consumer protection, employment and sector-specific rules such as OSFI guidelines.
What is the biggest AI due diligence risk?
Often it is training data: data used without a licence permitting commercial training, customer data used beyond contractual rights, or scraped personal information. Those problems can taint models the buyer is paying for. A model-by-model register with data sources and licences is the best starting point.
Does Quebec's Law 25 matter for AI targets?
Yes. It requires privacy impact assessments in specified situations, including before personal information is communicated outside Quebec, and notice when decisions are based exclusively on automated processing. The CAI can impose administrative monetary penalties of up to CAD 10 million or 2% of worldwide turnover.
Can AI issues be covered by R&W insurance?
Partly. Insurers may cover breaches of AI representations that were unknown at signing, but they often exclude known issues and may limit coverage for broad AI or IP risks. Known problems are better handled with special indemnities, escrows or pre-closing remediation such as retraining a model without a problematic dataset.
Can we use AI tools to review the data room?
Yes, with care. Check the confidentiality agreement and clean-team rules before uploading documents to third-party tools, protect personal information, and ensure that a person reviews every material finding. AI tools are good at triage and extraction but can miss unusual clauses or misread documents.
Sources
General information only, not legal advice. References to statutes, courts and authorities describe published law and practice at the date shown.


