Here is a problem almost every dealership has, and almost none has solved properly.
You have a hundred and fifty cars in stock. You advertise them across your own website, two or three portals, paid search, and some social. Each of those channels runs its own auction or ranking. Each one behaves differently depending on the car - price band, age, fuel type, how common the model is. Somewhere in the interaction between your stock and those channels there is an allocation of your marketing budget that is meaningfully better than the one you are running today.
You will probably never find it. Not because you are not capable, but because finding it requires being a specialist in online marketing, and you are already doing a full-time job that is not that. So the budget goes roughly where it went last year, adjusted by feel. That may be an excellent decision. It may be costing you a great deal. The honest position is that you do not know.
This is the exact shape of problem that AI is now extremely good at - and that most people fail to get help with, for a reason that has nothing to do with technology.
The instinct that gets in the way
Faced with a capable AI system, the reflex is to give it a task. Write me a marketing plan. Tell me the best channel for used electric cars. Compare these three portals.
Every one of those questions produces an answer. Most of the answers are generic, because the question contained no information about your business. You get the average of everything ever written on the subject, formatted confidently. You read it, think “that’s quite good, actually”, and do nothing with it, because it does not fit your situation closely enough to act on.
Then you conclude that AI is impressive but not that useful. This is the single most common failure pattern, and it is entirely self-inflicted.
The problem is that you asked it to answer. The move that works is to ask it to ask.
The move
Open a capable AI system and type something close to this:
I run a used car dealership. I have around 150 cars in stock. I advertise on several channels but I have no idea whether I am spending my marketing budget well, and I am not a marketing specialist. Do not give me a plan yet. Interview me until you understand my situation properly, and tell me when you have enough.
Then answer the questions. Including - especially - the ones you cannot answer.
What happens next is the part people do not expect. It will ask what you spend per channel. You may not know. Say so. It will ask where you would find out, suggest that the numbers are probably sitting in invoices in your email, and offer to help you look. It will ask what proportion of your website traffic comes from where. You may not know that either. It will tell you where that data lives and what to click.
At no point does it require you to already understand the problem. That is the whole point.
| INSTRUCTING IT | ASKING IT TO INTERVIEW YOU |
|---|---|
| "What is the best channel to advertise used electric cars?" | "I do not know if my marketing spend is well allocated. Interview me." |
| Answer arrives in twenty seconds | Twenty minutes of questions before any answer |
| Generic, correct, unusable | Specific to your stock, your spend, your market |
| You act on none of it | You act on it, because it describes your actual situation |
| You conclude AI is overrated | You conclude you had a data problem, not a marketing problem |
Why it works
Two reasons, and they are worth understanding rather than just accepting.
The first is that defining the problem is the hard part. This is true in engineering, in medicine, in law, and in running a business. Solutions are comparatively easy once the problem is stated precisely. Most business problems are never stated precisely, because the person closest to them has stopped seeing them clearly. An outside party asking naive, patient, sequential questions is the oldest debugging technique there is. AI is now very good at being that party, and unlike a consultant, it is available at 22:00 and does not charge by the hour.
The second is that these systems ask better questions than they give answers. That sounds like a criticism. It is not. A question is cheap to be wrong about - you just correct it. An answer built on a misunderstanding is expensive, because you act on it. Front-loading the questions moves the risk to the cheap end.
There is a related habit worth adopting: many systems have an explicit mode for exactly this, where the work is deliberately deferred while the shape of the task is agreed. If yours does, use it. If it does not, you can create the same effect by saying “do not start yet” and meaning it.
The part that feels bad
Now the honest bit, because it is the actual obstacle.
It is uncomfortable for a professional to type “I run a dealership with 150 cars and I do not really know what I am doing with my marketing.” You have run this business for years. You have staff. People ask you for advice. Writing that sentence - even to a machine, even with nobody watching - cuts against everything about being competent at your job.
And that discomfort is precisely why most people never open the door to the best help available to them. They stay in instruction mode, where they are the expert giving directions, and they get expert-shaped answers to questions they did not know how to ask.
The reframe that helps: you are not admitting incompetence. You are declining to pretend to a specialism you never claimed. Nobody expects a dealer principal to be an online marketing specialist, an employment lawyer and a data analyst. The reason those things go under-optimised in every business is not stupidity. It is that a day has a finite number of hours and you spent yours on the business.
There is a smaller version of the same discomfort worth naming too. When a system gives you an answer you do not follow, say so. I do not understand what you have written. Explain it as if I have never seen this before. Nobody is watching, there is no cost, and the alternative is nodding along and acting on something you have not understood.
When it gets to the answer
Once it genuinely understands your situation - what you spend, where, what comes back, what is on your forecourt - the output changes character entirely. It stops being an article about marketing and starts being a recommendation about your business: that cars above a certain price band are being advertised in the wrong place because the audience for them sits elsewhere, that a channel you have been paying for since 2021 is producing nothing you can trace, that a segment of your stock is invisible on the platform where it would sell fastest.
And you will have gone from an allocation you could not defend to one you can, in an afternoon, without hiring anyone.
The boundary
This method is superb for one category of work and dangerous for another, and the line between them is worth being precise about.
Use it freely for anything where the output lands on your own desk. Analysis, strategy, comparisons, understanding a market, working out what your own data says, preparing for a negotiation, learning a subject you have avoided. Nobody outside your business is affected if it gets something wrong, and you will catch it.
Do not use it casually for anything that becomes the customer’s experience of you. The moment output crosses the line from your desk to a buyer’s inbox, the standard changes completely. That is not a reason to avoid AI in customer-facing work - done properly it is the strongest use case in the industry - but “done properly” means curated, grounded in your dealership’s actual facts, tested before it goes live, and disclosed as AI. Since 2 August 2026, EU law requires the disclosure part.
The failure mode people fall into is using the casual method on the consequential work: pasting a car’s details into a chat window, getting a description full of weightless prose about summer drives, and publishing it. That is the same instinct that produces the generic marketing plan, except this time a customer sees it.
Ask it to interview you about the problem on your desk. Build something properly for the problem in your customer’s inbox.
Frequently asked questions
What does “ask it to interview you” actually mean in practice? Describe your situation and your uncertainty, then explicitly tell the system not to produce an answer yet, and to ask you questions until it understands. Answer honestly, including “I don’t know.” Let it tell you where to find what you don’t know.
Does this work if I do not have my data ready? Yes, and that is the main advantage. A good system will identify what data it needs, tell you where it typically lives - invoices in your email, your website analytics, your portal dashboards - and walk you through retrieving it. You do not need to arrive prepared.
How long should the interview go on? Longer than feels natural. Most people stop too early, at the point where they have said enough to get a plausible-sounding answer. The useful threshold is when the system can describe your situation back to you more accurately than you described it yourself.
Is this different from just using ChatGPT? It is a different way of using it. Same tool, opposite posture: you are being questioned rather than issuing instructions. The tool matters less than the posture.
Can I use this to write customer communications? Not directly. Use it to understand your business and prepare your thinking. Customer-facing communication needs to be grounded in your dealership’s real facts - stock, policies, availability - and, in the EU, disclosed as AI. That is a different kind of build.
Sources
- EU AI Act Article 50 transparency obligations, applicable from 2 August 2026: European Commission.
