The Adolescence of
Dealership AI

Opening

A vendor walked into the GM's office at a five-rooftop dealer group last month with a deck called "AI for your Dealership." Slide three was "Personalized Customer Journeys." Slide seven was "Hyper-Automation at Scale." Slide twelve had a robot icon. The GM sat through forty minutes of it and asked one question at the end. How does the system learn my store. The vendor said they were "exploring integrations." The GM thanked him, walked him out, and went back to a service drive with eighteen ROs open and a BDC drowning in messages from the weekend.

Some version of that meeting happens every business day in this country. The deck changes. The robot icon doesn't.

Most dealers are about to make a decision about AI they don't realize they're making. Some already made it and haven't noticed. This essay is the map I wish someone had handed me eighteen months ago.

Here is the honest version, before I write another word.

“

Dealers are overestimating what AI vendors can do for them. They are underestimating what AI can do in their hands.

The first group will pay too much, too late, for tools that don't quite work. The second group is already selling them the wrong thing and calling it a platform.

Both halves of that sentence matter. If you only hear the first half, you become the skeptic in the back of the 20 Group room who says "we've seen this movie" and waits, and the wait costs you. If you only hear the second half, you become the dealer who signed a six-figure annual contract on a vendor's word, because the demo was good and the sales rep was sharp, and you didn't read the data processing addendum. You will find out which one you were in about two years.

This essay is for the people who want to be neither.

A note on language before we start. You will not see the word AGI in this essay. I do not use it on stage either. AGI is a term invented by AI labs to raise money and frame regulation, and it has nothing to do with what happens between seven and nine on a Tuesday morning on your service drive. Dealerships do not need AGI. They need AI that knows the deal jacket. The rest is somebody else's marketing.

Section 1Where we actually are

We are not in the singularity. We are in the adolescence.

Dario Amodei at Anthropic wrote an essay last year called "Machines of Loving Grace"1 that I would encourage every dealer to read. Sam Altman at OpenAI wrote one too. Both describe a near-future where AI compresses scientific timelines and reorders entire industries. They are probably directionally right. They are also writing for a different audience than you. The question for a dealer in 2026 is not "will AI cure cancer." The question is "will the AI I am about to buy understand my deal jacket." Those are different conversations and the second one is the one that determines whether you have a business in 2030.

So let me describe what is actually happening in dealership AI right now, in stores I have personally walked, not the conference-stage version.

What is working.

Service appointment scheduling and confirmation flows. A well-implemented AI scheduler with clean shop capacity data and a halfway honest connection to the DMS can reduce no-shows by ten to fifteen percent and free up an hour and a half of an advisor's day. I have seen this in production. It works.

Voicemail triage and after-hours lead response. When the BDC closes at seven and the next ten leads come in at eight, nine, and ten p.m., an AI that reads the lead, drafts a reply in the dealership's voice, and either sends it or queues it for morning approval is a real win. Response time is the single largest variable in lead conversion. AI moves it from "next morning" to "two minutes." That math is durable.

Multi-point inspection upsell language. The technician sends a finding to the advisor. The advisor needs to convert that finding into a customer-ready estimate with photos, a clear explanation, and a payment option. The drafting of that explanation is exactly the kind of pattern-matching work AI is good at. Two of the three big inspection platforms have a usable version of this in 2026.

Internal documentation and SOP drafting. The store has a process. The process lives in someone's head. The AI interviews that person for twenty minutes and produces a draft SOP. Then a human cleans it up. This is the highest-leverage AI work in most stores and almost nobody is doing it.

What is sold but not working.

Most "AI BDC" tools sold as standalone products. The pitch is autonomous customer-facing chat that closes appointments. The reality, in most of the stores I have audited, is a noticeable drop in CSI scores, an uptick in complaint calls about "did a robot just text me," and a GM who eventually pulls the plug and blames AI as a category. The technology is not the problem. The deployment is. AI behind the rep works. AI in front of the customer, untuned, with no human handoff trigger, breaks the relationship that is the whole point of the dealership.

Generative descriptions on used inventory at scale. They sound the same. Google figured that out two model updates ago. Most of the dealers who paid for this product saw a brief lift and then a drop. The dealers who write better descriptions, longhand or AI-assisted with a human in the loop, do better than the ones who automated the entire thing.

Most "predictive analytics" dashboards. The math is real. The data feeding the math is not. We will get to that in the next section, because it is the largest hidden problem in dealership AI and nobody is being honest about it.

What is coming in the next twelve to twenty-four months.

Workflow-aware AI. Not chatbots floating on top of the work, but AI that lives inside the deal jacket, the RO, the pay plan, the compliance checklist. The AI knows where the deal is, what step comes next, and what the operator is supposed to do.

Owner-side intelligence. A chief-of-staff layer for the dealer principal that reads the entire store's weekend like a CFO reads a P&L. Not a dashboard. A briefing. This exists in pieces in 2026. By 2028 it will exist as a single layer in stores that have done the underlying work.

Custom-built tools by dealership employees. The most under-discussed shift in this industry. The tools that win the next decade in your store are probably going to be built inside your store, on top of foundation models, by someone you have not promoted yet. We will get to that person in Section 5.

That is the honest state. Real wins, real failures, real near-future. The rest of the essay is about why the gap between "what works" and "what is sold" is bigger than it looks. And it starts with a word the industry has been using wrong.

Section 2AI is not intelligence

There is a word the industry has been using wrong for three years, and the wrongness of it is the whole reason we are stuck in adolescence.

The word is AI.

What dealers actually need is not AI. It is intelligence. Those are not the same thing. AI is a technology. Intelligence is an outcome. A dealership can buy a great deal of AI and end up with no more intelligence about its operation than it had the day before. Most of them have.

Intelligence is what you have when your data is clean, owned, connected, and pointed at decisions someone in the building actually has to make on Monday morning. AI is one of the tools that turns data into intelligence. It is not the only tool. It is not even the most important tool in most stores in 2026. The most important tool is the discipline of knowing what data you have, what data you trust, and what decision the data is supposed to inform.

The vendor in the opening of this essay was selling AI. The GM was looking for intelligence. Neither of them realized they were having two different conversations. That mismatch is happening in every dealership in America, every day, and it is the single biggest reason this industry's first three years of AI spend has produced so little change in the way stores actually run.

Adolescence is the gap between the AI being real and the intelligence being real. The AI is real. I described what is working in the last section. The intelligence is mostly not real yet, because the substrate that intelligence requires has not been built. And the reason it has not been built brings us to the part of the AI conversation that nobody on the conference stage will say out loud.

“

Every AI failure in your store this year was probably a data failure wearing an AI costume.

The vendor told you the AI would book service appointments. It did not. Why. Because the customer records in your DMS have three phone numbers, two of them disconnected, and the AI tried the wrong one. The vendor told you the AI would score leads. It did. Badly. Why. Because the lead source field in your CRM has been miscoded for nine years and the model trained on garbage. The vendor told you the AI would identify in-equity customers for an upgrade offer. It found four hundred of them, three hundred of whom had already traded in eighteen months ago, because the deal history in your DMS was never reconciled with the service drive after the original sale.

This is the load-bearing wall of the entire essay. The dealership AI conversation in 2026 has been almost entirely a software conversation. It needs to be a data conversation. Because data is the substrate of intelligence and software is not.

Walk into any store in the country. Pick a dealership. Pull a sample of one hundred customer records from the DMS. You will find duplicate records under different spellings of the same name. You will find phone numbers that have not been valid since the customer's last carrier change. You will find RO histories that do not reconcile with the deal jacket from the original sale because the deal moved to a different DMS in 2018 and the migration was done on a budget. You will find customers marked active who died in 2019. You will find at least one record where the email field contains "none@gmail.com" because a service writer in 2014 needed to advance a screen and that was the workaround.

This is not a hypothetical. This is every store I have audited in the last three years. Including the good ones.

Why is it this bad. Three reasons.

The DMS was designed for accounting, not intelligence. Twenty years of integrations have been bolted onto a system whose original job was to print a deal jacket and post it to the books. Every CRM, every digital retailing tool, every BDC platform, every appraisal tool, every service scheduler has plugged into the DMS through a different pipe, on a different schedule, with a different definition of what a "customer" is. The result is exactly what you would expect.

The vendors who could fix it will not. Data hygiene is not billable. It is not a feature. It does not demo well. The DMS providers have no incentive to move quickly on this because a messy database is a switching cost. The CRM vendors will not fix it because their job is to write data in, not to clean data up.

Your team does not have the time. Cleaning a DMS is not a project. It is a discipline. It is a daily, weekly, monthly habit of validation, deduplication, reconciliation, and reasonable judgment about which record is the right record. Nobody in the store has been hired to do that work. Nobody's pay plan rewards it. And the most senior person in the building who could lead it is also the busiest.

This is what we mean when we say AI in 2026 is in its adolescence. The technology is real. The deployment is hard. And the hardest part of the deployment is not the AI. It is the substrate. The AI will only ever be as smart as the data you feed it. Feed it 2019 phone numbers, 2014 workarounds, and unreconciled deal histories, and the smartest model in the world will tell you confidently, in fluent English, the wrong thing.

There is a line I have been using in 20 Group rooms for two years.

“

Data is the foundation. AI is the multiplier.

A multiplier on zero is zero. A multiplier on garbage is faster garbage. A multiplier on clean, owned, connected data is a different store entirely. That different store is the one with intelligence. Everyone else has AI.

The dealers who win the next decade are going to spend the next twelve months doing the boring work first. Not because the boring work is glamorous. Because the boring work is what makes the AI work matter at all. The boring work is how you build the substrate that turns AI into intelligence.

Now we can talk about what right looks like when the boring work is done.

Section 3The dream

I started writing a book at the end of 2024 called The Intelligent Dealership, the week computer-use AI first shipped, because I could see that the rate of change in this industry was about to outrun the language the industry had to talk about it. It didn't come out until the end of January 2026. The long version of what right looks like is in there. The short version is this.

By 2028, in the well-run store, the AI handles the scheduling, the parts pre-pull, the loaner coordination, and the multi-point inspection upsell language, while the technicians and advisors do the work that requires hands and judgment. The math on service absorption changes. The dealer principal walks in on Monday morning and gets a short brief from a chief-of-staff layer that reads the entire store's weekend the way a CFO reads a P&L. Pay plans update themselves when a position changes, because the pay plan is a structured object inside the operating system and not a Word doc on someone's desktop. Compliance steps cannot be skipped, because the AI will not let the deal advance until the step is documented.

None of this is science fiction. Every piece of it exists in some form somewhere in the country in 2026. None of it exists together in one store yet. The work of the next thirty-six months is putting it together, on purpose, with the dealer principal in the driver's seat, on top of data that is clean enough to trust. The dream only happens in stores that did the data work first. Hold that prerequisite loosely while you read the next section, because the rest of the essay is about the five specific traps the market is currently walking into with its eyes closed. Every one of them gets worse if the data underneath is broken. Every one of them is a way the industry is currently selling AI without intelligence.

Section 4The five risks nobody is naming

For each risk: what it looks like in the wild, and what the defense looks like. These are not theoretical. I have watched every one of them play out in real stores in the last eighteen months. They are ordered deliberately. The first one is the precondition for the other four mattering.

Every dealership in America is generating training data right now. Every chat transcript, every RO note, every desk log, every recorded service call, every appraisal override, every BDC text thread, every digital retailing session. The question is not whether that data is building intelligence. It already is.

The question is whose.

Right now, mostly not yours. Your CRM vendor's model gets smarter on your customer conversations. Your chat vendor's model gets smarter on your sales objections. Your appraisal tool gets sharper on your trade decisions. The aggregator sitting on top of your DMS gets a clearer picture of how your store actually runs than you do. You get a dashboard.

The data you generated, the data that describes your operation in more detail than any internal report ever has, is being refined into intelligence and pointed back at the industry. Sometimes pointed at your competitors. Sometimes pointed at the OEM that has been waiting fifteen years for the right tools to disintermediate you. Sometimes pointed back at you, as a product you will eventually pay for, made from work you already did.

Every contract you have signed in the last three years has language about this. You probably did not read it. The data processing addendum, the sub-processor list, the model training rights clause, the aggregation and anonymization carve-outs. Those documents are the entire game and most dealers I know have never opened them.

Treat your DMS data the way a bank treats deposits. It is yours. It is also, in a structural sense, the most valuable asset you have that is not on your balance sheet. It is not yours to give away casually because someone showed up with a nice deck. Every contract, every integration, every "free trial" is a deposit decision. Most dealers are making thousands of those decisions a year without anyone in the building treating them as decisions.

Read the DPA. Then read the sub-processor list. Then read the model training rights clause. If you do not have someone in the building who can read those documents, hire a lawyer who can, by the hour, before you sign the next contract. The hour is cheap. The contract is not.

Negotiate the training rights. Most vendors will say their standard contract does not allow exceptions. Most vendors are lying. The dealers I have seen do this well have written addenda that explicitly forbid the vendor from using their data to train models that are then sold to other dealerships in their market. The vendors signed. They will sign for you too if you ask.

Inventory your data flows. Make a list of every system that touches customer data in your store. For each one, write down what data goes in, what data comes out, and where it lives after. Most dealers cannot complete that list. Completing it is the first day of getting control back.

If you do not own the loop, someone else does. And the someone else is not neutral. That is the precondition that makes the next four risks real instead of theoretical.

You sign a contract with an AI vendor. You pay a monthly seat fee. The AI works. Six months in, the vendor swaps out the model under the hood, because the one they were using got more expensive and they need to protect margin. Your appraisal accuracy drops eight percent. Your trade conversion follows it down. Nobody tells you. Nobody can tell you, because the vendor's customer success rep does not know either.

This is rented intelligence. You are paying for an outcome you do not control, on infrastructure you cannot see, running on a model that can change without notice. When it works, you cannot take it with you. When it stops working, you cannot fix it.

The pattern repeats across every category. CRM AI. BDC AI. Marketing AI. Inventory pricing AI. The dealer is renting the result. The vendor owns the production function.

Model transparency clauses in the contract. You have the right to know what model is running. You have the right to be notified before it changes. You have the right to test the new model against the old one before it goes live in your store. Vendors will resist this. Push anyway.

Build internal where the stakes are highest. Anywhere a model error costs you real money, appraisal, pricing, lead scoring, the version that runs on your data, against your workflows, in your environment, is worth the investment. The build is cheaper than you think. The lock-in is worth more than you think.

Name a Dealership AI Builder on your staff who can read the vendor's documentation and call the question. Most stores cannot tell rented from owned because nobody in the building speaks the language. Fix that first. The rest gets easier.

A bot answers a customer question about financing. The answer is plausibly accurate and legally exposed. A bot writes a follow-up text that crosses a TCPA line nobody at the vendor thought about, because the vendor sells in twelve industries and yours is the one with the regulators.

Here is the part most dealers have never had named for them. The DMS does not enforce the deal jacket. Or any other process, for that matter. These systems are not that sophisticated, which is exactly why a DMS will happily accept a duplicate record, an unreconciled deal, a service line with no advisor attached, or a customer file with three phone numbers and no indication which one is real. The compliance the industry counts on has never lived in the software. It has lived in the heads of the people who have been doing the work for twenty years, and in the muscle memory of a handful of templates that get printed in a specific order, and in the F&I manager who notices that step seven was skipped because she has been there since 2009.

That is the inheritance the AI is being layered onto. A category of software that has never actually enforced a process, holding compliance together with seasoned humans and habit.

Most of the AI being sold to dealerships in 2026 does not enforce any real process either. It floats on top of the workflow. It generates content, schedules appointments, drafts emails. It does not know what step seven of a compliant deal looks like. It will let you skip it. It will help you skip it faster. The category that was already compliance theater is being amplified by a new category that adds speed without adding enforcement.

This is the risk that takes a store off the road. Not gradually. All at once, when a regulator or a plaintiff's attorney finds the pattern in your texts or your call recordings or your automated lead responses, and the discovery process turns up a vendor contract that disclaims everything.

Treat compliance as a first-class object in your AI strategy. Every workflow your AI touches needs to have its compliance steps documented, enforced, and audited. If the AI cannot prove the step happened, the step did not happen. That is the standard the DMS never held itself to. The AI you buy is the first chance the industry has to actually hold the line in software. Insist on it.

Insist on audit trails. Every AI action in a customer-facing workflow needs to be logged with a timestamp, the input that triggered it, the output it produced, and the human, if any, who reviewed it. If a vendor cannot produce that log for an action that happened in your store six months ago, they should not be in your store at all.

Map the compliance steps before you map the AI. The dealers who get this right start by writing down every regulated step in every workflow, sales, F&I, service, parts, and then ask the vendor where in their product each one is enforced. Most vendors cannot answer that question. The ones who can are the ones worth talking to.

"The system suggests we price this car at $24,750." Why. The vendor cannot say. The model cannot say. The GM cannot say. The price gets used anyway, because the dashboard is pretty and the math is presumed sound, and three months later the inventory turn is off and nobody can reconstruct why.

This is the version of AI that takes authority away from your operators and gives it to a black box. It is the most seductive failure mode in the category, because the AI seems to be doing the work. It looks like leverage. It is actually abdication.

A good AI tool in your store should be able to do four things on demand, in plain English. Show the inputs it used. Show the comparable cases it learned from. Show the confidence interval on the recommendation. Accept an override from a human operator and learn from the override. If any one of those four is missing, the tool is not advising you. It is replacing you.

Explainability is not optional. If a vendor cannot show you why the model said what it said, in language your GM can understand and your operators can act on, the vendor is selling you a magic trick. You do not run a store on magic tricks.

Override-and-learn. Your operators have judgment. The AI should respect it. Every override a human makes should be a signal the system learns from, not a deviation the system flags. If your operators are constantly overriding the AI and the AI never gets smarter, the AI is not learning from your store. It is enforcing a vendor's idea of what your store should be.

AI as advisor, not authority. The recommendation goes to the human. The human decides. The decision and the rationale get captured. Over time, the AI gets better because it is learning from the people who actually know your store. That is the correct asymmetry. Reverse it at your peril.

The pitch is autonomous customer-facing chat that closes appointments and handles inbound leads. The reality, in most stores I have audited, is a noticeable drop in CSI scores in the months after deployment, an uptick in complaint calls about "did a robot just text me," and a GM who eventually pulls the plug and blames AI as a category.

This is the failure mode I worry about most for the industry, because it is the one that poisons the well. A dealer principal who has a bad experience with customer-facing AI in 2026 will not buy a different AI tool in 2028. They will write off the entire category. And the AI that would actually have helped their store, the AI behind the rep, the AI in the back office, the AI that amplifies the technician and the advisor, will never get installed because the AI that should never have been deployed in front of the customer made the whole category look like a scam.

Karen Hao's interview with the CLA CEO2 got this right. Even at hyperscale, customers still want humans for the moments that matter. The relationship is the product. The relationship is what the dealership has that the OEM does not. Pointing AI at the customer, untuned, with no clear human handoff trigger, is the dealer equivalent of giving away the only competitive moat you have.

AI behind the rep, not in front of the customer. In the workflows where AI is doing customer-facing work, there is a real person who owns the relationship and the AI is helping them do their job faster. The customer experiences a faster, sharper, more personal rep. They do not experience a chatbot wearing a name tag.

Clear human-handoff triggers. Any customer-facing AI in your store has documented criteria for when the conversation gets handed to a human, and the handoff is fast, clean, and contextual. The human is not starting from zero. The human has the entire thread, the customer's history, and a suggested next move.

VIP moments stay human. The trade-in negotiation. The F&I conversation. The complaint resolution. The thank-you call after delivery. These are not workflows to automate. These are the moments your store is paid to be present for. AI can help your people prepare for them. AI does not replace your people in them.

Section 5The economic risk and the Dealership AI Builder

There is an honest version of the economic conversation that very few people in this industry are willing to have on stage. I will have it here.

Some entry-level roles in BDC, finance, and admin are going to compress over the next thirty-six months. Not all of them. Not all at once. But the math is changing on which seats are worth which salaries, and pretending otherwise is not a kindness to anyone.

That is the hard half of the sentence. Here is the other half.

The most valuable person in your dealership in 2026 may already be on your payroll. They are probably under thirty. They have already built a couple of things with ChatGPT or Claude that saved a desk manager an hour a week. Nobody asked them to. They just did it. They currently have a different job title. Service writer, finance assistant, marketing coordinator, BDC lead, whatever. They are bored.

Find them. Give them four hours a week. Name the role, I call them the Dealership AI Builder, give them access to your operational documentation, and give them a problem worth solving. Watch what happens in ninety days.

The dealers spending $200,000 a year on AI vendors that do not understand their store are going to be embarrassed in 2027 by the dealers who promoted the right twenty-five-year-old in 2026. I have watched this happen three times. The pattern is reliable enough that I will say it on the record.

There is a cost-reduction story here too. The serious version is not "AI replaces people." The serious version is "AI changes the math on which seats are worth which salaries, and the dealers who restructure their pay plans around the new math will run more profitable stores than the dealers who do not." The Dealership AI Builder is the highest-leverage hire you will make this decade. They will pay for themselves inside six months. They are also the person who tells you when a $50,000 vendor contract is actually a $5,000 problem you could solve in-house.

Naming the role is half of creating it. Most stores have a version of this person already. They are just not titled, not paid, not given time, and not protected from the rest of their day job. Fix those four things and you have a different store inside a year. You also have the first member of your intelligence team, which is the team that turns AI into something your store can use.

Section 6The path

Mature dealership AI, in 2028, has four characteristics. Use these as a rubric every time a vendor walks in the door. Above all four, treat one precondition as load-bearing.

The precondition: clean, owned, connected data. No AI you buy or build matters if the data underneath is broken. The data work is unglamorous. It is also the entire foundation. If your DMS is a mess, your AI will be a more expensive mess. Intelligence requires a substrate. The substrate is data. Do the boring work first.

Workflow-aware. The AI knows where the deal is, what step comes next, and what the operator is supposed to do. It does not float on top of the work. It lives inside the work.

Owner-controlled. The dealership owns the data, the model behavior on its data, and the right to leave. If the vendor disappears tomorrow, the store keeps running and the intelligence stays with the dealer.

Compliance-enforcing. Every workflow the AI touches has its compliance steps documented, enforced, and audited. The AI helps the store stay legal. It does not help the store cut corners faster. This is the first generation of dealership software with a real chance to enforce process in code. Hold it to that bar.

Employee-amplifying. The AI sits behind the team, not in front of the customer. It makes the people in the building more effective at the work humans are uniquely good at. It does not replace the relationships. It funds them.

Close

We are in the adolescence of dealership AI. Adolescence is not a problem to be solved. It is a stage to be navigated.

The dealers who treat this stage as a rite of passage instead of a product purchase will own the next decade of automotive retail. The ones who keep buying solutions from vendors that do not understand their store will spend the next decade paying for tools that do not quite work, wondering why their margins are compressing while the dealer down the street keeps expanding.

“

The dealer down the street will not be winning because they bought better AI. They will be winning because they built intelligence.

Those are different projects. The first one is something a vendor can sell you. The second one is something only you can build, on your data, in your store, with the people you already employ, on a foundation you control.

The boring work comes first. Map your operation. Clean your data. Name your Dealership AI Builder. Read the DPA. Audit your vendor contracts for data ownership terms. Pick one workflow and build the internal version. Then pick the next one.

That is the adolescence well lived. The rest of it is somebody else's marketing.