Psychology-first, not automation-first
Most dealership AI is built by automating a process somebody already had. The better order is to work out what actually persuades a nervous buyer, then automate that.
There are two ways to build software that talks to car buyers, and the order you do them in determines almost everything about the result.
The common way is automation-first. Take the drip campaign a store already runs, make it faster, add some personalization tokens, and let a model write the sentences. The output is a quicker version of something that was not working very well to begin with.
The other way is to start with what actually moves a person who is anxious about their credit, and build the automation around that.
What psychology-first means concretely
Rapport before qualification
Every automated system wants to qualify immediately, because qualification produces data and data is easy to measure. But a subprime customer who is asked about income in the first two messages closes the conversation, because it confirms exactly what they were afraid of.
Read the room, then move
Not every customer is at the same rung. A person who is ready to book should be asked to book. A person who is still deciding whether you are safe to talk to should not. Running the same ladder at both is what produces the mechanical feeling customers describe as being processed.
Handle the objection that was actually raised
Price objections are frequently not about price. A customer asking for an out-the-door number is often checking whether you will be straight with them. Answering the literal question and missing the real one is the most common failure in the entire category.
An objection is information about what the customer is afraid of. Treating it as a hurdle to clear rather than a signal to read is why most scripts stop working after two exchanges.
Where the training data has to come from
This is the part that cannot be shortcut. Generic language models are trained on the internet, and the internet does not contain many subprime automotive finance conversations.
What the model needs is volume in the actual domain — millions of real automotive sales conversations, with the objections that actually get raised, the wording that actually lands, and the specific ways deals fall apart. In our case that is more than 8.9 million conversations, and the difference is not subtle. It is the gap between something that sounds plausible and something that sounds like it has done this before.
The rules get stricter, not looser
One counter-intuitive consequence: the better it is at persuasion, the harder the guardrails have to be.
A system that can genuinely build rapport with an anxious buyer is a system that could talk that buyer into something. Which is precisely why quoting a rate, claiming an approval, promising a vehicle and asking what car they want have to be structurally impossible rather than merely discouraged — and why they have to hold even when someone in the thread claims authority to override them.