Quick answerWhat if the partner relationship is public knowledge already, does the agent still need to say it? Yes. Public knowledge of a partnership existing in general does not substitute for the specific customer, in the specific conversation, being told that this particular recommendation is coming from that paid relationship. Should the disclosure differ if the fee only applies on a completed sale versus a flat referral fee? The disclosure language should stay the same either way; the customer's actual decision-relevant fact is that a financial relationship exists, not the specific mechanics of when your company gets paid.
Quick answer
If money changes hands when your AI agent recommends a specific partner, vendor, or product, the agent has to say so at the moment of the recommendation, not bury it in a footer link the customer never reaches. Treat the disclosure test the same way you would for a human salesperson receiving a commission: state that a financial relationship exists, in plain language, before or alongside the recommendation, not after the customer has already acted on it.
This is a different problem than an honest competitor referral
Our post on when an AI agent should recommend a competitor because they are genuinely the better fit covers a recommendation made with no money involved at all: the agent points a customer elsewhere purely because that is the right answer. A paid partner or affiliate relationship is the opposite case. The recommendation might still be correct and genuinely useful to the customer, but the incentive behind it is no longer neutral, and that incentive is exactly what disclosure law and basic customer trust require you to surface. Conflating the two means teams build a careful competitor-referral policy and assume it already covers the affiliate case, when the two need different controls.
This is also different from marketing your own capabilities honestly
Truth-in-marketing for AI agent capability claims governs what your agent says about your own product. A paid partner disclosure is about a third-party referral relationship: your agent recommending someone else's product or service because that other company pays you for the referral, not about your own capability claims at all. The two controls can both apply to the same conversation without either one covering the other.
What counts as a disclosable relationship
Any arrangement where your company receives money, credit, or a comparable benefit tied to the customer acting on the agent's recommendation counts: a per-referral affiliate fee, a revenue-share partnership, a placement fee for being the default suggested option, or a reciprocal promotion deal where the benefit is non-cash but still real. The test is not whether the fee is large; a small flat referral fee still creates the same disclosure obligation as a large revenue-share deal, because the customer's interest in knowing about it does not scale with the dollar amount.
Where and how to surface it in the conversation
Put the disclosure in the same message as the recommendation itself, in a short, plain sentence the agent always includes rather than a toggleable setting a customer has to know to ask about: something like stating plainly that the company has a paid partnership with the recommended provider. Do not rely on a general terms-of-service clause covering this, since the relevant regulatory and platform-policy tests generally look at the specific moment of recommendation, not whether a disclosure exists somewhere in a document the customer never opened. Log every instance where the agent surfaces a paid-partner recommendation, since that log becomes your evidence trail if a regulator or platform ever asks whether disclosure actually happened in practice, not just in policy.
What happens when the agent is uncertain whether a fee applies
Build a lookup, not a judgment call, into the agent: every partner or affiliate relationship your company holds should live in a structured registry the agent checks before recommending, so the disclosure decision is deterministic rather than left to the model inferring from context whether money is involved. An agent guessing wrong in either direction, disclosing when no fee exists or silently skipping disclosure when one does, is a governance failure in the registry, not a prompt wording problem, and should be fixed by correcting the underlying partner data source.