Your Next Inventory Manager Might Be an AI Agent

Before I let an AI agent lower the price on a used car, I would ask it a question.

Why has this vehicle not sold?

Show me the comparable units.

Show me the listing.

Show me the leads.

Show me when it actually became ready to retail.

Then tell me what you think we should do.

Because a vehicle reaching 45 days tells me how long we have owned it.

It does not tell me what went wrong.

That is where I think the conversation about AI in used car operations needs to go.

We already have tools that help us price inventory. I want to know whether the next generation can help us investigate the problem, get the right work done, and follow up without someone having to restart the process every morning.

Inventory software already does more than display book values.

vAuto describes ProfitTime GPS as providing guidance across acquisition, pricing, and selling, with recommendations that respond to market and dealership conditions. Its product materials also describe connections to merchandising and reconditioning workflows. [1]

So adding an AI label to a price recommendation is not much of a thesis.

The more interesting question is what happens around that recommendation.

Can the system investigate a weak listing?

Identify information it cannot access?

Prepare the next action for approval?

Check whether that action happened?

Anthropic draws a useful technical distinction: workflows follow predefined paths, while agents can choose their next steps and tools as they work toward a goal. [2]

Inside a dealership, I would judge that flexibility by something practical.

Can it handle a vehicle whose problem does not fit the standard aging report?

The example below is a proposed workflow. It is not a claim that one vendor currently delivers every part of it.

I think the opportunity is moving from software we consult to software we can assign work to.

Take one specific assignment:

Review the vehicles that need attention today. Investigate the likely problem. Bring the manager a recommendation with the supporting evidence.

That assignment crosses several systems.

Inventory has one part of the story.

The website has another.

Recon has another.

The CRM may explain what shoppers said after they inquired.

The potential value is in bringing those pieces together before the manager makes a decision.

And the limitation is just as clear.

An agent cannot investigate information it cannot access. A website scan does not tell it what happened during yesterday's appointment.

If the data is missing, I want that written into the recommendation.

Imagine a used SUV with 45 days in stock and weak shopper activity.

This is a hypothetical example.

The easy recommendation is a price reduction.

Maybe that is exactly right.

But I would want the agent to investigate three things first.

Is the comparison fair?

Same trim? Similar mileage? Comparable condition and history? Similar equipment? Does certification explain part of the price difference?

A cheaper vehicle is not automatically a better comparable.

Has our vehicle had a fair chance to sell?

When did recon finish? When did the photos go live? Is the listing complete? Is the vehicle showing correctly on the channels we use?

Forty five days owned and forty five days properly merchandised are different facts.

Both matter. Time spent in recon still costs the store money.

What happened when someone showed interest?

Were inquiries answered? Did customers object to price, condition, equipment, or something else? Was an appointment lost because the vehicle was unavailable?

Silence in the CRM does not prove that nothing happened. It may mean nobody recorded it.

I would want the agent to distinguish an observed problem from an unanswered question.

Suppose the investigation finds incomplete photos, an equipment omission, and comparable listings that do not clearly support an immediate reduction.

The agent could prepare this recommendation:

Now there is something to evaluate.

The manager can inspect the evidence, challenge the conclusion, and approve a next step.

With the right permissions and integrations, the agent could then create the photo task, draft corrected copy using verified vehicle information, and check whether the changes went live.

If the task is overdue, it could flag it.

If the update fails, it should report the failure.

Sending a task is not completing the work.

That last check is where I would spend a lot of time during a demo.

Eventually, the agent will recommend changing a price.

That is where the conversation gets real.

What can it change?

Who approves it?

Which cost figure is it using?

What happens when the recon estimate changes?

For an initial pilot, I would let it investigate and draft recommendations. The manager would approve any public price change.

Later, I might allow limited actions within written rules, enforced by the software that executes the change.

A minimum price. A maximum adjustment. A limit on frequency. Clear exceptions requiring approval.

Those limits should not depend on the agent remembering a paragraph in its instructions.

Every change should leave a record of the evidence, approval, previous value, and new value.

The dealership needs to know what happened without reconstructing a conversation.

This is the phrase I would push hardest on.

The vendor says the system learns from every sale.

Fine.

What exactly does it learn?

If we lower a price and the vehicle sells tomorrow, that does not prove the reduction caused the sale.

The buyer may have been working with us for a week. A competing vehicle may have disappeared. The salesperson may have finally reached the customer.

And if the system is rewarded only for selling vehicles faster, we should not be surprised when it keeps recommending lower prices.

I would want results evaluated across gross, time in stock, carrying cost, and the store's inventory strategy.

Track what it recommended. Track what the manager approved. Track what actually changed.

Then evaluate whether the pattern supports expanding its authority.

One sale is an outcome.

It is not proof of a better decision process.

1. Pick One Narrow Assignment

Start with a small group of aging vehicles and ask for a daily investigation brief.

Keep the scope specific enough that the manager can check the work.

2. Confirm What It Can Actually See

List its data sources, refresh times, and missing inputs.

Ask the vendor to demonstrate access using an actual vehicle from your store.

Do not accept “we integrate with your systems” as the entire answer.

3. Run It Without Publishing Changes

Have the manager assess the same vehicles independently before reviewing the agent's recommendations.

Compare the findings. Investigate disagreements.

Did the agent catch something useful? Did it invent a fact? Did reviewing its work take longer than doing the work?

4. Measure Execution as Well as Recommendations

Record manager review time, factual errors, useful issues found, and whether approved tasks were completed.

Follow sales and gross over a longer period with a reasonable comparison group. A small pilot can expose workflow problems before it can establish a profit lift.

5. Expand Only What Earns It

If listing checks work, expand listing checks.

If price recommendations remain unreliable, keep them under review.

Trust should be earned by task.

My expectation is that dealers will see more products promising to coordinate work across existing systems.

Some will come from vendors already in the store. Others will be new companies connecting the pieces.

For anyone building in this category, that raises the bar.

A morning summary is useful only if it improves what happens afterward.

Can your system get reliable data? Can it respect the store's authority limits? Can it confirm that an approved action actually happened?

Those are the questions I would ask before discussing how many agents a product has.

And I would still check the tools the dealership already owns.

A fixed reminder or existing workflow may solve part of the problem perfectly well.

Use an agent where the investigation requires judgment and adaptation. Keep the straightforward work straightforward.

Your next inventory manager might include an AI agent doing part of the job.

The part I would assign first is the daily investigation: which vehicles need attention, what the evidence says, and which approved actions remain unfinished.

That could give a good manager more time for the decisions that deserve their attention.

But I would judge the system by the work it completes and the decisions it improves.

Show me the vehicle.

Show me the evidence.

Show me who approved the action.

Then show me what happened.

That is the demo I want to see.

If an AI agent reviewed your inventory before tomorrow's morning meeting, what would you trust it to do on its own, and what would still require your approval?

Follow Rooftop Insider on LinkedIn and join the conversation.

I'm Kamil Grzych.

I work inside automotive retail, with experience across dealership sales, BDC operations, digital retail, customer experience and technology implementation across Toyota, CDJR and Lexus dealerships.

I started Rooftop Insider to examine what is changing inside automotive retail from the perspective of people actually working inside dealerships.

Rooftop Insider is an independent automotive retail publication covering AI, technology, inventory, pricing, digital retail, customer behavior and dealership operations.

The goal is simple.

Separate the signal from the noise and explain what changes actually mean for the people running dealerships.

Rooftop Insider articles are based on independent analysis and real world dealership experience. No company referenced in this article paid for or approved its inclusion.

The vehicle scenario, proposed permissions, pilot design, and outlook are Rooftop Insider's analysis. They are not reported dealer results or a description of one available product.