Selling a Model Fine-Tuning Services Business in Canada

A model fine-tuning services business has one preparation problem most sellers do not anticipate: the deliverable it sells — a customized model, tuned on a customer’s own data — sits inside a web of permissions from two different directions at once, the customer whose data was used and the foundation-model vendor whose base model was adapted. An owner who has decided to sell gets a materially cleaner process by sorting out both sides of that before listing, rather than discovering mid-negotiation that a term buried in a vendor agreement blocks part of what the buyer thought they were acquiring.

What to fix before listing

• Review every active customer fine-tuning agreement and confirm the weight-ownership clause is consistent, current and actually reflects how the engagement was delivered
• Put a documented data-retention and deletion process in place for customer data and model checkpoints once an engagement ends, if one does not already exist
• Read the foundation-model vendor’s own commercial or reseller terms for language restricting resale, transfer or assignment of fine-tuned derivatives on a change of control
• Confirm every contractor ML engineer who did tuning work signed a written intellectual-property assignment covering that specific work

The data-use licence gap most sellers do not think to check

The “what to fix” list above catches weight ownership, retention and vendor terms, but there is a fourth paper gap that surfaces just as often and is easy to miss because it looks, at a glance, like the same issue as weight ownership: whether each customer engagement actually documents a written licence to use that customer’s data for the specific training performed, separate from any clause about who owns the resulting weights. A services agreement that authorizes the engagement in general terms is not automatically the same thing as a data-use licence authorizing exactly what training was done with the customer’s information, and a seller who has never had this pointed out tends to assume the two are the same document. This is ordinary contract and confidentiality law rather than an AI-specific statute, so there is no regulator filing or special form involved — it is a matter of pulling every active customer file and confirming the actual written scope of what the business was permitted to do with that customer’s data, engagement by engagement, rather than assuming a standard template covers every case. Where the business has reused techniques or patterns learned on one engagement while serving a different customer, that reuse should also be checked against what the first customer’s agreement actually permitted. A seller who resolves this before listing removes a specific, recurring finding from the buyer’s due-diligence list rather than negotiating around it after an offer is already on the table.

The foundation-model vendor problem specifically

Because a fine-tuning business builds its product on top of someone else’s foundation model, the vendor’s own commercial terms function almost like a licence the business depends on, even though it is contractual rather than regulatory. Anti-assignment language in a supplier or platform agreement is common in software generally, and it applies here with particular force: if the foundation-model vendor’s terms require its consent before the business — or the specific reseller or fine-tuning-service arrangement — can transfer to a new owner, that consent needs to be part of the sale timeline from the start, not an afterthought discovered during the final week.

Confidentiality matters more here than in most sales

The data a fine-tuning business works with is frequently a customer’s own competitive secret — proprietary process knowledge, internal documents, or specialized domain data the customer would not want disclosed. That makes discretion during a sale process doubly important: a leak that a customer’s fine-tuning vendor is changing hands can itself alarm a customer who is sensitive about who has access to their data, independent of any actual change in how it is handled. A tightly controlled information flow, with a signed non-disclosure agreement ahead of any technical walkthrough, protects both the deal and the customer relationships the buyer is paying for.

What the buyer will ask for

Expect requests for the actual fine-tuning agreements with each customer, read specifically for weight-ownership and data-use language rather than taken on faith from a summary. Expect a request to see the foundation-model vendor’s commercial terms directly, not a paraphrase of them. And expect questions about the evaluation methodology itself — how improvement over the base model is actually measured and documented — because that methodology is frequently a meaningful part of what makes the business worth buying rather than replicating from scratch.

What commonly delays a close in this sub-sector

The most frequent delay is discovering, partway through diligence, that the foundation-model vendor’s terms require a formal re-approval or consent before the fine-tuning relationship can transfer, which then has to be sought without unsettling the vendor relationship the business depends on. A close second is a customer contract silent or ambiguous on weight ownership, which forces a scramble to get a written clarification from the customer before closing rather than after. Sellers who resolve both issues, or at least map them clearly, before going to market close meaningfully faster than sellers who leave them for the buyer to discover.