How to make a car dealership AI-native: a playbook

The $1 Tahoe was a stunt and nobody drove away with the truck, but it showed the weak spot: the bot could state a price nobody at the store had approved. The costliest technology failure so far was not AI at all. It was two weeks without CDK, when stores that run deals, repair orders and accounting through one DMS went back to paper.

Meanwhile the economics have turned. New-car gross at AutoNation fell from $1,061.8 million in 2023 to $664.8 million in 2025, and its SG&A rose from 63.4% to 67.9% of gross over the same years[2]. With new-car gross shrinking, more of the store's profit comes from the service drive and the F&I office, and both depend on people answering phones, writing up repair orders, chasing stips and building menus.

The rules have moved too. Since 1 October 2026, in California, the first written reply about a specific car must carry its total price and be kept for two years[5]. The FTC sent 97 dealer groups warning letters on advertised prices in March[7]. Its recent settlements, Lindsay and Manchester City Nissan, are about fees and add-ons charged without consent[10,11]. A lead agent quotes prices and a menu tool lays out add-ons, so both now write the content the FTC is checking.

One way to tell them apart: a buyer says the chat quoted $31,000 out the door and the buyer's order says $33,400. An AI-native store can show, in minutes, which price record the agent used, who approved it, whether the doc fee was in it, and whether the agent was even allowed to state an out-the-door number.

No dealer becomes AI-native by replacing its DMS, CRM or scheduler. It adds agents that work across them: one on the service line, one replying to leads, one checking deal jackets. Each agent needs the store's rules, a manager who approves what matters, limits on which customer data it can see, and a record. The DMS records the deal; it does not record what a chat agent said about the price yesterday, or who allowed it to say it.

OrchKernel is built to be that layer. It connects to the systems the dealer already runs and does not replace them; the DMS stays the system of record. Details are in the OrchKernel blueprint.

NADA counts 16,990 franchised light-vehicle dealerships in the US, with $1.3 trillion of sales in 2025. Service and parts were 13.3% of those sales dollars industry-wide; the average store wrote 16,252 repair orders with 16 technicians and employed 65 people[26]. NADA's free report no longer shows net profit, so the public groups are the best margin evidence. Asbury breaks out its mix clearly:

The service drive and the F&I office carry the store. Together they earned about 71% of Asbury's gross profit on about 18% of its revenue (our arithmetic). New cars, more than half the revenue, earned a fifth of the gross. Warranty labor and parts, paid by the OEM under its claim rules, were about 19.5% of franchised service and parts sales in 2025 (our arithmetic)[26].

Margins are thin at every public group. In fiscal 2025 their gross margins ran from 15.2% to 17.9% and net income from 0.8% to 2.9% of revenue (our arithmetic)[4]. SG&A ate most of the gross:

People are the cost, and most of their pay moves with volume. Salespeople, F&I managers and service advisors are largely on commission; technicians are often paid flat rate. That matters for the AI business case: an hour an agent saves a commissioned BDC rep or advisor does not turn into a saved dollar unless the pay plan, the headcount or the volume changes.

Technicians are the hard constraint. They are 25.2% of dealership employees[26], and BLS projects about 66,200 openings a year for automotive technicians and mechanics[25]. No agent turns a wrench. What agents can do is keep bays full, get inspection findings in front of customers faster, and keep advisors at the counter selling approved work instead of on hold with the phone.

A dealership runs two loops on one customer record and one DMS. The sales loop runs from the lead to the funded contract; the service loop from the appointment to the warranty claim. The trade-in joins them: it becomes a used car that the shop reconditions and the used-car manager prices.

The DMS is both the record and a gatekeeper. In 2017 a federal court found enough evidence of "a per se illegal horizontal conspiracy" between CDK and Reynolds to block third-party data integrators and ordered them to stop blocking Authenticom; CDK settled in 2020 (reported via a secondary source)[37]. In practice, an AI layer reaches dealer data through the DMS vendor's partner programs and APIs, through scheduled reports, or through the CRM. The dealer should own the authorization for every connection and be able to revoke it.

The OEM is a second boss. AutoNation lists "restrictions imposed by, and significant influence from, vehicle manufacturers" as a risk factor[2]. Warranty claim rules, rebate eligibility, co-op advertising and certified pre-owned programs all come from the brand.

Dealers are interested and wary. In Cox Automotive's survey of 537 franchise dealership leaders, 81% said AI is here to stay, 60% were "beginning to test" and about 15% were embedding AI into workflows and decisions; 74% worried about accuracy and errors[28]. CDK's own study says 39% of dealers use AI and another 27% plan to within a year, without showing its sample[30]. Both companies sell AI to the same dealers.

Buyers are ahead. 25% of new-vehicle buyers used AI tools in their purchase, in Cox's fall 2025 survey of 2,300 buyers (vendor-run research)[29]. A growing share of shoppers arrive having asked a chatbot about price and availability first, and will compare that answer with the store's.

The public groups say little. AutoNation's 10-K mentions AI only as a cyber risk[2]. Carvana says it uses AI "to power chatbots that streamline customer communication paths"[3]. None reports an AI productivity number, so every outcome figure below is a vendor claim.

Vendors are named as examples, not recommendations. Most of the traction is on the service phone, which is also where we suggest starting, for reasons the next sections give.

One challenger went after the DMS itself. Tekion claims more than 3,000 dealerships[31]. Asbury had moved 38 stores from CDK to Tekion by the end of 2025, calling CDK's architecture "fragmented 'bolt-on'" and citing Tekion's "unified solution"[1].

Point tools went after the phone and the inbox. Toma, which builds AI voice agents for dealers, raised a $17 million Series A led by Andreessen Horowitz, and Holman Automotive is among its backers[32]. Numa and Impel sell service and lifecycle agents into existing stores[33,34].

The online-native comparison is Carvana. It calls itself "a technology-native retailer," sold 596,641 retail units in 2025 and cut SG&A per retail unit from $5,741 in 2023 to $3,868[3]. The numbers are not comparable with a franchised store's: Carvana sells only used cars and has no franchise service department, while a franchised store's SG&A also carries the shop.

We found no AI-native franchised dealer at scale. State franchise laws and OEM agreements decide who can own a franchise and how it runs, so a new entrant cannot simply open an AI-first Toyota store. The change will happen inside existing stores and groups, which is who this playbook is for.

Six stages. Stage 0 comes first everywhere because every later stage repeats its price and consent records. Service phones come next: the largest share of gross, the clearest bottleneck, and less pricing and credit law per message. Sales replies follow once the price record can carry the total price into every message. F&I comes last because that is where enforcement concentrates. Durations are rough, for a group of 3 to 30 rooftops.

Stage 0 for two rooftops, then Stage 1 on their service lines, then the first sales replies. Ideally one of the two stores is in California, so the hardest price rule is tested from the start.

This is our reading, not a survey finding; we found no study of how dealership jobs change with AI. AI phone and text agents make a central desk practical for a five-rooftop group without a large call center.

Compensation was 43.9% of gross profit at AutoNation and personnel costs 41.7% at Asbury in 2025[1,2], and much of it is commission. If an agent books the service appointment at 11 p.m., does the advisor who writes it up the next morning get the same credit? If a lead agent sets the sales appointment, who is paid on the sale? Settle it before agents go live; a BDC that thinks the agent is taking its spiff will route around it. We found no published model for this.

Asbury chose to move its system of record to a single platform that, in its words, is "expected to make it easier to enhance technology"[1]. Other groups will keep their DMS and connect a control layer to it. Either way the questions are the same: who holds the rules for agents, how they are checked across every tool and rooftop, and what happens to the agents when the DMS is down. The CDK outage showed that one vendor can stop every store at once.

An agent can prepare each of these. A named manager with authority at that rooftop decides, and the decision is recorded.

The rule a dealer's AI meets first is a California pricing law that never mentions AI. The FTC's national CARS Rule was struck down in January 2025, yet 2026 still brought pricing FAQs, 97 warning letters and two settlements over fees and add-ons. Six groups of rules recur below, each followed by what it means for an agent.

The lead-reply log is evidence you must keep under the CARS Act, while the credit application next to it should be disposed of on schedule under the Safeguards Rule. One retention setting cannot do both; set it per record class.

Real cases first, with what each teaches a dealership group.

We found no public enforcement action naming an AI agent or chatbot at a dealer, and no TCPA suit over a dealer's AI voice agent. We do not claim one exists. Every case above turns on a price, a fee, an add-on or a consent record; the scenarios below show how an agent could produce the same problem.

Each is a scenario, not a reported case, and each maps to one of the control points.

OrchKernel is the layer between AI agents and the systems your stores run on, as drawn in the missing layer. Agents send their actions through it; it checks the rules for that rooftop, holds what needs a manager, and records what happened.

What it is not. OrchKernel is not a DMS, a CRM, a phone agent or a pricing tool, and it does not decide prices or credit. The DMS stays the system of record for deals, repair orders and accounting. OrchKernel governs the actions sent through it. A vendor chatbot that answers shoppers on its own, or a CRM feature that sends texts directly, is outside OrchKernel unless those actions are routed through it; deciding which ones must be is part of the Stage 0 inventory.

Where a dealership group needs a control whatever tools it uses, who owns it, and what enforces it.

OrchKernel is source-available under the Business Source License and runs on your own servers, so you can read the code that enforces these controls.

Record your baseline before Stage 1, then track the same numbers at each stage. Public benchmarks exist for service volume, F&I per vehicle and SG&A share of gross. For lead response, contracts in transit, warranty rejections and recon days we found none; NADA's 20 Group data and the mystery-shop scores are private.

Open questions from our research. If you have good data on any of them, write to support@prefero.ai.

Sources were read in October 2026; dates are publication or data dates.