Few operations have as many systems for an AI agent to cross as a car dealership. The customer asking about the SUV on WhatsApp in the middle of the night has a car in the service department, a consortium installment coming due (consortium plans are a group purchase model common in Brazil) and a used car to trade in. Each piece of information lives in a different system, and almost never where the conversation is happening.
That is why the useful question is not whether AI can talk to the customer. It can. The question is what the agent reads and what it writes in the DMS and the CRM, and what stays out of its reach. For whoever answers for a dealership's operation or IT, that answer is what decides whether the project reaches production.
Where each piece of data lives
Before any agent, draw the map. Each category of data has a source of truth, and an agent that consults the wrong source confidently answers something that is no longer true.
| System | What it holds | The agent reads | The agent writes |
|---|---|---|---|
| DMS | Vehicle inventory, repair orders, parts, invoicing, accounts receivable, vehicle history. | Availability, list price, repair order status, parts, open receivables. | Little: service appointments and, if allowed, parts sales. |
| CRM | Leads, source, funnel stage, assigned salesperson, interactions. | Whether the customer already exists and at what stage. | The lead, a summary of the conversation, the stage, the appointment with the salesperson. |
| Service schedule | Capacity per day and per service. In the DMS or in a separate tool. | Open slots for the requested service. | The appointment, with the car's details. |
| Lenders and banks | Terms and credit analysis. | Installment simulation. | Nothing equivalent to approving credit. |
| Used-car appraisal | Price reference and inspection report for the pre-owned car. | A reference range, if the policy allows it. | The appraisal request. |
| Consortium administrator | Groups, quotas, installments, draws, bids. | Balance, installments, draw, bid result. | Nothing that changes the quota. |
Two questions resolve half the design. Is the CRM a separate system or a module of the DMS? The agent needs to know where each piece of data comes from, or it starts creating duplicate customers. In a group with several stores and brands, is there a single DMS or does each store have its own database? The answer defines how the agent queries the network's inventory and which store it routes the lead to.
What the agent cannot look up, it should not assert. If inventory is not integrated yet, the agent does not say the car is available: it qualifies, books and leaves the confirmation to the salesperson.
Reading can be opened early. Writing is where the risk lives
- Broad reading from the start. Inventory, price, repair order status, parts, open receivables, consortium balance. Looking something up changes nothing, and it is what makes it possible to answer with real data.
- Narrow writing, one action at a time. Creating the lead, logging the conversation, booking a test drive or a scheduled service. Limited effect, easy to undo.
- Writing that stays with a person. A discount outside the policy, reserving a unit, tax registration data, canceling a repair order, a payment agreement outside the policy. The agent prepares and routes; whoever has the authority decides.
Every write needs two protections: no duplicates (if the agent retries, the CRM does not get two leads nor the service department two appointments) and a credential scoped to the minimum the task needs, never the manager's user account. Before promising a deadline, classify what each system exposes (a documented API, database reads or only the screen), because the schedule is that of the most closed system. The details are in how to integrate AI agents with the ERP, the CRM and WhatsApp.
From the first contact to the test drive
A lead that arrives in the middle of the night and only gets an answer the next day goes cold in the queue. The flow that works has seven steps, and each one touches a different system.
- Identify before creating. Whether the lead comes from a listing portal, the website, Instagram or WhatsApp, the agent looks the customer up in the CRM by phone and email before opening a new record.
- Check inventory during the conversation. The unit comes from the DMS while the customer waits, not from a spreadsheet copied that morning. In a group with several stores, the agent says which store has the car.
- Handle the trade-in without promising a value. The agent collects the used car's model, year, mileage and condition and books the appraisal. A reference range, if the store allows sharing one, comes from the appraisal tool, never from an estimate by the AI model.
- Simulate without approving. The installment uses the lender's terms, and the agent makes clear that approving credit is the lender's decision.
- Book with the right salesperson. A visit or test drive on the salesperson's calendar, by store, brand and availability, recorded in the CRM.
- Pass on a ready lead. The salesperson receives what the customer wants, what was looked up and what is still open. If they need to ask everything again, the design failed.
- Follow up with those who did not close. With defined criteria and frequency. On the official WhatsApp, following up more than 24 hours after the customer's last message requires a message template approved by Meta; see AI agents on WhatsApp.
Service department: inbound and outbound scheduling
On inbound, the agent collects the license plate or VIN, the mileage and the service, checks capacity and books. An oil change and a full scheduled service take up the schedule differently, so the agent needs the same time reference per service that the service advisor uses. During the service, the status comes from the repair order in the DMS, and the customer receives the actual stage.
On outbound, the dealership reaches out to the customer when the scheduled service is due. The sale date and the date of the last repair order are in the DMS; the current mileage is not, because the system only knows the mileage from the last visit. The agent estimates from the time elapsed and the history, and confirms with the customer before calling them in.
Parts, collections and consortium
Parts. Stock and price come from the DMS on the spot. When model, year and trim are not enough to identify the part, the agent asks for the VIN or passes to the parts counter: a wrong part becomes a return.
Collections for service and parts. The agent reads the open receivables and collects within a policy set by the store: when to remind, what to offer, how far to negotiate. Outside the policy, a person decides. Whether the customer has already paid is checked at the moment of sending, so nobody who settled that morning gets a collection message.
Consortium. The main data is not in the DMS: group, quota, installments, draws and bids sit with the administrator, and that is where the agent reads from to answer questions and to act on the first installments, where dropouts weigh most. After the award, the agent goes back to the DMS and the CRM, because the awarded credit letter becomes a sales lead within the dealer network itself.
When the agent passes to a person
Sales, service, parts, finance and consortium each have their own queue. The handoff takes the conversation to the right queue, with the full history. The criteria need to be written down, and the text on human-in-the-loop details where to place human approval:
- An explicit request. The customer asked for a person. The agent does not insist.
- Time to negotiate. A lead with a visit booked, or asking for special terms, goes to the salesperson.
- Outside the policy. A discount, term or agreement the policy does not cover.
- Low confidence. A part that cannot be identified, a poorly described defect, a customer not found in the CRM.
- A complaint. A car held past the promised date, a disputed charge. Transferring quickly is worth more than trying to resolve it.
What needs to exist before going live
- Personal data handled with criteria. The financing simulation involves the customer's CPF (Brazil's individual taxpayer ID) and income. Define what the agent asks for and what is masked before it reaches the model; see AI governance and the LGPD (Brazil's data protection law).
- A trail for every action. Lookups and writes recorded with parameters and return, to reconstruct the conversation when the customer says the agent promised a price.
- A written policy. Price, discount, collections policy, handoff criteria. What lives only in the manager's head, the agent has no way to follow.
- Behavior on failure. With the DMS down, the agent says it will confirm and transfers, instead of answering as if it had looked it up.
- An owner after go-live. Price lists, campaigns and inventory change. Someone reviews a sample of the conversations every week.
How to measure
Conversation volume says little. The indicators that matter already exist at the dealership; measure the baseline before turning the agent on.
- Time to first response, by channel and time of day, including nights and weekends.
- Conversion from contact to visit or test drive, and the show rate for appointments.
- Service schedule utilization, and attendance at the scheduled services customers were called in for.
- Amount recovered in collections, and delinquency in the first consortium installments.
- Transfers by reason, because each reason calls for a different fix.
What to ask before signing
- Which of my systems do you already integrate with, and with which read and write operations?
- For whatever has no ready integration, who builds it, how long does it take, and is it in the budget?
- Is inventory queried at the moment of the conversation or from a periodic copy?
- How does the agent avoid a duplicate lead when the customer comes in through two channels?
- In a group with several stores and brands, how do routing and queries to the network's inventory work?
- When the DMS goes down, what does the agent answer, and where do I see that?
- Who looks after the agent after go-live?
How Runflow works with car dealerships
The car dealerships page describes 14 agents for the journey in this text: first contact on WhatsApp, lead qualification, service, parts, insurance, collections, commissions and consortium, from the sale to the post-award journey. The same page lists integrations with Linx DMS, TOTVS DMS, DealerNet, NBS, Salesforce and your CRM, AutoAvaliar, banks and lenders, consortium administrators, WhatsApp, portals and Instagram.
Customer service runs on the Conversation Hub, on Meta's official WhatsApp, with AI and human agents in the same inbox and a queue per department. The platform is for building, operating and measuring each agent: it records every execution reproducibly, sanitizes sensitive data before it reaches the model and allows human approval where the risk demands it. The Forward Deployed engineers integrate the agents with your systems, including whatever has no ready connector. At Osten Group, the result was +30% conversion from contact to visit, with pre-sales service going from 7 people to 2.
