How to Evaluate an AI Agent For Your Sales Floor (What It Automates, What It Doesn't)

Every operator I know has the same problem. A real day with a hundred small handoffs, ten of which quietly drop, and a vendor cycle promising AI agents that will fix it. The demos are smooth. The case studies look good. And the question nobody in the room asks out loud is what the thing actually does when a real day at your sales floor starts.

I sell cars for a living. Twenty years in automotive retail. No tech background, no computer science degree. And for the last 18 months I have shipped 22 production AI agents on evenings and lunch breaks that sit in the gaps of my actual operator day. The honest answer to what they do, and do not do, is more specific than most pitches admit.

What does an AI agent actually automate for a sales-floor person?

Not the relationship. That is the starting point.

The genuine value of a good AI agent for an operator is in the capture and route layer. When you finish a phone call at 11:47 AM between two appointments, something needs to remember the follow-up, push it to the right place, and trigger the next step without you opening anything. Most of us cannot do that ten times a day at any consistent quality. A capture-and-route agent can.

The tasks an operator-grade AI agent genuinely handles:

• Voice capture. Speak the work you just finished, the agent transcribes it, parses it, and commits it to the right destination.
• Routing between tools. One prompt reaches Gmail, your CRM, your calendar, your file storage, and back, without you switching tabs.
• Drafting the next message. A reply, a summary, a status note, in your brand voice, ready for one tap of edit.
• Closing routine tickets. A small script that closes a flag, logs a task as done, or moves an item between stages.
• Reporting that runs while you sleep. Weekly pulls of GA4, ad performance, or pipeline activity, formatted for a five-minute scan.

That list is real and it is the unglamorous core of what AI agents do well right now. Most operators are slowest exactly when they are between conversations, which is when the small handoffs go quiet.

What does an AI agent not automate?

The things vendors do not lead with are just as important to know.

An AI agent does not close a deal. It does not handle the customer who has a specific concern about their down payment. It does not replace the trust you have built with your service customers over years. And it does not remove the operator from the loop. Anyone selling you that frame is selling a category that does not exist yet at the operator level.

What you are building or buying is a quiet layer of leverage that compounds across a busy week. That is valuable, and worth doing carefully. But the moment a workflow demands a judgment call, a well-designed agent hands the work back to the human who knows the floor. The test for any agent build is whether it hands off cleanly, or whether it keeps the work in the model layer past the point it should.

An AI agent is not a closer. It is a quiet layer of leverage that compounds, and the test is whether it hands off cleanly the moment a human should be driving.

What integration questions should I ask before I commit?

This is where most builds go wrong. Operator workflows already span four or five tools by the time you start. Email lives in one place, the CRM in another, the calendar somewhere else, the team-chat in a fourth. An agent that cannot reach all of them through one routing layer will break the first time you ask it to do something real.

Before you commit time or money to an agent build, walk this flow on paper:

Phone call ends

→

Capture (voice or text)

→

Parse into structured fields

→

Route to the right tool

→

Human picks it up

The seam where most agents stop is the routing layer. Without something that holds your tools together in one auth surface, every prompt that needs more than one tool breaks. That is the problem the Model Context Protocol was designed to solve, and the boring piece that quietly determines whether your agent stack will run for a year or for a week.

The specific questions worth pressing on before you commit:

• Name your three most-used tools. Can the agent reach all three through one auth flow?
• When a capture happens, where does it land, and how do you see it the next morning?
• When the agent makes a mistake, where does the audit trail live?
• What is the hosting cost at your level of traffic? An honest agent build runs on less than $50 a month for a single operator. If a vendor pitch needs four figures of monthly hosting at your scale, the math is wrong somewhere.
• If you stop using the agent, do you keep your data?

A builder who knows their stack can answer these on the first ask. A pitch will tell you it works and circle back later. Later is when you find out what you built or bought.

How should I measure whether an AI agent is working?

Before you build or buy, write down your numbers. Not estimates. Your actual ones.

If the agent is supposed to capture follow-ups, count how many follow-ups you dropped last month. If it is supposed to route work between tools, count how many minutes a day you currently lose to tab-switching. If it is supposed to draft a weekly report, count the hours that report currently costs you.

Those numbers are the before. Anything the vendor or the build demo shows you is meaningless without them.

When you do your ninety-day review, the test is simple: did your numbers move? Did you drop fewer follow-ups, switch tabs less, or get your weekend back? If the agent's dashboard says yes and your numbers say no, the dashboard is the problem.

What is the right question to start with?

It is not "which AI agent should I build." It is "where in my actual operator day does work go quiet, and at what stage?"

Some operators have a capture problem. The next call ends and the follow-up disappears. A voice-capture agent solves that directly. Some have a routing problem. The work exists, it is just sitting in the wrong tool. A routing layer solves that. Some operators have a writing problem. The drafts go in a queue that never moves. A content agent solves that.

Map the failure point in your real week before you evaluate tools. That is the sequence. Orientation before execution. The work continues to evolve in the field; broader industry context from organizations like the Anthropic developer documentation and the open Model Context Protocol spec is genuinely useful for understanding what the layer underneath your agent looks like. But neither will tell you which problem is yours. That you have to look at directly.

The builders who earn a second conversation are the ones who tell you, honestly, which of your problems an agent solves and which ones it does not. If a vendor or a builder tells you it solves all of them, that is the pitch talking, not the tool.

AI agents at the operator level are real. The gap between what they handle and what they promise is where the evaluation happens. Ask for both sides of that line before you commit.

If you want the full breakdown of how I evaluated and built all 22 of mine, the playbook is in the 10 agents lead magnet. The community where we ship in public is at skool.com/agent-empire-4291. Free.

While I sell cars for a living.

Michael