The Legal AI 'Affiliate' Definition Problem: Why Contract Review Tools Flag Cross-Affiliate Restrictions Without Checking Whether the Definition Has a Threshold That Changes the Entire Scope
There is a specific failure mode in AI contract review that senior in-house counsel and M&A associates are beginning to notice, complain about quietly, and then work around manually — which is precisely the kind of failure that should not require a workaround. The tools...
There is a specific failure mode in AI contract review that senior in-house counsel and M&A associates are beginning to notice, complain about quietly, and then work around manually — which is precisely the kind of failure that should not require a workaround. The tools flag affiliate-related provisions. They summarise what those provisions say. And they almost never check whether the contractual definition of "affiliate" actually covers the entity you are asking about.
This is not a minor gap. In a material number of commercial agreements, the affiliate definition is the substance of the risk, not background boilerplate.
What the Tools Actually Do
Virtually every AI contract review product on the market — whether a standalone tool or embedded within a broader CLM workflow — processes affiliate-related covenants through a pattern-recognition layer. The tool identifies clauses containing affiliate language: confidentiality obligations extended to affiliates, licensing grants that include or exclude affiliates, non-compete restrictions that purport to bind the whole corporate family. It then summarises the provision as written.
What it does not do, with any reliability, is trace the affiliate definition and apply it to the specific entity topology at issue. The summary tells you that affiliates are bound by the confidentiality restriction. It does not tell you whether the definition of "affiliate" is pegged to majority ownership, beneficial control, a 20% threshold, or some bespoke formulation that carves out joint ventures entirely. The flag looks the same regardless.
Why the Definition Is Usually Where the Issue Lives
The threshold problem is real and surprisingly common. A confidentiality restriction that extends to "affiliates" sounds broad. If the agreement defines "affiliate" as an entity in which a party holds more than 50% of the voting securities, that definition excludes, by its own terms, every minority-owned subsidiary, every preferred-share structure that lands below the majority line, every 50/50 joint venture, and every entity where control is achieved through board seats or contractual arrangements rather than equity ownership.
The practical consequences are not hypothetical. A manufacturing conglomerate licensing enterprise SaaS across its operating group may assume the agreement covers its 40%-owned regional venture. It does not. An acquirer reviewing target contracts in an M&A process may identify what looks like a broad affiliate confidentiality obligation binding the whole corporate tree and miss that the definition excludes entities in which the target holds less than a controlling interest — including the most commercially active JV in the portfolio.
Non-compete provisions present the sharpest version of this problem. A seller's agreement may contain a non-compete that purports to bind affiliates. If the definition uses a majority-ownership test, the seller's 30%-owned associate — which operates in exactly the same market — is outside the restriction. That is not a technicality. It is the outcome the parties negotiated, or failed to negotiate. Either way, an AI summary that reads "non-compete extends to affiliates" is affirmatively misleading if the affiliate definition means something narrower than the reader assumes.
Enterprise SaaS, Licensing, and M&A: The Three Contexts Where This Hurts Most
In enterprise SaaS agreements, affiliate definitions govern whether subsidiary entities can access the platform under the master agreement or whether each requires a separate order form. A majority-ownership threshold in the affiliate definition means that a partially-owned joint venture cannot be added to the enterprise license without a new contract. Counsel relying on an AI summary that flags affiliate access rights without parsing the threshold may advise clients that coverage exists when it does not.
In licensing deals — particularly in pharma, technology, and media — the affiliate definition determines whether sublicensing restrictions bind the whole group or only those entities above the ownership threshold. This matters acutely when the most commercially significant activity happens in a partially-owned entity specifically because of a prior structuring decision designed to manage regulatory exposure or third-party approval requirements.
In M&A, the affiliate definition determines which existing obligations follow the acquired entity and which do not. A change-of-control that drops the acquirer's ownership of the target to 49% as part of a carve-out structure may move the entity outside its own affiliate definition, triggering neither the benefits nor the obligations of the affiliate provisions. This is not an exotic scenario.
What a Tool Would Actually Need to Do
Solving this properly requires something that current tools do not implement: cross-reference resolution with definitional inheritance tracking across the full agreement.
A tool that genuinely addresses affiliate-scope risk would, at minimum, do three things. First, it would extract the affiliate definition and represent its parameters — threshold, control test, exclusions — as structured data, not running prose. Second, it would link every affiliate-adjacent provision back to that structured representation and flag where the provision's scope depends on a threshold that may not be satisfied. Third, where the relevant entity's ownership structure is known — because it has been provided as input, because the tool has access to cap table data, or because the M&A process has produced a corporate chart — it would apply the definition to the actual facts and tell you whether the entity in question is in or out.
None of the widely deployed tools do this. The better ones extract and display the affiliate definition alongside flagged provisions, which at least surfaces the definition for human review. That is useful but insufficient. It is the difference between a research assistant who finds the relevant clause and one who actually answers the question.
The Problem With "Trained on Millions of Contracts"
When vendors describe their training datasets, they are usually describing the system's ability to identify where affiliate language appears, not its ability to reason about what that language means given a specific definition and a specific entity. Pattern recognition scales. Definitional reasoning across a variable corporate topology does not scale the same way, and no vendor has demonstrated that it does reliably.
The risk for senior counsel who have begun trusting AI contract summaries is not that the tools hallucinate provisions that do not exist. It is that they accurately report what the provisions say while silently omitting the definitional context that determines whether those provisions apply. That kind of confident, structural omission is harder to catch than an obvious error — which is why it is the failure mode worth worrying about.
Until a tool can show that it has resolved the affiliate definition, parameterised its scope, and applied that scope to the specific entity you are asking about, treat every AI-generated affiliate-risk summary as an observation about the contract's text, not an answer about your client's exposure.