The AI Sales Agent Handoff Boundary
Search for an AI sales agent and you get capability lists. Enrich the lead, draft the first email, chase the follow up, book the meeting, update the record, score the pipeline. Every vendor page carries a version of that list, they are all broadly accurate, and reading three of them teaches you almost nothing.
The list is not the decision. Most of it is table stakes, and the parts that differ between vendors are largely the parts you stop noticing a month after the thing is running. What you will still be dealing with a year later is the question none of those pages opens: where the agent stops.
Sales is not support with a different label, and the difference is exactly what makes this hard. A support agent's worst realistic output is an unhelpful answer. A sales agent's worst realistic output is a commitment. It quoted a price. It agreed to a timeline. It described a feature that does not exist yet. It said yes to a term your contracts people would have said no to. Those are not bad answers, they are statements a customer can reasonably act on, and somebody on your side owns them.
The boundary is a product decision, not a prompt
The reflex is to handle this in the prompt. Tell the model not to quote prices, tell it to escalate anything contractual, tell it to be careful. This feels like a control and it is not one. A prompt is a strong preference expressed to a system that generates text, and the failure you are worried about is precisely the case where it generates text that did not follow the preference.
Enumerate instead, before you pick a vendor, in the same place your team keeps the rest of its commercial rules. For everything the agent might say, decide which of three categories it belongs to.
Free. The agent may say it unattended, in its own words, to anyone. Published positioning, what the product does today, links to documentation, its own availability for a call.
Templated. The agent may say it, but not in its own words. Pricing, terms, service levels, anything a customer will hold you to later. The agent selects an approved block of text and inserts it. It does not compose it, and it does not paraphrase it. This is the category most teams discover late, usually just after an agent has helpfully reworded a price list into something nobody approved.
Escalated. The agent may not say it at all. It recognises the request and hands over. Anything bespoke, anything conditional on a negotiation, anything depending on a date your delivery team has not agreed to, anything a customer would sensibly read as a promise.
The useful property of this shape is that it is enforceable outside the model. Templated means a retrieval step returning a fixed string, not a generated one. Escalated means the only tool the agent holds for that path is "create handoff" rather than "send reply". Neither depends on the model behaving well, which is what makes them controls rather than hopes. The same argument, applied across a whole agent estate rather than to one job, is the governance piece on enterprise agents.
The handoff is the part that actually breaks
Teams that get the enumeration right still lose deals at the handoff, because they build it as an exit rather than as a transition. The agent reaches something escalated, stops, and drops a notification into a queue. From the prospect's side, a conversation that was answering in seconds goes quiet, and the silence arrives at the exact moment they asked the question they cared most about.
Three things make the difference, and none of them is model quality.
The first is that the agent says what is happening. Not a generic apology, but the specific shape of it: that this particular question needs a person, that a person has it, and roughly when. A prospect who knows they have been escalated waits. A prospect who thinks they have been ignored leaves.
The second is that the human receives the context rather than the alert. The transcript, what the agent already committed to, what it declined to answer and why. Without that, the first thing your salesperson does is ask the prospect to repeat themselves, which undoes whatever goodwill the fast part earned.
The third is that the boundary is reversible. Once a person is in the conversation, the agent should not resume on its own and start answering over them. That sounds obvious and it is a common failure, because the agent has no notion that anything changed unless somebody built one.
Measure the boundary, not just the pipeline
The metrics that come with these tools are the ones the tools are good at: replies, meetings booked, records touched. Useful, and they say nothing about whether the boundary is in the right place.
Two numbers you will have to build yourself are worth more than the whole dashboard. How often the agent escalated something it could have handled, which tells you the boundary is too tight and you are paying for an agent while doing the work. And how often it answered something it should have escalated, which is the one that costs money rather than efficiency. The second is harder to see, because nobody reports it. It surfaces later, as a customer quoting something back at you.
Sample the transcripts by hand at first, weekly, a handful at a time. Everyone wants to skip this step and automate the review with another model. That can come later. Early on the point is that a human who understands your commercial rules reads what the agent actually said, because the boundary you wrote down and the boundary you shipped are rarely the same thing in the first month.
What this changes about buying
Read the enumeration back and most vendor comparisons answer themselves.
If a platform cannot return a fixed, approved string for anything in your templated category, it is not usable for pricing conversations regardless of how good its writing is. If it has no first class concept of handing a conversation to a person and staying out of it afterwards, you are going to build that yourself. If it cannot tell you what it said to whom, in a form you can read months later, you have no way to review the boundary at all.
Those three questions cut the field faster than any feature matrix, and they are worth asking before a demo rather than after one. The types of AI agents piece sorts agents by what they are allowed to touch and how long they run, which is the same instinct applied one level up, and what separates a conversational agent from a chatbot covers where accountability lands once a system is genuinely holding a conversation.
Start narrow, on purpose
The version of this that works is smaller than the version in the sales deck. One job, with the free category defined generously, the templated category defined exactly, and everything else escalated on day one. That configuration will escalate too much. That is the correct direction to be wrong in, because an over eager escalation costs a salesperson a few minutes and an over eager answer costs you a commitment you did not make.
Then move the boundary deliberately, one category at a time, on the evidence from the transcripts. Every widening should be a decision somebody made and can point at, not a drift that happened while the prompt was being tuned. If you are still sizing the running cost of any of this, the LLM cost calculator is the quicker way to get a number than a vendor quote.
The capability lists will keep converging, because everyone is building on the same models and the same handful of patterns. The boundary is the part that stays yours, and it is the part nobody can sell you.