How Do You Build a Lead-Capture Agent for a Dealership Floor?

I sell cars for a living. Twenty years in automotive retail, no tech background, no computer science degree.

Here is a thing that happens on a dealership floor every single day.

Someone walks on. You greet them. You are already moving them toward a vehicle. And the whole time, a small voice says write this one down.

You don't. Because writing it down means stopping, opening the CRM, and typing a record while a live customer stands there waiting.

So you tell yourself you will enter it after.

After is three more customers, a phone call, and a manager pulling you into a deal. The walk-in is gone. You never wrote the name down.

That lost walk-in is the task. This build log is the agent I built to stop losing it.

What does a lead-capture agent for a dealership floor actually do?

It saves the walk-in before the walk-in walks off.

Not a full CRM record. Not the address, the trade, the credit app. Just enough to follow up, captured in the seconds you have while you greet someone and start moving them toward a car.

The whole job is this. A customer walks on. You say a short note out loud, a new one just walked on, wants a used truck, first name is the name they gave, here is the number. The agent turns that into a small lead and stores it. You keep walking the customer to a vehicle. The lead is already saved.

That is the difference between a lead you captured and a lead you meant to type up later.

Here is the trap, and it is the same trap that eats every floor. Capturing a lead competes with helping the customer, and the customer always wins. If capture costs you a full form, it loses to the live person in front of you, every time. So it never happens.

A lead-capture agent wins by making capture almost free. A spoken note you can say while you keep moving. That is the entire idea.

What fields should the agent capture, and what should it ignore?

Three fields. A name, one contact method, and the vehicle they came for. That is the whole floor version.

A name, even just a first name, so the follow-up is not cold.

One contact method, a phone or an email, so you can actually reach them.

The vehicle interest, what they came in for, so the follow-up has a reason to exist.

Everything else is enrichment you add later. The trade, the timeline, the budget, the exact stock number. All of that matters for a deal. None of it fits in the seconds of a greeting.

Greet the customer

→

Speak a short note

→

Agent pulls 3 fields

→

Lead saved to your list

The reason to cap it at three is not neatness. It is physics. A greeting gives you seconds. If you try to capture the trade and the budget and the timeline in those seconds, you capture nothing, because the customer moves before you finish. Three fields fit. Six do not.

Capture the three now. Enrich the record later, when the customer is sitting down and engaged and you have a minute. A partial lead you saved beats a perfect record you never started.

This is the same lesson as building any first agent. You do not build the whole thing. You build the smallest version that is actually useful, and you ship that. If you want the longer version of that argument, I wrote about where to start building AI agents when you have a day job.

How do you build the smallest useful version, step by step?

Here is the order I built mine in. Each step is small enough to ship on its own.

Step one. Pick the three fields and nothing more. Name, one contact method, vehicle interest. Write those down as the only fields the floor version captures. The discipline is in what you leave out. Every field you add past three is a field that makes capture too slow to happen.

Step two. Capture by voice during the greeting. The agent takes a spoken note the moment you greet the customer. Capture lives inside the interaction instead of competing with it. A one-tap voice note beats a form you promise to fill out after, because on a busy floor the after never comes. If you want the microphone side of this, it is the same one-tap capture I covered in how to capture work between customers without typing anything.

Step three. Turn the spoken note into a small structured lead. The agent pulls the three fields out of what you said into a tiny record. Name here, contact here, vehicle here. Keep the shape dead simple so a missing field is obvious at a glance.

Step four. Write the lead to one destination you already check. Save it to a single place you look at every day. One list. One channel. Not a shiny new tool you will forget by Friday. A lead in a place you already watch gets followed up. A lead in a tool you check once a week is a lead you lost slower.

Step five. Read the lead back before you trust the save. After the agent writes the lead, it reads the record back out and confirms the three fields are there. A silent save failure caught in the moment is a save you can retry. A silent save failure found three days later is a walk-in gone for good.

What breaks when you build it, and how do you fix it?

Three failures, in the order they hit me. This is the part the demos skip.

One. The missing contact. I would speak a note in a rush, say the name and the vehicle, and forget the number. The agent happily saved a lead I could never follow up on. A name and a truck with no phone is not a lead. It is a regret. The fix is to make the contact method required. If the note has no phone and no email, the agent asks for it back before it saves, instead of storing a lead that is dead on arrival.

Two. The wrong vehicle from a fuzzy note. I said wants a truck and the customer had actually pointed at an SUV, and my early parser guessed the wrong one and saved it confidently. Now the follow-up references the wrong vehicle and I look like I was not listening. The fix is to keep the vehicle field as close to what I actually said as possible, and not let the agent get clever and infer a model I never spoke. Capture the words. Do not invent the intent.

Three. The save that never happened. The agent called the save, got no error, and told me the lead was captured. The write had actually failed. I found out two days later when I went to follow up and the list was empty. No error is not the same as saved. The fix is the read-back from step five. The agent does not trust the write. It reads the lead back out of the destination and only tells me captured when the record is actually sitting there.

A partial lead you actually saved beats a perfect record you never started. Capture three fields during the greeting, not ten fields after the customer leaves.

How do you keep this small enough to ship this week?

You capture one lead type first. Walk-ins. Not phone-ups, not internet leads, not service. One.

Pick the walk-in, because it is the one you lose most and the one nobody types up. Give it the three fields. Capture it by voice in the greeting. Write it to the one list you already check. Read it back before you trust the save.

That is the whole v1 of a lead-capture agent, and it is enough to stop losing walk-ins.

• Capture exactly three fields, a name, one contact method, and the vehicle interest.
• Take the note by voice during the greeting, not a form after.
• Require the contact method, and ask for it back if it is missing.
• Save to one destination you already check every day.
• Read the lead back out of the destination before you report it captured.

Run that for a week on walk-ins before you add a second lead type. A week of real greetings tells you whether your read-back actually catches a failed save, which it will, the first time a token quietly expires or a list gets renamed. If you want the deeper version of making sure a save lands where you meant, I built that separately in how to make sure a voice agent sends work to the right destination.

Only after that week do you add the second lead type. The capture shape and the read-back you built first make each new one cheap, because a new lead type is the same three fields and the same verify-and-confirm you already trust. That is the compounding that lets an operator stack grow. Five rivers, twenty-two agents, all built the same way, one small verified capture at a time, on less than fifty dollars a month in hosting.

The reason to build this, if you take nothing else from the log, is that a lost walk-in is not a small miss. It is a customer who came to you ready to look, and left without a trace, while you were busy doing the job. The whole point of a lead-capture agent is that the greeting itself saves the lead, so helping the customer and capturing the customer stop competing.

When the note has to travel from your phone into a saved record, the capture leans on the Model Context Protocol, the open standard for how an AI reaches a tool, so the same spoken note can write to a list or a channel without you wiring each one by hand. And if you want the plain research on why fast follow-up matters at all, the Harvard Business Review study on lead response time is the honest source, firms that reach a new lead within an hour are far more likely to have a real conversation than firms that wait. A lead you captured in the greeting is a lead you can reach while it is still warm.

The community where operators ship these in public, bugs and all, is the free build log at skool.com/agent-empire-4291. Come capture your first walk-in before it walks off. Build log, not a landing page.

While I sell cars for a living.

Michael