Hand the rep the answer while the call is still open

Yes, AI can help a rep during a live sales call, but only in one narrow way: when a known objection lands, it puts one line on the rep's screen that has worked before, in a real rep's words, and nothing else. It never speaks, never quotes a price, never shows a claim without a source, never changes who is recording, and stays silent unless it is confident. That narrowness is the product. It is also why this is the last piece of the system to build.

What does the rep see during the call?

Twenty-three minutes into a later call, Dana says what the rep has been waiting for her to say: the sync comes free with the recorder they already pay for. It is the competitor's price, and the price is zero. Marcus, the account executive, has a second or two before the silence gets long.

The card does not invent an answer. It matches the sentence against every earlier recorded call where a buyer raised the same objection about the same product, typed by the extraction from part 2, and joined to how each of those deals ended. The responses are ranked by outcome, deals that closed above deals that did not, and the top one is shown in the words of the rep who said it. That is the same corpus part 9 reads to fix the deck, used while the call is still open.

What makes a live card usable?

• One line, not a paragraph. A rep cannot read prose while someone is looking at them. One move to make and one fallback if the buyer pushes.
• Silent. No chime and no pop-up that steals focus. It is there if they glance and invisible if they do not.
• Their words. The phrasing comes from a rep who said it on a real call and won. Model prose sounds like model prose, and the buyer hears the difference.
• Ranked by outcome. Sorted by what closed, not by what reads well.

What should a live sales assistant never do?

The guardrails are the design. Each one is a hard rule in the build, not an instruction in a prompt.

Two of these come straight from earlier parts. No live pricing is the live version of the boundary in part 3, where the system argues about the amount but never sets it. No claim without a source is the traceability rule from part 7, applied to a card that is on screen for a few seconds.

Why does the moment inside the call matter?

Because that is where buyers now bring what AI told them. Gartner surveyed 645 B2B buyers in August and September 2025 and reported in May 2026 that 69% turn to sales reps to validate AI-generated insights. Gartner's Robert Blaisdell described reps supporting decisions "at critical moments in the journey." A buyer who arrives with a comparison an assistant wrote for them is asking the rep to check it, out loud, in the meeting. The recap that goes out tomorrow is too late for that question.

Why is live assist the last thing to build?

Everything else in this series is forgiving. A wrong CRM field gets corrected before it is saved. A wrong draft never gets sent. A wrong card in front of a live buyer costs the rep the room, and they will never open the assistant again.

So it ships last, behind a confidence threshold high enough that it stays silent for most of the call, and it earns its place by being right about the handful of objections that actually recur. If the only objection it ever handles well is the competitor's price, that is still worth building. For a seat-priced product sold into a team that already owns a recorder, "we already have something that does this" is the objection that decides the deal, and it usually arrives with a price of zero attached.

The threshold logic is the same one that routes inbound email. In the outbound series, a reply the system cannot answer from approved material with confidence goes to a person with the context attached. Live, there is no time to route, so below the threshold the answer is simply nothing.

What does it need before it can work?

• Recordings with outcomes. Calls have to be joined to how the deal ended, which means the CRM record from parts 1 to 7 comes first. Without outcomes there is nothing to rank by.
• Enough instances. A team of four account executives might see a dozen handled instances of its most common objection in a quarter, an illustrative figure. That supports one or two cards, not a library.
• A live transcript. The recorder has to expose the transcript while the call is running. Not every recorder or plan does, so check yours before planning around it.

Who this is not for: a team that records only some of its calls, or cannot yet tie calls to deal outcomes. Build the record and the recap first. Live assist is the reward for having them.

Next in the series

Part 11, The harness: the same webhook twice does nothing twice, goes underneath all of it: how the runtime is built so that repeated events, failed runs and over-eager prompts cannot do damage, and which steps should stop being agents once their shape settles.