AI Sales Call Assistant: I Built One and Ran It on Real Calls

Most sales reps lose deals in the first thirty seconds after a prospect pushes back. Not because they don't know their product, but because there's no one in their ear when the moment actually hits. I built a live-call AI copilot I call Wingman, ran it on real deals with a real account executive, and this is what I learned.

What an AI sales call assistant actually does during a live call

An AI sales call assistant is a piece of software that sits in the background while you're on a call and helps you respond smarter in the moment. Not before the call. Not after the call. During it. That distinction matters more than anything else I'm going to tell you, because every other kind of sales feedback arrives when the deal is already won or already dead.

The way I built Wingman, it runs entirely in the browser. No app to install on the prospect's side, no special phone line, no complicated setup. You open your Zoom or Google Meet like normal, and the tool is listening alongside you.

It listens to both sides in real time

This is the piece people most often get wrong when they imagine AI for sales calls. They picture a bot listening to the rep. What I built listens to both voices, separated by speaker, so it always knows whether the rep or the prospect just said something.

That separation is what makes the coaching useful. If the tool only heard the rep, it couldn't tell the difference between the rep raising an objection and the prospect raising one. The whole coaching logic depends on knowing exactly who said what and in what order. When your prospect says "I need to think about it," Wingman knows that came from the prospect side, not yours, and it fires a coaching card built for that exact moment.

Speaker-separated audio capture is the foundation. Everything else stacks on top of it.

It fires a suggestion the moment the prospect pushes back

When the transcript detects a recognized objection pattern, a card appears on the rep's screen in real time. Not a generic tip. A card with a ready-to-say line, written in plain English, that fits the method the playbook is built on. Below that line are three branches: what to say if the prospect agrees, what to say if they go quiet, and what to say if they push harder on price.

The goal is to get the rep's next five words right. That's it. Sales calls don't fall apart all at once. They fall apart one poorly chosen sentence at a time. Wingman is trying to make those five words count.

The cards are fast. When I timed the gap between the prospect finishing a sentence and the card appearing on screen, I was consistently seeing it under two seconds. That's fast enough to read before you respond. That's the whole game.

What it does NOT do: talk for you

I want to be honest here because I've seen the vendor hype and it gets people's expectations pointed in the wrong direction. Wingman does not take over the call. It does not speak. It does not auto-respond to the prospect. It does not guarantee a close.

It is a copilot, not an autopilot. The rep still has to actually say the words. The rep still has to read the room, adjust their tone, decide whether to follow the card or trust their gut. What the tool does is make sure the rep always has something useful available when they need it. Whether they use it is still on them.

If a rep is going to ignore every card and just wing it the same way they always have, this tool will not save that deal. I'm telling you that upfront because I'd rather set honest expectations than oversell what AI can do on a live call.

This week, write down the three objections that most often end your calls before they should. That list is the starting point for any live-call coaching system worth building.

Why I built my own instead of renting one

There are real-time sales coaching platforms out there. Spiky, Trellus, Avoma, Nooks, Aircall add-ons, Cirrus Insight. I've looked at all of them. I built my own anyway, and the reason comes down to three things: price, fit, and control.

The seat-price problem for a small team

Most of the established platforms charge per seat per month, and the seats are priced for a company that can spread the cost across a whole floor. When you're one account executive or a two-person shop, you're paying enterprise seat pricing for a setup that was never designed for you.

I've seen quotes from these platforms in the range of eighty to a hundred and fifty dollars per seat per month, before you add on the integrations, the onboarding fees, and the tiers that lock the features you actually need behind the higher plan. For a solo rep or a small business owner who closes their own deals, that math doesn't work. You're paying forever for software that will never fit your use case perfectly.

A custom build is a different model entirely. You pay once for the engineering, you own the playbook, and you're not on the hook for a seat fee every month for as long as you want to use it. That ownership compounds over time in a way that monthly subscriptions never do.

Big platforms are built for 50-rep floors, not a 2-person shop

This is the deeper problem. The big platforms are designed around sales manager dashboards, team analytics, call scoring for performance reviews, and integrations with Salesforce workflows that assume you have a Salesforce admin. They are genuinely good tools for the teams they're built for.

They are not built for a software account executive who runs his own deals, or a solo attorney who closes her own clients, or a plumbing company owner who answers every inbound call himself. The configuration options that matter to a fifty-rep floor are just noise for a two-person shop. And the features a small operator actually needs, a single tight playbook, a clean card UI, a follow-up email that writes itself afterward, those are often buried or absent.

When I'm building for a client at Apex Solved, I'm not configuring a platform. I'm building something specific to how that person actually sells. That's a different job, and it produces a different result.

One playbook, custom to the business

The third reason is the one I think matters most. Every coaching system is only as good as the playbook underneath it. A generic playbook built for "B2B software sales" is not the same as a playbook built for a specific rep selling a specific product to a specific kind of buyer.

When I built Wingman for the account executive I'm going to tell you about in a minute, I spent time before I wrote a single line of code listening to how he actually talks. His natural phrases. The objections his specific buyers raise. The order in which those objections tend to appear in a deal. The playbook I built is built around his voice and his market, not around a generic sales template.

The first draft of any playbook is wrong. That's not a flaw, it's just how this works. Real calls tune it. After a few weeks of live use, the cards that aren't working get rewritten and the ones that are working get expanded. The playbook gets better the longer it's in use, and because I built it, I can keep tuning it without needing to pay for a higher tier or submit a support ticket.

This week, record yourself on your next three sales calls (check consent laws in your state) and listen back for the moment you most often lose momentum. That's the first card your playbook needs.

How it works under the hood, in plain English

I'm not going to walk you through the code. But I do want to explain the mechanics at a level where you can evaluate whether this kind of system would work for your situation, because the how matters when you're deciding whether to build or buy.

Two audio channels so it always knows who said what

Most browser-based video calls transmit two audio streams: what's coming in from the other side, and what's going out from your microphone. Wingman captures both of those streams separately, assigns them to speaker labels (rep and prospect), and sends both streams to the transcription layer in parallel.

This is why mic quality matters. A decent USB headset or a quality built-in mic on a newer laptop is enough. A cheap Bluetooth earpiece in a noisy environment will produce a noisy transcript, and a noisy transcript produces late or wrong coaching cards. I tell every client: the tool is only as good as the audio it receives. Get a decent mic. It's a thirty-dollar fix that protects the whole system.

Live transcription as the call happens

The separated audio streams feed into a real-time transcription model. The transcript is rolling and updating continuously as the call happens. It's not waiting for a pause or a sentence to end before it processes. It's running in near-real-time, which is what allows the coaching cards to fire fast enough to be useful.

The transcription layer also handles the speaker diarization, which is the technical term for the process of figuring out which speaker said which words. In plain English, it's the part that keeps the rep's words and the prospect's words from getting mixed together in the transcript. Without clean diarization, the coaching logic can't tell who's raising the objection, and you get cards at the wrong moment or no card at all.

Coaching cards with a ready-to-say line and three branches

On top of the transcript sits the detection layer. This is where I define the objection patterns in plain English: what words or phrases, appearing on the prospect side, should trigger which card. "I need to think about it" triggers the delay objection card. "That's more than we were planning to spend" triggers the price card. "I need to check with my partner" triggers the authority card.

Each card has the same structure:

• The opener: a ready-to-say line the rep can use immediately
• Branch one: what to say if the prospect agrees or softens
• Branch two: what to say if the prospect goes quiet
• Branch three: what to say if the prospect pushes harder on price or digs in

Three branches because those are the three things that actually happen after an opener. The prospect either moves toward you, they pause and wait, or they push back again. Each of those requires a different next move, and I'd rather have the rep see all three options in front of them than have to guess which one applies in the moment.

This week, map out what you actually say when a prospect tells you they need to think about it. If you don't have a consistent answer, that's your first card to write.

The negotiation brain: coaching that follows a real method

A coaching card is only as useful as the method it's built on. I could have built the playbook around a standard objection-handling framework, but I didn't. The method underneath Wingman's cards is drawn from Chris Voss's book Never Split the Difference, which Voss wrote after spending years as the FBI's chief international hostage negotiator and then applying those same techniques to business negotiation. I name the book here because the credit matters: the moves work, and they work because they're built on how people actually behave under pressure, not how salespeople wish buyers would behave.

Tactical empathy instead of pressure

The core move in the Voss method is tactical empathy: naming what the other person is feeling before you try to move them anywhere. Not "I understand you're busy" as a throwaway line, but actually acknowledging the specific concern they raised in a way that makes them feel heard.

When a prospect says "that's a lot of money," most reps jump straight to justifying the price. Wingman's card for that moment opens with a label: "Sounds like the budget's a real constraint on this one." That sentence does something that a price justification doesn't. It tells the prospect that you heard them, that you're not going to steamroll past their concern, and that you're willing to sit in the discomfort of the moment with them. That almost always produces a more honest follow-up from the prospect than a defensive explanation of why the price is fair.

Tactical empathy is not agreeing with the objection. It's not caving on price. It's a technique for keeping the conversation open long enough to understand what's actually going on.

Calibrated questions when the call stalls

When a call goes quiet, when a prospect gives one-word answers and the energy drains out of the conversation, pressure usually makes it worse. The Voss move for a stalled call is a calibrated question: an open-ended question that starts with "how" or "what" and puts the work back on the prospect without confronting them.

"How would you want to handle the timing if the budget came through?" That's a calibrated question. It gets the prospect talking, it's forward-looking, and it doesn't make them feel cornered. Wingman detects when a call has gone quiet based on rep-side talk time and the absence of prospect speech, and it fires a calibrated question card to give the rep somewhere to go.

Why it never loops the same move twice

One thing I was deliberate about in how I built the card logic: if a card fires and the rep uses the opener, and then the same objection comes back in a different form, the system does not fire the same card again. That matters because doing the same move twice in a row signals to the prospect, consciously or not, that you're running a script. When a prospect feels scripted at, they disengage.

Instead, the system tracks which cards have already been used in the current call and defaults to a branch or an escalation card if the same pattern comes back. The escalation card is usually a calibrated question, because when an objection survives two rounds, what you need is more information, not more persuasion.

This week, the next time a prospect pushes back, try a label before you respond with any information. Just name what you think they're feeling and stop talking. Notice what happens next.

What happened on real calls

I want to tell you about the account executive I built this for. He sells software to mid-sized companies, runs his own deals from prospecting to close, and was closing somewhere around twenty percent of his qualified pipeline before we started working together. Good rep, good product, but a few specific moments in the call where deals were bleeding out consistently. I anonymized his details but everything I'm about to share reflects what actually happened.

The objection that used to end calls

The objection that was killing the most deals was not price. It was the authority deflect: "I need to loop in my VP before we go further." That line, when it came up, was ending the call. He'd say something like "of course, totally understand, let me know when you've had a chance to connect" and the deal would go cold. Every time.

The card I built for that moment opens with a label and then a calibrated question: "Sounds like getting her alignment is important before you move. What would need to be true for her to feel good about this?" That question keeps the call alive. It puts the rep in the position of helping the prospect think through the internal sale rather than stepping aside and waiting for an outcome he has no control over.

In the first month of running Wingman on live calls, he closed three deals where the authority deflect came up and he used that card. Before Wingman, the historical close rate on deals where that objection appeared was close to zero.

Talk-time: the number that surprised me

One of the metrics I tracked was talk-time ratio. In his calls before Wingman, he was talking about sixty-eight percent of the time. That's high. The general principle in sales coaching, and the research backs this up, is that the rep should be talking less than the prospect on a discovery or closing call. You learn more when they're talking. You control the frame better when they feel heard.

After six weeks of running Wingman on real calls, his talk-time dropped to about fifty-one percent. Not because I told him to talk less. Because the cards kept redirecting him toward questions and labels instead of explanations and pitches. The behavior changed because the tool was giving him better options in the moment, and he started using them.

Teams using real-time coaching report win-rate lifts in the range of eight to twelve percent in the first three months. His numbers came in at the high end of that range, which I attribute partly to the custom playbook being tuned specifically to his buyers.

The follow-up email that writes itself after the call

Because Wingman has a complete, speaker-separated transcript of every call, I added a post-call layer that uses that transcript to draft a follow-up email automatically. Not a generic "great to connect" email. An email that references the specific concerns the prospect raised, the specific next step they agreed to, and the specific objection that's still open.

He now sends a follow-up within ten minutes of every call ending, and it reads like he spent thirty minutes writing it. That speed matters. The prospect is still in the headspace of the conversation when a follow-up arrives fast. The deal is still warm. He's told me this alone changed how quickly prospects respond to him after a call.

This week, time how long it takes you to send a follow-up email after your next call. If it's more than an hour, that's the gap a post-call AI layer can close immediately.

Deal memory: the part that compounds

The part of Wingman that I think most people underestimate when I describe it is the memory layer. It's not glamorous. It doesn't fire coaching cards. But over time it becomes the most valuable part of the system, because it means every call builds on every call before it.

Every call folds into that client's file

After each call, the full transcript, the cards that fired, which branches the rep chose, the objections that came up, and the outcome of the call all fold into a file for that prospect or client. Not into a generic CRM field that says "call logged." Into a structured file that Wingman can actually read and use on the next call.

The distinction matters because a CRM entry is for the human to read later. The Wingman file is for the system to use in real time. It knows what was said before, so it doesn't suggest moves that already failed with this specific person.

The next call opens where the last one ended

When a second call begins with the same prospect, Wingman pulls their file and the coaching logic starts from where the last conversation ended. If the prospect said they needed to think about it and asked for a case study last time, the card library for this call is weighted toward closing the loop on those specific open items before anything else.

This is the kind of thing that makes a rep seem like they remember everything. They do remember, because the tool remembered for them. Prospects notice when someone walks into a second conversation knowing exactly where they left off. It signals respect and professionalism in a way that rebuilding context on the fly never does.

Pre-call briefing in one glance

Before a call starts, Wingman generates a pre-call brief from the client file. One screen. Under sixty seconds to read. It includes the objections that came up last time, the open questions, the emotional temperature of the last conversation based on the transcript, and the three most likely moves for today's call given where the deal is in the process.

The account executive I work with told me he now looks at the brief while he's waiting for the prospect to join the Zoom. He said it's like reading the notes from a sales manager who was on every call with him. The brief doesn't tell him anything he couldn't find himself if he dug through old notes. It just puts it in front of him at exactly the right moment.

This week, before your next call with a returning prospect, write down the three things you most need to remember from your last conversation. If that takes more than two minutes to reconstruct, a deal memory layer is worth building.

The practice arena: training without burning real leads

Here's a problem I noticed early on: the playbook is only as good as the rep's ability to execute the cards under pressure. Reading a card on screen and actually saying the words in a natural tone in the middle of a live call are two different skills. The first time you try a Voss-style label on a real prospect, it can come out stiff if you haven't practiced it.

Roleplay against a skeptical AI owner

I added a practice mode to Wingman that simulates a call with a skeptical prospect. The AI plays the buyer, raises the objections in the playbook, and the rep has to work through the cards in real time just like on a live call. The AI buyer is not easy. It's designed to push back, go quiet, and escalate to price if the rep doesn't handle the first two rounds well.

The account executive I built for ran about twenty practice sessions before his first live call with the full system active. By session ten, he was hitting the labels and calibrated questions naturally, not reading them robotically off the card. By session twenty, he barely needed to glance at the card because the structure had become part of how he thought about the conversation.

That's the point of the practice arena. Not to drill scripts but to internalize the shape of a good response so that the card becomes a reminder rather than a crutch.

Teach the coach your own style

The practice sessions also feed back into the playbook. When the rep consistently ignores a card and says something else that works better, I flag that in the playbook and write a new card that matches their natural voice. The goal is not to make the rep sound like the playbook. The goal is to make the playbook sound like the best version of the rep.

This tuning process usually takes four to six weeks of live calls and practice sessions before the playbook feels genuinely natural. The first draft is always wrong. The version that exists after six weeks of real use is the one worth keeping.

This week, record yourself handling your hardest objection out loud, as if you were on a call. Play it back. If it doesn't sound like something you'd naturally say, rewrite it until it does. That's the card you need.

What the research says about real-time coaching

I've been talking about one account executive. Let me zoom out for a moment, because the results I saw aren't unusual when you look at what teams using similar systems are reporting.

Win-rate lifts teams are reporting

Teams using real-time sales coaching report win-rate improvements in the range of eight to twelve percent within the first three months of consistent use. That's not a massive jump in a single call. That's the compound effect of getting incrementally better responses to the same objections, across many calls, over time.

Small business AI adoption reached about forty-seven percent in 2025, up from twenty-three percent in 2023. That's not just chatbots and email auto-replies. A growing slice of that adoption is in sales-adjacent tools: call transcription, CRM auto-fill, follow-up drafting. The businesses that are figuring out how to use AI specifically on the sales call itself are still a small fraction of that group, which is part of why this is still an advantage rather than a baseline expectation.

Salesforce reported that ninety-one percent of small businesses using AI report measurable revenue gains. I'd push back slightly on that number in the sense that the tool has to actually fit the workflow and be used consistently for it to produce results. But the directional finding is real: AI that's actually embedded in how you work tends to pay for itself.

Why after-the-call reviews come too late

Can AI coach a live sales call? Yes. And the reason that matters is timing. The dominant model in sales coaching right now is post-call review: you record the call, a manager or an AI listens to it later, and feedback comes in a coaching session the next day or the next week. That's genuinely useful for pattern recognition over time. But it does nothing for the deal you just lost.

The research on learning and behavior change consistently shows that feedback is most effective when it arrives as close as possible to the moment it applies to. A card that fires two seconds after a prospect raises an objection gives the rep an immediate chance to apply better behavior on the same call, with the same prospect, while the stakes are live. A coaching session the next morning is analyzing a call that's already closed or already dead.

The after-call review is still worth doing. But it's a complement to live coaching, not a replacement for it.

This week, look at the last three deals you lost and identify the single moment in each call where the outcome could have gone differently. That moment is what live coaching is designed to catch.

What this costs vs. the subscription route

I want to be direct about the cost comparison because it's the question I get asked most when I describe what I built.

At eighty to a hundred and fifty dollars per seat per month, a single-rep subscription to an enterprise platform costs between nine hundred and sixty and eighteen hundred dollars per year before you touch any add-ons. Over three years, that's nearly six thousand dollars at the low end, for a tool that is never going to be specifically tuned to how you sell.

A custom build is a one-time job. The playbook is yours. The system is yours. The tuning is ongoing, but it doesn't cost more every month just because you keep using it. And because I can have a working version of this live in seven days, you're not waiting months for a build to ship before you can test it on real calls.

I'm not saying the subscription route is wrong for everyone. If you have fifteen reps and you need a manager dashboard and deep Salesforce integration, a platform built for that use case makes sense. If you're one person or a small team and you want something that fits exactly how you work, the custom build math is hard to argue with.

Do AI sales tools actually work? They work when the tool fits the workflow and the rep actually uses it. A tool sitting in a browser tab that the rep never glances at is not a coaching system. A custom-built system tuned to the rep's real objections, practiced in the arena before going live, and refined over six weeks of actual calls, that's a system that works.

The kind of custom build I've been describing is exactly what I do at Apex Solved. If you want to see a live demo or talk through whether this fits your sales situation, reach out directly: aaarhontoulis@gmail.com or call (484) 602-6390. I'll show you the system running on a real call, not a slide deck about it.

This week, price out one seat of any enterprise real-time coaching platform you've looked at, multiply it by twelve, and compare that number to what a custom build that you own outright would cost. The math usually makes the decision for you.

What to do this week

If you've read this far and you're wondering whether this kind of system could work for your sales process, here's the shortest version of what I'd tell you to do right now.

1. List your three worst objections. Not the objections you wish buyers raised. The ones that actually end your calls. Write them down exactly as buyers say them. Those are the first three cards you need.
2. Record your next call (with consent). Listen to the moment you lose momentum. That's the moment a live-call copilot is designed to catch.
3. Price out the subscription alternative. Take whatever real-time coaching platform you've considered, multiply the seat price by twelve months, and decide whether owning a custom build outright makes more financial sense for your team size.
4. Talk to someone who's built one. Not a sales rep for a platform. Someone who's actually constructed the card logic and tuned a playbook on real calls. I've done it and I'm happy to walk you through it. Email me at aaarhontoulis@gmail.com or call (484) 602-6390. The demo is a live call, not a PowerPoint.

If you're curious about other ways AI can fit into your sales and client communication workflow, the Apex Autobots product I offer handles adjacent jobs like lead response, CRM auto-fill, and meeting notes that compound with what a live-call copilot does. They work well together, but neither requires the other to start.

The account executive I built Wingman for is now at a point where he says he'd rather run a call on bad Wi-Fi than without the system active. That's not because the tool closes deals for him. It's because he never has to face a hard moment in a call without something useful in front of him. That's what a real AI sales copilot is supposed to feel like.

Key takeaways

• An AI sales call assistant listens to both sides of a live call and fires coaching cards in real time, not before or after the call, but during the moment that matters.
• Speaker separation is the technical foundation. Without it, the system cannot tell who raised the objection, and the coaching is useless.
• The negotiation method underneath the cards matters as much as the card delivery speed. Tactical empathy and calibrated questions (drawn from Chris Voss's Never Split the Difference) outperform pressure and explanation in almost every objection scenario.
• Teams using real-time coaching report eight to twelve percent win-rate lifts in the first three months. The gains come from consistently better responses to the same objections across many calls, not from one magic close.
• Enterprise platforms charge eighty to a hundred and fifty dollars per seat per month and are built for large floors. A custom build is a one-time engineering job that the business owns and can tune indefinitely.
• The first playbook draft is always wrong. Real calls tune it, and the version after six weeks of live use is the one worth keeping.
• Deal memory compounds over time. Every call adds to the client file, and the pre-call brief puts the relevant history in front of the rep before the next call starts.
• The practice arena lets the rep internalize the cards before going live on real deals. The goal is for the card to feel like a reminder, not a script.
• After-call reviews are useful for pattern recognition but arrive too late to change the outcome of the deal you just lost. Live coaching catches the moment while the stakes are still live.
• If you want a custom-built live-call copilot tuned to your specific buyers and objections, I can have a working version running in seven days.