Quick answerAllow it only after the original support issue is fully resolved, only when the suggestion is directly and obviously tied to something the customer just described, and never more than once per conversation. Require the agent to frame it as a single, easily dismissed suggestion rather than a pitch, and drop it immediately and permanently for that conversation the moment the customer shows any sign of not wanting to hear it.
Why this needs its own rule, separate from negotiating a deal
This is not the same governance question as how far an AI agent should go negotiating price or terms with a customer. That question assumes a deal is already on the table and being actively discussed. This one is about an agent introducing a brand new commercial idea into a conversation that started as something else entirely, like a bug report or a billing question, which raises the pushiness question much earlier, before the customer has signaled any interest in buying anything at all.
The resolution first rule
The single most important gate is sequencing. An upsell suggestion offered before a customer's actual problem is resolved reads as the company caring more about revenue than about the issue in front of it, and customers notice immediately. The agent should complete the support resolution, confirm the customer is satisfied with the outcome, and only then consider whether a relevant suggestion is worth surfacing. Reversing that order is the single most common way this goes wrong.
What counts as directly and obviously relevant
A customer who spent the conversation working around a limitation that a higher tier plan would remove is a legitimate, low pushiness upsell moment, because the connection is obvious and the customer just experienced the gap firsthand. A customer who mentioned nothing about limitations, tiers, or cost is not a legitimate moment, even if a human salesperson might have found an angle. The bar should be that the customer could immediately see why the agent brought it up, not that the agent had to construct a clever justification.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Frequency and exit signals
One mention, once, is the ceiling. An agent that reintroduces the same suggestion later in the same conversation, or brings it up again in a follow up conversation about something unrelated, crosses from helpful into pushy regardless of how relevant the original suggestion was. The agent also needs to treat any hesitation, a change of subject, or a flat no as a hard stop for that conversation, not a cue to rephrase the pitch and try again.
Where this differs from deciding to reach out at all
This is narrower than the broader governance question of whether an AI agent should proactively reach out to a customer first, which covers initiating contact that did not exist yet. An in conversation upsell happens inside a conversation the customer already started, which is a lower bar to clear commercially, but the pushiness risk is arguably higher precisely because the customer did not come in expecting a sales moment at all. Design the guardrails for this case independently rather than assuming your outbound contact policy already covers it.
What happens when the line gets crossed anyway
Even with good guardrails, some customers will feel pressured by a suggestion that was, by policy, well within bounds. Have a clear path for that complaint that does not require the customer to prove intent, similar in spirit to how you should respond when a customer claims your AI agent pressured them into a purchase, since perception matters here as much as policy compliance.
FAQ
Should the upsell suggestion ever be phrased as a question instead of a statement?
Yes, a question invites a genuine no in a way a statement does not, and it is generally the lower pressure framing to require by default.
Does this apply to email or SMS follow ups after the conversation ends?
No, a follow up sent after the conversation has closed is a separate outbound contact decision with its own consent and timing rules, not an in conversation upsell.