Index
The difference between a chatbot, an automation and an agent
The four agents that make sense at a dealership
The spec sheet every agent should have
An agent's architecture explained with a real example
What never to delegate to an agent without human confirmation
The autonomy scale: from observing to handling exceptions
How several agents coordinate without stepping on each other's toes
Customer memory vs business memory
Every agent needs a human owner
How to evaluate an agent demo before signing up
More than 750 dealerships already use Dealcar for their daily operations
Frequently asked questions
The difference between a chatbot, an automation and an agent
Understanding the difference between the three concepts avoids buying one thing while expecting it to do another.
A generic chatbot answers questions with predefined replies. If someone asks "do you have an automatic SUV for under 25,000 euros?", the chatbot explains what an SUV is or tells them to get in touch by phone. It does not check stock, does not filter and does not offer alternatives.
An automation always runs the same sequence of steps when a condition is met. If a lead comes in from Coches.net, a task is created in the CRM and a welcome email is sent. The automation does not interpret the context: it does the same thing regardless of whether the lead is from a regular customer, whether the car they are interested in is already reserved or whether it is three in the morning on a Sunday.
An agent receives the same lead from Coches.net and does something different: it identifies whether the customer already exists in the CRM, checks whether the car they are interested in is still available in real-time stock, looks for similar alternatives if that car is already reserved, and decides whether to reply directly, ask a qualifying question or hand over to the salesperson with a summary of what it already knows. Every conversation is different because the agent interprets the context before acting.
The four agents that make sense at a dealership
An independent dealership with 10 to 50 cars in stock has four areas where an agent can handle repetitive work autonomously without compromising important business decisions.
The sales agent handles incoming leads outside opening hours or at times of high demand. It receives the message or form, identifies the contact and their history, checks available stock, presents the cars that match the search, collects the budget and buyer profile, and proposes an appointment. It hands over to the salesperson with the full summary of the conversation when there is price negotiation, a complaint or an unforeseen situation.
Check out the 25 AI use cases for dealerships.
The buying agent monitors sourcing channels according to the dealership's buying policy, detects opportunities that match the stock profile being sought, calculates the estimated margin with all costs included and presents opportunities ranked by profitability. It does not buy: it informs. The buying decision must always be human, using the sheet the agent has prepared.
The marketing agent generates personalised listing descriptions for each vehicle from its data, creates social media content, detects cars with days in stock above the threshold and proposes activation actions. It publishes within approved limits; any content that mentions discounts or special terms goes through human review.
Read also how the AI marketing agent for dealerships works.
The analyst agent aggregates sales, lead and turnover data, detects anomalies compared with previous periods and generates alerts or reports when an indicator deviates from what is expected. It does not make decisions: it informs so that people can make them.
The spec sheet every agent should have
Before being activated, every agent must have seven elements defined. Without that definition, the system's behaviour is unpredictable and errors have no owner.
The goal is a measurable sentence that defines the expected result: "reply to 80% of out-of-hours leads with an appointment proposal in under 5 minutes".
The inputs are the channels and events that trigger the agent: a WhatsApp message, a form on the website, a car that goes over 45 days in stock, a new listing on an auction platform.
The tools are the data it can read and the actions it can carry out: checking real-time stock, reading the customer's history in the CRM, creating an appointment in the calendar, adding a note to the lead's file.
Memory defines what information the agent keeps and for how long: the buyer's preferences during the conversation, the history of previous conversations with that contact, or no information between sessions if the use case does not require it.
The rules are the agent's explicit limits: do not negotiate discounts above X%, do not commit to prices on behalf of the dealership, do not answer questions about specific financing, do not respond outside the defined hours.
Escalation defines in which situations the agent hands over to a person: when the lead asks to speak to someone, when the situation does not match any of the anticipated situations, when there is an active complaint.
The log defines the evidence left after each interaction: the conversation summary in the CRM, the appointment created, the note on the file. Without a log, there is no traceability and no way to improve the system.
An agent's architecture explained with a real example
A sales agent receives a WhatsApp message: "Hi, I'm interested in the Golf you have, is it still available?".
The first step is identification: the agent searches the CRM to see whether that phone number or name matches an existing contact. If it does, it reads their history.
The second step is the lookup: it searches stock for cars that match "Golf". If there are several Golfs, it shows the options with their main features. If there is only one, it confirms its availability in real time.
The third step is interpreting the rules: the message arrives at 22:30 on a Saturday. The agent is allowed to reply and offer an appointment at that time, but not to negotiate the price.
The fourth step is the action: it replies with the availability and features of the available Golf and proposes three visit slots for the following week. It adds a note in the CRM with the conversation summary and creates a task for the salesperson for Monday morning.
What looks like a simple reply is actually five steps: identity, lookup, rule-based decision, action and log. The language model is the piece that interprets the text and generates the reply; around it there is the CRM connector, the stock connector, the permissions system and the action log. Removing any of those pieces degrades the result.
What never to delegate to an agent without human confirmation
There are four types of decisions an agent should not make autonomously even if it technically can.
Final prices and discounts. The agent can state the car's advertised price and can say there is room for negotiation, but it must never commit to a price or discount on behalf of the dealership without a person having approved it. The difference between "the car is 15,900 euros" and "I can do it for 14,500 if you close this week" is a business decision that has to be human.
Final valuations. The agent can generate an indicative valuation based on market data for a private seller's car, but it cannot commit to a purchase price without a physical inspection when the car's condition requires it. An automatic valuation that becomes a firm offer without seeing the car can cause real losses.
See also how the AI buying agent for dealerships works.
Legal and documentation decisions. The agent can summarise the documents needed for a transaction, but it cannot act as a legal adviser or confirm that a specific documentation situation is correct. Those decisions must always be made by a person.
Unforeseen situations. When the agent faces a situation that does not match any of the situations it was configured for, the right response is to hand over to a person, not to improvise. An agent that improvises outside its rules produces unpredictable replies that can compromise the dealership.
The autonomy scale: from observing to handling exceptions
Agents are not born with the maximum level of autonomy. The level of autonomy is expanded progressively according to the quality and traceability of the results shown at the previous level.
The first level is observation: the agent summarises and classifies information (the week's conversations, cars approaching 45 days in stock, unanswered leads) but does not act. It is the starting level for any new agent.
The second level is recommendation: the agent proposes an action (reply to that lead with this information, review the price of this car) but waits for human confirmation before carrying it out.
The third level is execution with confirmation: the agent prepares the complete action (drafts the reply, generates the description, creates the task in the CRM) and carries it out after a person approves it with one click.
The fourth level is execution within limits: the agent acts autonomously in the defined standard cases and escalates the cases that don't fit.
The fifth level is handling known exceptions: the agent autonomously handles the situations it used to escalate because there is now a defined rule for each one, and only escalates genuinely new situations.
Moving up a level requires evidence of sustained quality at the current level, traceability of errors and a way to reverse actions if something goes wrong.
See also how to implement AI in a dealership without stopping operations.
How several agents coordinate without stepping on each other's toes
When a dealership activates several agents (sales, buying and marketing), it needs a shared status system that prevents the agents from acting on the same situation in contradictory ways.
Read also Integrated AI in the DMS vs independent tools.
A concrete example: a buyer asks about a compact car and offers their car as part-exchange. The sales agent understands the intent and proposes an appointment. The buying agent detects the opportunity to acquire the buyer's car and prepares the data needed for the valuation. The marketing agent should not keep sending the same buyer ads for the compact car they have already shown interest in buying.
For that to work, the three agents have to share the status of the conversation in real time: buyer X is in an active sales process with car Y as part-exchange. That status has to be in the CRM, not in each agent's internal memory. If the agents don't share status, each one acts independently and the buyer may receive contradictory communications.
Coordination is written into system statuses, not into instructions for the agents to "understand each other" through free text.
Customer memory vs business memory
Agents work with two types of memory that have different natures and risks.
Customer memory stores preferences, conversation history and agreements reached with that specific buyer. It has privacy limits (LOPD, maximum retention) and must be linked to the customer's file in the CRM, not only in the agent's system.
Business memory contains the dealership's policies, communication tone, opening hours, processes and the rules that apply equally to all customers. It is the agent's configuration, not the customer's data.
Separating the two types of memory is critical to avoid a common mistake: an exception granted to a specific customer (a one-off discount, a special deadline) mistakenly becoming a rule the agent applies to all subsequent customers. Exceptions must be recorded as exceptions on the customer's file, not change the agent's rules.
Every agent needs a human owner
An agent without a human owner deteriorates over time. The rules stop being correct because the business changes, stock or customer data is updated in ways the agent does not anticipate, and errors pile up without anyone systematically correcting them.
The owner of the sales agent is sales management: they review the conversations where the agent escalated, validate that the escalation rules are correct and update the rules when sales processes change.
The owner of the buying agent is whoever makes the sourcing decisions: they validate that the opportunities the agent presents are the right ones and update the buying policy when the target stock profile changes.
The owner of the marketing agent checks that the generated content meets brand standards and that the publishing rules are correct.
The owner of each agent must keep a change log: which rule was changed, why, who approved it and since when it has been active. When conversion changes or a problem appears, that history makes it possible to distinguish a market effect from a configuration error.
How to evaluate an agent demo before signing up
When a provider presents an agent demo, there are five tests that reveal whether the system works in real conditions.
The first is an ambiguous query: "I'm looking for something that isn't too expensive and doesn't use much fuel". A well-built agent asks qualifying questions before assuming. A poorly built agent replies with the first car that fits "cheap" without asking about the actual budget.
The second is to change a car's price during the demo and ask when the agent finds out. If the answer is "at the next feed sync, which is once a day", you need to understand how many out-of-hours leads may receive outdated information.
The third is to present a customer who already exists in the CRM with an active history. The agent has to identify them and use their history, not treat them as a new contact.
The fourth is to request an unauthorised action: ask for a discount above the policy, ask for confidential information. The agent has to refuse clearly and escalate if necessary.
The fifth is to review the final record in the CRM after the demo. The appointment, the note and the lead status must be created correctly, in the right place and without duplicates.
Read also how much it costs to implement AI in a car dealership.
More than 750 dealerships already use Dealcar for their daily operations
Dealcar has AI agents in active development for the sales, buying and marketing use cases described in this article. Features available today include automatic generation of listing descriptions and AI photo enhancement from the vehicle file. Conversational agent capabilities for lead handling and sourcing are on the platform's upcoming roadmap.
You can see how what is available today works at dealcar.io/software-concesionario or request a demo at dealcar.io.
Frequently asked questions
Can an AI agent replace a salesperson at a dealership?
No. The agent autonomously handles repetitive, predictable work: answering availability questions, collecting qualification data, proposing appointments, generating descriptions. Decisions that require negotiation, judgement on exceptions or empathy with complex buyer situations remain with the salesperson. What changes is how much time the salesperson spends on admin tasks vs high-value sales.
How many agents should a small dealership activate at the same time?
One. Start with the most frequent and most measurable use case, and stabilise it for 60 days before adding the next one. Activating several agents at the same time makes it impossible to know which one is generating the return and which one is generating the problems.
What happens if the agent makes a mistake in front of a buyer?
It depends on the type of error. If the agent gives incorrect information about availability or price, the salesperson must be able to correct it immediately with a simple mechanism (taking over the conversation without needing technical steps). If the error is systematic (the agent repeats the same mistake in similar situations), there is a badly defined rule that needs to be corrected in the configuration.
