AI content generation tool due diligence

By the time you are under an LOI on an AI content generation tool, the diligence work is not about deciding whether to buy anymore — it is about confirming the business is what the seller represented it to be, and finding that out before closing rather than after. The findings that actually derail deals in this category cluster around three things: where the training data came from, what the customer contracts say about output ownership, and how dependent the product really is on one outside vendor. Everything else in a standard technology-diligence checklist still applies, but these three carry disproportionate weight here.

Documents to request before you rely on anything else

Ask for the licensing records behind every training and fine-tuning dataset the current models were built on, not just a summary — you want the actual licence terms, not a seller’s characterization of them. Pull every active customer subscription agreement and any separate output-licensing terms, and read them for consistency: contracts drafted at different points in the company’s history sometimes say different things about who owns generated content, which is itself a finding worth flagging before you even get to the substance. Where fine-tuned models, style layers or prompt libraries are represented as transferable assets, get the underlying data-licensing terms that would actually have to permit that transfer.

Registry and public-record checks worth running

Run a Canadian Intellectual Property Office search on any trademarks, and confirm the copyright position on the software and any registered assets through CIPO’s transfer-of-ownership process rather than taking the seller’s word on what is actually registered and to whom. Confirm the corporation’s standing and, where the deal involves a share purchase, that there is no undisclosed CRA debt or judgment against the corporate entity you would be inheriting along with everything else — quick checks relative to the technical and copyright work, but ones that surface real problems often enough to be worth running as a matter of course.

Confirm what can actually transfer and what needs a fresh consent

Not everything the seller describes as an asset moves automatically to a new owner, and this category has more of that problem than most software businesses. Confirm directly with the underlying foundation-model vendor, any cloud-hosting provider and any marketplace or plugin-store platform the product is distributed through whether the existing agreement can actually be assigned to you, or whether each one requires its own fresh application that the platform has no obligation to approve — a marketplace listing with an established review history and customer base is worth far less if it cannot move with the sale. Get the licensing terms behind any fine-tuned models, style layers or prompt libraries the seller represents as transferable, and read them for whether the underlying data licence itself actually permits that kind of transfer, rather than assuming a licence that allowed the seller to use the data also allows the seller to hand it to you. Where the business has built content libraries or templates in-house, confirm the corporation actually owns them outright rather than a contractor or a departed employee who built them and never formally assigned the rights over.

Findings that actually kill this kind of deal

The findings that most often end a deal outright are training or fine-tuning data with unclear or unlicensed provenance — particularly where scraped copyrighted material is involved — customer contracts that are silent or contradictory on who owns generated output, the absence of any indemnity or clarity on copyright-infringement risk passed through to customers, and an undisclosed single foundation-model dependence for the core generation engine the seller did not flag up front. None of these is automatically fatal if disclosed early and priced or fixed, but discovering any of them late in the process, after the seller represented otherwise, is what actually kills deals.

What a finding means when you actually see it

An unclear answer on output ownership does not mean the business is worthless — it means every existing customer relationship carries a dispute waiting to happen if a customer later claims ownership over content the vendor also claims, or vice versa, and fixing it after closing is your problem, not the seller’s. Undisclosed single foundation-model dependence does not mean the product does not work — it means the vendor could change pricing, access terms or availability at any time and the business has no fallback, which changes how you should think about the durability of the earnings you are paying for.

Privacy and marketing-claim exposure to verify

Where prompts or generated outputs include personal information, PIPEDA applies everywhere in Canada, and Quebec’s Law 25 layers on additional obligations specifically for Quebec-based operations or Quebec residents’ data — confirm which framework actually governs the business you are buying rather than assuming PIPEDA alone covers it where there is any Quebec exposure. Separately, review any marketing claims the seller has made about originality, freedom from copyright risk, or human-equivalent output quality; unsubstantiated claims of that kind draw Competition Bureau scrutiny under its deceptive-marketing-practices framework, and that exposure transfers to you along with the brand.