Data-labelling and annotation business due diligence

Under an LOI on a data-labelling business, the diligence process is not about whether the client list looks good anymore — it is about confirming the contracts behind that list are real and durable, and that the workforce and data-handling practices behind the delivery are what the seller represented. The findings that actually derail these deals cluster around contracts, workforce classification and client-data handling, in that order of frequency, and a diligence process that treats all three as first-tier issues rather than boilerplate catches problems before they become your problem.

Documents to request before you rely on anything else

Ask for every master service agreement and per-project contract with the business’s AI-lab and enterprise clients, not a client list — you need the actual terms, including assignment clauses, termination rights and any exclusivity commitments, rather than a summary taken at face value. Request the documented annotation workflows, taxonomies and quality-control tooling that supposedly make the operation repeatable, and test whether that documentation is actually current or describes a process the business has since drifted away from. Pull the individual employment or contractor agreements governing the annotator workforce as well, since the terms on paper are what you will be testing against actual working conditions next.

Verifying how the workforce is actually engaged

Worker classification is governed by ordinary provincial employment-standards law, tested on factors like control over how the work gets done and integration into the business, not by any AI-specific rule — so the diligence question is not whether contracts say "contractor," it is whether the actual working relationship matches that label. Interview a sample of annotators about how their schedules, tools and supervision actually work if you can, since a contractor relationship that functions like an employment relationship in practice is exactly the kind of finding a written agreement alone will not reveal.

Registry and record checks worth running

Confirm the corporation’s standing and check for any outstanding CRA debts or judgments against it, particularly given how payroll-adjacent this business is — a workforce reclassification dispute can generate exactly the kind of tax exposure a corporate-status and debt check is designed to surface. Where the business has represented that it holds proprietary labelling or quality-control software, confirm that IP is actually owned by the corporation and not by an individual contractor or a departed founder who built it and never formally assigned it over.

Findings that actually kill this kind of deal

The findings that most often end a deal outright are client datasets that were retained, reused or repurposed beyond the scope of the original engagement — a serious breach of client trust that can trigger contract termination on its own — no confidentiality or data-handling terms on file for sensitive client data such as health, biometric or financial information, worker-classification exposure serious enough that it could convert the contractor workforce into employees retroactively, and revenue concentrated in one or two large clients who have the technical capability to bring the work in-house. Any one of these, undisclosed until diligence uncovers it, tends to do more damage to the deal than the underlying fact itself.

What a finding means when you actually see it

A retroactive worker-reclassification exposure does not just mean paying more going forward — it can mean liability for past unpaid entitlements calculated back over however long the misclassification ran, which is a materially different number than a forward-looking cost adjustment. A client-data handling lapse does not necessarily mean the data was misused — it can simply mean nobody can currently prove it was not, which is enough on its own to jeopardize a sensitive-sector client relationship regardless of what actually happened.

Privacy and cross-border data exposure to check

Where client datasets include personal information, PIPEDA governs the business’s obligations everywhere in Canada, and Quebec’s Law 25 adds further obligations specifically where the operation is Quebec-based or the data belongs to Quebec residents — confirm which framework actually applies given where the business and its workforce operate. Where any part of the annotation workforce operates outside Canada, also check what the client contracts actually promise about data residency and access, since those commitments are typically addressed contractually rather than by statute and are easy for a seller to have made informally without documenting properly.