How to Evaluate an Automotive AI Vendor

AI is a label, not a capability. Here is how to find out what a vendor's software actually does before you sign for it.

Every automotive vendor is selling AI right now, and the word covers products that have almost nothing in common. Under the same three letters you will find rules based automation that has existed for fifteen years, statistical scoring models, speech to text transcription, large language models writing messages, and offshore staff typing replies in a chat window. All five can be described honestly as artificial intelligence by a marketing department. Only some of them do what you think you are buying.

So the first move in any evaluation is to stop asking whether a product has AI and start asking what happens mechanically. When a lead arrives at two in the morning, what fires? Is it a template with merge fields, a model generating new text, or a person? When it decides which lead to work first, is that a trained model, a point system somebody configured, or the order they came in? When it summarizes a phone call, is it transcribing audio and summarizing the transcript, or is a person listening?

Vendors who have built something real answer those questions immediately and in detail, because they are proud of the mechanism. Vendors who have not will redirect to outcomes, case studies and language about learning from your data. That redirect is information. It is not proof of fraud, and plenty of useful automotive software is rules based and works fine, but you should know which thing you are buying before you price it.

The second move is to ask what it does badly. Any vendor who cannot name a weakness in their own product either does not understand it or is not being straight with you.

Take these into every demo, including ours. Write down the answers, because a vendor comfortable being pinned down in writing tends to be the one still worth having in year three.

What exactly is automated, and where does a human step in? Ask for the specific handoff rule. If a customer says something the system does not understand, what happens? If a customer asks for a price the store has not authorized, what happens?

Does it write messages that are sent without review? This is the single most consequential question for a generative product. A system that drafts for approval and a system that sends autonomously carry very different risk, and vendors blur the distinction routinely.

What did it train on, and does my data train it for other dealers? Get this in writing.

What does it do when it does not know? The right answer is escalate to a person with a clear signal. The wrong answer is confidently make something up about your inventory, your pricing or your financing.

Can I see the log? Every message sent, every call handled, every decision taken, with timestamps, in a form a manager can review. If you cannot audit it, you cannot manage it.

What is the cost at volume? Per lead, per conversation, per minute and per seat pricing behave very differently in a good month. Model the busy month, not the average one.

What happens to my data if I leave? Ask before you sign, not during your exit. Our vendor question list covers the wider contract side.

A vendor demo is a rehearsed path. You learn very little from watching software handle the situation it was built to handle. Insist on the ugly cases, and bring them yourself so they cannot be prepared for.

Give it a customer who asks about a vehicle that sold yesterday. Give it a customer who says stop texting me but keeps replying. Give it someone who asks what their trade is worth with no vehicle details, then changes the details halfway through. Give it a message with two questions in it, one of which is about financing. Give it a wrong number that answers politely.

Watch what the product does at those moments, because that is what your customers will experience. In particular, watch whether it invents. A system that cheerfully quotes a payment it has no basis for, or confirms a vehicle is in stock when it is not, will do that to a real shopper at scale and you will be the one apologizing.

Then ask to see it running on a dealership that is not a reference account. References are curated. If a vendor cannot show you the product working somewhere ordinary, you are looking at a pilot dressed as a platform.

One more test that costs nothing: send the vendor's own after hours lead form a question at eleven at night and see what comes back. Companies selling responsiveness are often quietly bad at it themselves.

Your customer database is one of the more valuable things your store owns, and AI contracts are where it quietly changes hands. Three specific things to settle in the paperwork rather than the sales call.

Ownership. The contract should say the data is yours, unambiguously, including conversation history and recordings. Vague language about a license to use for service improvement can cover a lot of ground.

Export. Ask what you can take with you, in what format, and whether it includes communication history rather than just contact records. A CSV of names and phone numbers is not your history. Test the export during the pilot, not at the end of the relationship. There is a working list at CRM data export checklist.

Secondary use. Does your data improve a shared model that competitors also benefit from? There is no single right answer here, but there is a right way to find out, which is asking directly and getting it in writing.

Also worth confirming: where the data physically sits, who at the vendor can access it, and whether subprocessors are involved. Under the FTC Safeguards Rule you carry obligations around vendor oversight, and a vendor who cannot answer basic questions about their own security posture is making that harder for you. The related reading is at FTC Safeguards Rule compliance and dealership data governance.

Automated outreach at volume is exactly where compliance problems become expensive, and the risk lands on the dealership, not the vendor. Work through these before the contract, because retrofitting them is painful.

How is texting consent captured, stored and evidenced? Can you show, for a specific customer, what they agreed to and when? Are opt outs honored instantly and across every channel and every rooftop, or only in the system that received them? A customer who opts out at one store and gets a message from another store in the same group is a real problem.

How does the system identify itself? Several states regulate disclosure when a consumer is talking to an automated agent rather than a person, and the rules are moving. Ask the vendor how their product handles disclosure and how quickly they can change it when a rule changes.

For call handling, ask how recording consent is managed in two party consent states, and whether the announcement is configurable by location.

On the financing side, campaigns cannot narrow an audience by age, gender, income, marital status, household size, education, language or ZIP. If a vendor presents targeting around those as an advantage, they are handing you liability and calling it a feature. Treat that as disqualifying information about the vendor. Related: TCPA compliance for dealership texting and the AI compliance checklist.

Finally, write your own internal rules before deployment rather than after the first incident. A short policy covering what the system may say, what it must escalate, and who reviews the log weekly, is worth more than a long one nobody reads. There is a starting point at dealership AI policy template.

It would be inconsistent to write a page about vendor honesty and then be vague about our own product, so here is the specific version.

Call transcription. Calls recorded in the platform are transcribed, which means a manager can read what happened on a twelve minute call in about thirty seconds instead of listening to it. This single capability changes coaching more than anything else on this list, because reviewing calls stops being a chore nobody has time for.

Document reading and VIN scanning. Document AI reads uploaded documents and scans VINs from a phone camera, so a driver licence, an insurance card or a VIN plate becomes structured data without retyping. On mobile that means a salesperson standing next to a car can capture the vehicle in a few seconds.

Then there is a large amount of automation that is genuinely useful and is not AI, and we would rather say so: follow up processes, drip sequences, task automation, lead distribution rules, appointment reminders, IVR and call routing. These are deterministic. They do what you configured, every time, which for compliance sensitive outreach is a feature rather than a limitation.

What we do not sell is an autonomous agent that holds unsupervised sales conversations with your customers and speaks for your store without review. If that is your requirement, evaluate the vendors who build it, and take the questions above into those meetings. Our take on that trade off is at human BDC versus AI BDC.

Pricing models in this category are unusually varied, and the differences only show up once volume moves. Model three scenarios before you sign: a slow month, a normal month, and your best month of last year.

Per conversation or per lead pricing punishes you exactly when things go well. Per minute pricing on voice punishes long calls, which are often your best calls. Per seat pricing is predictable but pushes stores to under license and share logins, which destroys your reporting and creates a security problem. Flat platform pricing is the easiest to plan around and is usually the most expensive at low volume.

Ask what the price is at renewal, not just for year one. Ask what triggers an increase. Ask whether implementation, training and integration work are included or billed. And ask what happens if you need to pause, because seasonality is real in this business and a vendor with no answer for a slow February is telling you something.

For what it is worth, our own pricing is deliberately boring: month to month with no long term contract, CRM Only from $199 a month, programs including exclusive local leads from $799, and a Skip A Month option from $199 for exactly the seasonality problem above. Full figures on the pricing page. If you want to model the return side before committing, the AI ROI calculator walks through illustrative math you can replace with your own numbers.

Most evaluations end in an argument between the person who liked the demo and the person who did not. A pilot ends in a number. Design it before you start or it will drift.

1. Pick one team or one lead source, not the whole store. Point a defined slice of flow at the new system and leave everything else alone so you have a control group.
2. Write down the three metrics before day one. Time to first human contact, follow up touches per lead before the trail goes cold, and appointments set that actually show. Everything else is secondary.
3. Capture a baseline for the same three metrics on the same period last month. Without it you will be comparing feelings.
4. Read the logs weekly. Have a manager read fifty actual conversations, not a summary report. This is where you find out what your customers are being told.
5. Test the export in week three, while you still have leverage.
6. Decide on a written threshold. Agree in advance what result means expand and what result means stop.

Expect a productivity dip in the first two weeks whatever you buy, because people are relearning habits. Judge the trend, not the first Monday.

None of these are automatically disqualifying on their own. Two or three together usually are.

A vendor who will not put capability claims in writing. A vendor who cannot show you an audit log. A vendor who requires a multi year commitment before you can evaluate, which is asking you to take a risk they are unwilling to take themselves. A vendor whose references are all in one market or all provided by the same regional manager. A vendor who claims a specific lift percentage across all dealers, since results depend on your market, your pricing and your people, and nobody can honestly promise a number. A vendor who offers to help you target financing offers by income or ZIP.

And the quiet one: a vendor whose product cannot be turned off by a manager at the store level. If a general sales manager cannot pause automated outreach in under a minute when something goes wrong on a Saturday, you have bought a system that outranks your own people.

We would rather you use this list against us too. If we are not the right fit, say so early and we will tell you honestly. Tell us what breaks today and we will give you a straight answer about whether we solve it.

What is the single best question to ask an AI vendor?

What happens when the system does not know the answer. The right response is that it escalates to a person with a clear signal. The wrong response is that it generates something confident about your inventory, your pricing or your financing terms.

Should AI messages send without human review?

That is a risk decision, not a technical one, and it should be made deliberately. At minimum, insist on a full audit log, a manager level off switch that works in under a minute, and written rules about what the system may say and what it must escalate.

Does LeadLocate sell an autonomous AI sales agent?

No. We provide call transcription, document reading and VIN scanning, plus a large amount of deterministic automation: follow up processes, drips, task automation, lead distribution rules and IVR routing. We do not sell an agent that holds unsupervised conversations for your store.

How do we know if the results came from the software?

Run a control. Point one team or one lead source at the new system and leave the rest on your current process for the same period, then compare time to first contact, follow up depth and appointments that show. Without a control you are measuring the season.

What compliance issues come with automated outreach?

Consent capture and evidence, instant opt out handling across every channel and rooftop, disclosure that a consumer is talking to an automated agent where required, recording consent in two party states, and fair lending rules on any financing message.

Can a vendor promise a specific lift in appointments?

No vendor can honestly promise a number, ours included, because results depend on your market, your pricing, your inventory and how fast your people respond. Treat a specific promised percentage as a reason for more scrutiny rather than less.

Ask us what is automated, what is deterministic, what we do not do, and what you can export. We will answer all four in writing. Month to month, no long term contract.