Conversational AI platform due diligence

Diligence on a conversational AI platform has to verify claims that are unusually easy to overstate in good faith — a resolution rate measured against a friendly test set rather than live customer traffic, a "we don’t misuse your data" assurance with no contractual backing behind it at all. A buyer under a letter of intent needs a checklist built specifically for those failure modes, not a generic software-acquisition template borrowed from an unrelated deal.

The documents to request

• Data-processing agreements with every customer, confirming what the platform is contractually permitted to do with conversation data, including any training or fine-tuning use of it
• Signed IP assignment agreements from every contractor involved in model fine-tuning or prompt engineering, closing off any question of who actually owns the work product
• The foundation-model vendor’s current terms of service, including any resale, embedding, change-of-control or termination provisions that could affect the business post-acquisition
• Source data behind any containment, deflection or resolution-rate figures used in sales or marketing materials, broken out by customer rather than presented as one blended number
• Documentation of the fallback and human-handoff design, and, where Quebec customers are involved, the process for honouring a human-review request under Law 25

Registry and public-record searches

A trademark search through the Canadian Intellectual Property Office confirms the platform’s brand and product names are actually registered, or at least filed, in the corporation’s name rather than a founder’s — a detail easy to miss on a company that grew quickly and never cleaned up its early paperwork. An execution and judgment search through the relevant provincial registry confirms there’s no outstanding judgment against the corporation that could attach to the business after closing and complicate the transfer. Neither search is a formality on a software business; unresolved trademark ownership in particular is a common, fixable-but-annoying finding on companies that moved fast in their early years and treated branding as an afterthought.

Verifying the numbers independently, not just requesting them

Requesting a resolution-rate report from the seller is a start, not an end, because the report itself can be built on a definition of "resolved" that flatters the platform without technically being false. A thorough technical review pulls a sample of the raw conversation logs directly — not the seller’s aggregated dashboard — and has an independent reviewer classify a subset by hand to see whether the seller’s own definition of resolution matches what a reasonable buyer would call resolved. Where access can be arranged, it’s worth cross-checking API usage logs against the foundation-model vendor account, since actual query volume should roughly track the customer and conversation counts the seller is representing; a mismatch between the two is one of the more reliable early signals that the numbers being presented don’t fully reconcile with what the platform is actually doing in production. None of this replaces the seller’s own reporting, but it turns a figure the buyer is being asked to trust into one the buyer has actually tested.

Findings that commonly kill conversational AI deals

Customer conversation data used to train or fine-tune models without contractual permission is the finding that ends the most deals in this category, because it means the buyer would be inheriting a liability toward the platform’s customers — and, by extension, those customers’ own end-users — with no clean way to make it right after the fact. A close second is discovering that the foundation-model vendor’s terms of service would bar the platform’s actual use case, or the specific deal structure being contemplated, once read closely rather than assumed to be fine. A missing IP assignment from a contractor who did meaningful fine-tuning work is a third recurring finding, leaving open whether the company actually owns what it’s selling to the buyer.

What a finding actually means when it appears

A resolution-rate figure that turns out to be measured differently than marketed — against a narrower ticket category, say, rather than the full support volume — is usually a repricing conversation, not a collapse, provided the underlying product still performs well. A training-data consent gap is more serious and depends heavily on scope: a small pilot customer with no clear agreement is a fixable pre-closing condition, while a data-use practice that touches most of the customer base without consent is a structural liability that belongs in the price, the indemnity, or the decision to proceed at all. Getting a straight answer on which situation you’re actually in is what technical and legal diligence run together, rather than in sequence, are for.

Foundation-model dependency as a diligence finding

Reliance on a single foundation-model vendor isn’t just a strategic risk to note in a summary memo — it’s a contractual fact to pull and read directly. Confirm what the vendor’s current terms actually permit, what happens if pricing or terms change without much notice, and whether the platform has any technical or contractual path to a different model if it needs one. A buyer who finds no such path isn’t necessarily walking away from the deal, but is pricing in the real cost of eventual re-platforming, and that cost belongs in the offer, not left as an unpriced risk discovered only after closing.