Saturday, 22:14. An enquiry arrives from a portal on a car you have had for six weeks.
Cheapest?! Contact me.
That is the whole message. No name, no context, three words and two punctuation marks.
On Monday morning it is somewhere in a queue behind forty others, several of which are considerably more appealing to open. Whoever gets to it eventually has a decision to make about how much of a Monday to spend on it, and the honest answer is not much, because there is no way to tell from three words whether this is a serious buyer or somebody idly comparing prices at the end of an evening.
Now consider what a genuinely useful reply to that message would contain.
It would confirm the car is still there. It would establish whether they are local. It would find out what they are actually looking for, because “cheapest” is a signal about a budget rather than about that specific car. It would ask whether there is a car to trade in, and if so, gather enough detail for your team to price it properly: make, model, registration, mileage, condition, photographs. It would say plainly that price is a conversation for a colleague, rather than pretending the question was not asked. And if the exchange warmed up, it would propose a time - Wednesday at 13:00, consistent with your opening hours and your test drive policy - for you to confirm.
Here is the thing worth noticing about that list.
Writing it is trivial. Any capable AI system available today will produce a well-structured, courteous, appropriately warm version of that message in about four seconds, in any language you like.
Almost none of it is possible. Because every useful sentence in it depends on a fact about your dealership that a general-purpose AI system has no way to know.
The category is mis-framed
Most of what is sold into this industry under the heading of AI is sold on the quality of its replies. Natural language. Human-like. Trained on millions of conversations.
All of which is true, and none of which is the constraint any more. Language stopped being the hard part somewhere around 2023. Every serious system on the market can write a good email. Buying one for its writing is like buying a calculator for its arithmetic - you are paying for the part that became free.
What did not become free is knowing things.
| THE REPLY NEEDS TO KNOW | WHERE THAT LIVES | AVAILABLE TO A GENERAL AI? |
|---|---|---|
| Which car in your stock this enquiry is actually about | Your live stock feed | No |
| Whether it is still available | Your live stock feed | No |
| Whether a test drive is possible, and on what terms | Your test drive policy | No |
| What your published warranty terms say | Your warranty documentation | No |
| When you are open, including next Thursday's half-day | Your opening hours and holiday calendar | No |
| Which slot to propose for Wednesday | Your schedule | No |
| What trade-in details your team will need to price the car | Your buying process | No |
| Which subjects it must hand to a person | Your rules | No |
| How to phrase all of the above politely | The model itself | Yes, trivially |
Eight rows where the answer is no. One row where it is yes. The industry has built its marketing almost entirely around the last row.
What “relevance” actually means
There is a temptation to treat relevance as a soft quality, a matter of tone. It is not. It is a hard, specific, checkable property, and it is what separates a reply that moves an enquiry forward from one that merely acknowledges it.
Consider the difference in practice.
Reply A: Thank you for your interest in the Volkswagen Passat. We would be delighted to discuss the price with you. Please let us know when it would be convenient to call.
Correct. Courteous. Achieves nothing. The buyer is exactly where they were, minus the seconds spent reading it, and the next dealer to reply - with anything more useful - takes the conversation.
Reply B: Thanks for getting in touch about the Passat - it is still available. On price, that is a conversation for one of my colleagues rather than me, but I can get everything else in place so it is a short one. Two things that would help: are you within reach of the showroom, and do you have a car you would want to trade in? If you do, send me the registration, roughly what it has done and a couple of photographs, and I will pass it to the team so they can come back with a proper figure. We are open Wednesday from 09:00, and 13:00 looks like a good slot if you would like to see the car - say the word and I will get it confirmed.
Same model. Same four seconds of generation. Every difference between them is a fact about the dealership.
That is the whole argument. Relevance is not a writing quality. It is an access question.
Why this is genuinely difficult
If it were simply a matter of connecting a system to a stock feed, everybody would have done it. It is harder than that in three specific ways.
The knowledge is scattered and mostly unwritten. Stock lives in one system, the diary in another, the warranty terms on a page of the website nobody has revised since 2023, the test drive policy in the sales manager’s head, and the fact that you close early on the Thursday before a bank holiday nowhere at all. Assembling that into something a system can use is not a technical problem so much as an organisational one, and it is the part of the work that is consistently underestimated.
It changes. A stock feed that was correct last Tuesday is wrong today. A warranty policy that was updated centrally in a group last month is still being described the old way by half the branches. Any system built on a snapshot degrades from the moment it goes live, which means the design question is not “how do we load this” but “how does this stay true”.
Your own people have the same problem. This is the part worth sitting with. Ask yourself honestly how many of your salespeople could accurately state your current warranty terms, or know what changed on the website last month, or could tell a customer on the phone what your policy is on unaccompanied test drives. In a larger group, decisions get made centrally and communicated once. Everybody means well. Nobody has read it.
So the alternative to a well-grounded system is not a perfectly-informed human. It is a busy person guessing, on a Monday, from three words.
What to ask a vendor
The category is crowded and getting more so. Most of the demos look similar, because they are all showing you the part that is easy - a nicely written reply appearing in a nicely designed inbox.
Four questions cut through it.
What does this know about my dealership specifically? Not what it can be told. What it is connected to. If the answer is a document somebody uploads at go-live, it is a snapshot, and snapshots rot.
How does it stay current when I change something on Thursday? Stock, prices, opening hours, policies. If the answer involves a person remembering to update something, assume it will be wrong within a month.
Can it identify the specific car in my stock that this enquiry refers to? This is a good proxy for whether it genuinely has your inventory or is just parroting the enquiry back. A system working only from the text of the message is not connected to anything.
What does it refuse to do? A system with no refusals has not been thought about. The right answer includes prices, negotiation, discounts, trade-in valuations, financing and interest guarantees, interpretation of warranty or insurance terms, mechanical diagnosis, guarantees of availability, and exceptions to your policies - because those are decisions with money or liability attached and they belong to a person.
Where this leaves the salesperson
Worth being clear about, because the objection lands immediately and it is a fair one.
None of this removes the sales role. It removes the part of the sales role that nobody chose the job for: reading forty ambiguous messages on a Monday, guessing which strangers are serious, chasing the missing details, and discovering three exchanges in that the budget was never there.
What survives is what a good salesperson is actually good at. Standing there with the keys. Reading a person across a desk. Knowing when to suggest a different car and when to stay quiet. Building enough trust that somebody hands over a very large sum of money. Simon-Kucher’s 2025 global study found 82% of buyers prefer to finalise their purchase at the dealership, and that is not a technology lag waiting to be closed. That is what buying a car is.
The gain is that the conversation arrives already warm. The person walking in on Wednesday at 13:00 has already said what their budget is, what they are trading in, and what they are looking for. The salesperson opens with all of it in front of them.
That is a better job than the one that starts with “Cheapest?!” at the top of a Monday queue.
Frequently asked questions
What makes an AI reply to a car enquiry actually useful? Access to the dealership’s own facts: live stock, opening hours, warranty and test drive policies, and a real schedule. The quality of the writing is no longer a differentiator - every capable system writes well. Whether the reply knows anything specific about your business is what determines whether the enquiry progresses.
Is this just an autoresponder? No. An autoresponder sends the same message to everyone and ends the exchange. A grounded assistant conducts a conversation, asks what is missing, adapts to the answers, retrieves facts about your dealership, and hands over an enquiry in a state a salesperson can act on.
Do I have to replace my CRM? Not necessarily. CoDriver can operate as a light lead management system in its own right, or as a layer in front of an existing CRM, delivering qualified enquiries into the system your team already uses.
What should an AI assistant never do in a car sale? Commit to a price, negotiate, offer discounts, put a figure on a trade-in, guarantee financing terms, interpret warranty or insurance conditions, diagnose a mechanical fault, guarantee availability, or make exceptions to dealership policy. Those carry money and liability, and they belong to a person.
Do customers have to be told they are talking to an AI? In the EU, yes. Article 50 of the EU AI Act has applied since 2 August 2026 and requires clear disclosure at the start of the first interaction. CoDriver identifies itself as an AI from the first message.
Sources
- Share of enquiries arriving outside opening hours (58%, peak 21:00-23:00): AutoUncle data, 2026.
- Share of consumers reporting no reply two days after enquiring (26%): AutoUncle lead survey, 2026.
- Preference for finalising the purchase at the dealership (82%): Simon-Kucher, Automotive Shoppers Take the Wheel, global consumer study, 2025.
- EU AI Act Article 50 transparency obligations, applicable from 2 August 2026: European Commission.
