Quick answerWhen a customer states or otherwise makes clear they are recording or livestreaming their interaction with your AI agent, the agent should not change its underlying answers, but it should adjust its behavior in two specific ways: acknowledge the disclosure plainly rather than ignoring it, and apply a slightly more conservative tone-and-precision standard to anything the agent says, since a recorded interaction can be clipped, shared, and judged out of context far more easily than a private conversation. Do not build a policy that tries to detect recording covertly or treats a disclosed recording as adversarial by default; the goal is consistency and care, not suspicion.
This is a different problem than the agent replying in public itself
The policy for whether your AI agent should ever reply in public, in a forum or on social media, treats the agent's own reply as a publication the moment it is posted somewhere visible beyond the customer it was written for. A customer recording a private, one-to-one conversation with your agent is the reverse situation: the conversation itself is still private in the sense that the agent is responding to one customer in what looks like a normal support channel, but the customer has independently decided to capture and potentially publish that private exchange.
The agent cannot, and should not try to, control what the customer does with the recording afterward. What it can control is how it behaves once it knows, or reasonably suspects, that this particular exchange might end up being watched by an audience well beyond the customer who is currently in the conversation.
Acknowledge plainly, do not perform or clam up
When a customer states outright that they are recording or livestreaming, whether they are testing the agent for a video, documenting a complaint, or simply narrating their day, the agent should acknowledge the disclosure in a brief, neutral way, something like noting that it understands the conversation is being recorded and it is happy to continue helping, rather than ignoring the statement entirely or making the acknowledgment the centerpiece of the response. Ignoring a direct statement about recording reads as either not listening or evasive, and either read makes for a worse recorded interaction than a simple acknowledgment would.
Avoid two failure modes at once here: do not have the agent become noticeably more guarded, defensive, or scripted-sounding once it knows it is being recorded, since that shift is itself often what gets highlighted and mocked in a shared clip, and do not have it perform for the camera, becoming unusually chatty, promotional, or eager to please in a way that would read as inauthentic if the same question came from a customer not recording. The right target is the same helpful, accurate answer the agent would give anyone, delivered with a brief acknowledgment of the disclosure and nothing more.
Raise the precision bar slightly, without changing the substance
The practical adjustment worth making is not to what the agent says, but to how carefully it says it: a recorded interaction can be clipped down to a single sentence and shared without the surrounding context that made that sentence reasonable, so err toward slightly more complete, less easily misquoted phrasing when the agent knows it is being recorded. This is a small, general precision adjustment, not a rewrite of the agent's substantive answers, and it should not cause the agent to hedge on questions it would otherwise answer directly, withhold information it would otherwise share, or become noticeably more cautious about topics unrelated to the fact of being recorded.
This differs meaningfully from the customer who tells you they're a competitor, or a journalist, where the disclosure itself signals a specific kind of scrutiny the agent should be aware of but should not treat as grounds for being cagier. A customer recording their conversation is very often just an ordinary customer documenting their own experience, with no adversarial intent at all, so the recording-disclosure policy should be built around care and clarity, not around the adversarial-audience posture that applies to a self-identified competitor or reporter.
Do not build covert recording-detection into the agent
Resist the temptation to have the agent try to infer, from cues in the conversation, whether it is likely being recorded even without an explicit statement, and adjust its behavior covertly based on that inference. This is a bad idea for two reasons: the inference will frequently be wrong, changing the agent's behavior for ordinary customers based on a false-positive guess, and if it were ever discovered that the agent behaves differently based on suspected surveillance rather than disclosed surveillance, that discovery itself would be a worse trust problem than anything a recorded conversation could reveal.
Build the policy strictly around explicit disclosure: the customer says they are recording, or the channel itself makes it obvious (a livestream platform's own interface, for instance). Absent that explicit signal, the agent should behave exactly as it would in any other conversation, which is also the simplest, most auditable version of this policy to actually implement and verify.
This explicit-disclosure-only rule is deliberately narrower than what an AI agent should do when it's being tested, not served, which covers a different, broader category of the agent being probed rather than simply observed. A customer recording a genuine support conversation is not testing the agent in that sense at all, even though both policies ultimately concern an audience beyond the customer currently being helped.
FAQ
Should the agent ever decline to continue a conversation because it is being recorded or livestreamed?
No, declining to help a customer because they disclosed a recording treats a neutral, common behavior as if it were adversarial, which is worse for trust than just answering normally. The only case where declining makes sense is the same case where it would make sense regardless of recording, if the conversation itself has moved into territory the agent is not authorized to discuss.
Does this policy change if the recording is clearly for a negative review or complaint video rather than neutral documentation?
The agent's core behavior should stay the same either way, since trying to distinguish intent behind a recording in real time is unreliable and risks exactly the kind of defensive shift that makes for a worse recorded interaction. If the conversation itself turns adversarial or the customer becomes abusive, handle that through the agent's normal de-escalation and human-escalation paths, independent of whether the interaction is being recorded.
Should customer service or trust teams review recorded interactions that get published?
Yes, treat a published recording of a customer interacting with the agent as useful, real-world QA data, worth a periodic review the same way transcript sampling reviews other live conversations, since it surfaces exactly how the agent performs under conditions the team cannot fully simulate internally.