Ask a dealer principal whether they use AI and the answer is now almost always yes. Ask what that means in practice and it is usually the same thing: somebody opens a chat window, types a question, reads the answer, and gets on with their day.
That is a real use of AI. It is also the first rung of five, and the gap between it and the top of the ladder is not a matter of degree.
“It can be difficult to understand, especially if you’re mostly accustomed to AI chat and not to the agentic way we’re using AI.” — Johan F. Schjødt, co-founder of AutoUncle This matters commercially, not just conceptually. A dealership standing on rung one is being sold products designed for rung four, and is evaluating them against the only experience it has. That is like judging a delivery service by how good you are at driving. The comparison does not hold, and the purchase goes wrong for reasons that have nothing to do with which vendor was chosen.
So before the next demo, it is worth knowing which rung you are actually on.
The five rungs
Rung 1 - Asking. You open a window, type a question, read the answer, and move the result somewhere yourself. You are the transport between the AI and your business. Almost everything described as “we use AI” is this.
Rung 2 - Teaching. Same posture, better inputs. You have stopped starting from nothing: you have given it examples of your own work, saved a profile it applies every time, told it who buys what in your shop. Niels Kristian’s ask it to interview you method sits here, and so does the listing approach in Buyers can tell - thirty of your own best descriptions instead of a spec sheet pasted into a box. The output improves dramatically and the position is unchanged: you still open it, you still carry the result.
Rung 3 - Connecting. Now it can see things. Your stock, your opening hours, your warranty terms, your calendar. This is a large step and an under-appreciated one, because it is where the answers stop being plausible and start being true. It is also mostly not a technology problem: the knowledge is scattered across systems and heads, and it changes on a Thursday. Anyone can reply to an enquiry sets out why this, rather than writing quality, is the actual difficulty of the category.
Still, on rung 3, a person starts it.
Rung 4 - Delegating. Nobody opens it.
That is the whole definition, and everything else follows from it. The system is triggered by an event rather than by a person - an enquiry arrives, a lot appears at auction, a lease expiry comes into range - and it acts, and then it hands something to a human who was not present when it happened.
Rung 5 - Chaining. Several agents handing work to each other across a longer process. This exists and it works, and the honest advice for most dealerships is that it is not the next step. A single delegated process, running properly, is worth more than three half-built ones.
| Rung | Who starts it | What it looks like in a dealership | Where dealers are |
|---|---|---|---|
| 1. Asking | You, every time | Pasting a car's specification into a chat window and asking for a description. Typing "which channel is best for used electric cars". You copy the answer out yourself. | Most |
| 2. Teaching | You, every time | Thirty of your own best descriptions saved as a profile it applies to every car. A session where you asked it to interview you about your marketing spend instead of instructing it. The heat-pump question you know Northern European buyers ask about a southern-market EV, written down where it can use it. | Some |
| 3. Connecting | You | It can name which car in your stock an enquiry is about, and whether that car is still there. A lot's registration and inspection history turned into a one-page digest with a maximum bid attached. A listing's photographs checked for a trim level the description never mentioned. | Few |
| 4. Delegating | An event | Your buying box held against every auction platform overnight, with only the matches landing in Saturday's calendar. An enquiry answered at 22:40, the missing details gathered, Wednesday at 13:00 proposed for you to confirm. Recurring registration and tax paperwork prepared for a signature. | Rare |
| 5. Chaining | An event | Auction monitoring handing to research, research handing to a market read, a bid recommendation arriving as one chain. Nothing published in this series goes here yet. | Almost none |
The useful question is not “how do we get to rung 5”. It is “which rung are we on, and is the thing we are being sold one rung up or three?”
What actually changes at rung 4
Two things, and the second is the one nobody selling this will lead with.
The first is the obvious gain. A delegated process is present when your business is not. It works at 22:47 on a Saturday because that is when the enquiry arrived - evening is when consumers do this, with the peak falling between 21:00 and 23:00 and well over half of enquiries landing outside opening hours entirely.
But presence is not the biggest change, and this is where most vendor pitches stop.
“A lot of this can easily be replaced and done very consistently to the satisfaction of the consumer.” — Johan F. Schjødt, co-founder of AutoUncle Consistency, not speed. A capable salesperson handling an enquiry on a quiet Tuesday morning asks the right questions and answers in full. The same person on a Friday evening with a heavy week behind them asks fewer and skims. Same competence, different output, and nobody records the thin reply as a loss. A delegated process asks the same questions of the promising enquiry and the three-word one, every time. For most dealerships that is the larger prize, and it is invisible, which is why it is undersold.
And the part that changes for the worse
Here is the thing worth knowing about agentic AI that does not appear in product marketing.
It fails differently, and the difference is not about how often - it is about who finds out.
On rungs 1 to 3, every mistake lands on your screen. The system produces something, you look at it, and you decide whether to use it. A bad answer costs you the seconds it took to read. You are the last check, and you are always there.
On rung 4 you are not there. The mistake has already happened, unattended, at an hour when nobody was watching - and the person who found it is your customer.
| Rungs 1 to 3 | Rung 4, at 22:47 |
|---|---|
| The system writes an answer | The system writes an answer |
| You read it, and decide | No one reads it. There is no decision point. |
| It leaves the building | It goes straight to the customer |
| The mistake lands on your screen. It costs you seconds. | The mistake lands in a customer's inbox. You read about it later. |
That inversion is the single most important thing to understand before stepping onto this rung. It does not make delegation a bad idea. It makes three things non-optional that dealers are routinely tempted to treat as extras:
Real access, because a system without it guesses. On rung 1 a guess is harmless; you catch it. On rung 4 a guess is a wrong answer delivered confidently to a buyer. Writing about how these systems fail, Niels Kristian puts the problem precisely: “there is no signal in the output that says ‘there is something I was not shown’. Confidence is uniform whether the picture is complete or not.” When you are the one reading every answer, that is a nuisance you absorb. When nobody is reading them, it is the entire risk of the rung. He develops it in The most impressive answer is rarely the correct one.
A written stop-list, because some mistakes cost money rather than credibility. Anything carrying a price, a commitment or a liability has to stop and wait for a person: what the car costs, what the trade-in is worth, what the finance terms are, whether the warranty covers a given fault. A system with no refusals has not been designed, it has been switched on. And the test for whether a delegated system is serving your customers or just absorbing them is a simple one: does it settle the matter, or does it hold you?
Disclosure, because the customer needs to know they can ask for a person. EU law has required this since August 2026 - Article 50 of the AI Act, clearly, at the start. But the practical reason is better than the legal one. A system that is pretending to be a person cannot gracefully say this is beyond me, a colleague will call you. One with nothing to hide can, and that sentence is what makes an unattended process safe.
There is one more property of these systems that only shows up once you delegate, and it runs in the opposite direction from what most people fear.
“The agent is always very positive and brave - so it is good to be there all the way in the beginning, to make sure it does not go too far.” — Johan F. Schjødt, co-founder of AutoUncle The worry going in is usually that an autonomous system will be unhelpful: stiff, evasive, refusing things a customer plainly needs. What actually happens is the reverse. It is eager. It wants to resolve the matter, it wants to say yes, and it will keep going further than you intended in order to be useful - reassuring where it should hedge, committing where it should hand over, filling a gap with something plausible rather than stopping at the edge of what it knows.
| Where an agent naturally goes | What that sounds like |
|---|---|
| Safe to state | "The car is still available." "We are open Wednesday from 09:00." "The warranty runs for twelve months." |
| Should hedge | "That model usually holds its value well." "I think we can find something in your budget." |
| Must hand over | "We can do it for 18,500." "Your car is worth around 9,000." "Yes, financing will be approved." |
This is worth separating from the sycophancy problem described in The most impressive answer is rarely the correct one. There, the system is agreeable towards you and a long session drifts. Here it is bold towards your customer, and there is no session to notice drifting in - because you were not in it.
Which produces the single most practical piece of advice in this article, and it is not a feature to buy. Sit with it the whole way at the beginning. Read every exchange for the first weeks, not a sample. You are not checking whether the writing is good; you are looking for the places where it went one sentence too far, and each one you find becomes a line on the stop-list. That is not a lack of trust in the technology. It is how the boundary gets drawn, and it can only be drawn by watching a real conversation with a real buyer.
The test before you buy
Knowing which rung you are on is diagnosis. This is the decision.
“The issue can be if dealers, in the hype of AI, end up buying a tool which is an addition. It’s adding more stuff. Where AI is best is optimising and improving and, in some cases, taking over processes that already exist.” — Johan F. Schjødt, co-founder of AutoUncle Anything on offer does one of three things, and only two are worth money.
It adds: a login, a dashboard, settings to maintain, and no task removed from anyone’s day. It looks like progress on the day you buy it and becomes a tab nobody opens.
It takes over: something a person does today stops being done by a person.
Or it makes a step unnecessary - which is the outcome nobody shops for, because it is an absence rather than a feature:
“Let’s say that you spend time showcasing a car where you could find out that the potential customer doesn’t have the budget for that car. So that is a way where you could save time as well.” — Johan F. Schjødt, co-founder of AutoUncle Nobody automated the viewing. The viewing stopped needing to happen, because something upstream had already settled it.
One question puts the three in order at the moment it matters: after this is installed, what will nobody be doing any more? If there is no answer, the rung has not changed. You have bought a better version of rung 1.
Where to go next, in order
-
Locate yourself honestly. If every AI interaction in your business begins with a person opening something, you are on rung 1 or 2, whatever the invoice says.
-
Do not skip rung 3. Delegating a process that cannot see your stock, your diary or your policies is delegating guesswork. Access first, autonomy second.
-
Pick one process to delegate, not four. For most dealerships, written enquiries - high volume, heavy work, and losses that show up nowhere.
-
Write the stop-list before it goes live, not after the first awkward message.
-
Then be present for every exchange in the first weeks. Not a sample - all of them. You are looking for the sentence that went too far, and each one you find is a rule you did not know you needed.
-
Judge it on consistency, not on the impressive first week. Ask whether the reply on a Friday evening looks like the reply on a Tuesday morning.
The dealerships that get value from this are not the ones highest up the ladder. They are the ones who know which rung they are standing on.
Frequently asked questions
What is agentic AI, in plain terms?
A system that runs without being opened. It is triggered by an event - an enquiry arriving, a car appearing at auction - rather than by a person typing a prompt, and it acts and then hands the result to someone. If a human has to start it, it is not agentic, however capable it is.
How is agentic AI different from a chatbot?
A chatbot answers a visitor who chose to open it, and its scope is that conversation. An agentic system works across a process, unattended, using your dealership’s own data, and produces something a colleague picks up later.
Why does agentic AI need access to my systems when chat does not?
Because nobody is checking the output. When you read every answer before using it, a gap in the system’s knowledge costs you a moment. When it answers a customer at 23:00, that gap becomes a wrong statement delivered on your behalf.
What should a delegated AI system refuse to do?
Anything carrying money or liability: quoting a price, negotiating, valuing a trade-in, committing to finance terms, interpreting warranty cover, diagnosing a fault, or guaranteeing that a car is available. Those belong to a person, and the refusals should be written down before the system goes live.
Should a dealership aim for multiple AI agents working together?
Not yet, for most. One delegated process that runs properly is worth more than several partly-built ones, and the failure modes multiply faster than the benefits.
What is the most common surprise when a dealership first delegates to an agent?
That it is too willing rather than too cautious. An agent wants to be useful, so it pushes further than intended - reassuring where it should hedge and committing where it should hand over. Read every exchange for the first few weeks and turn each overreach into an explicit rule.
Sources
-
Timing of consumer enquiries (peak 21:00-23:00; majority outside opening hours): AutoUncle data, 2026.
-
EU AI Act Article 50 transparency obligations, applicable from 2 August 2026: European Commission.
