The AI project you were pitched is a process problem

Sooner or later the proposal lands on the desk: an artificial intelligence project for the dealership, with a promised saving and a number attached. These notes are for the person who has to approve that number.

The useful thing to know before signing is that almost none of these projects fail on the model. They fail earlier, on something far less interesting: the process they were meant to lighten had never been written down, and nobody knew what it cost. AI laid over a process nobody has mapped automates the steps that should have been removed, and makes them harder to remove afterwards.

The saving shows up on the process map, before anything is built

The way we work on this has two tempi, and the order matters more than the content.

The first tempo is mapping and cutting. Mapping means writing the process down as it actually runs, step by step, with the time and the cost against each step, asked of the people who do the work rather than read off the org chart. The real map is almost always longer than the one the managing director expected, and the difference sits in steps invented to work around problems somebody else has since fixed. Cutting those costs nothing.

The second tempo is building on whatever survives the cut. Only on that.

Reverse the order and you buy the automated version of the wrong process, then keep it because it was expensive. Which is why most of the saving is visible at the end of the first tempo, before anything has been built.

A back office that hires every time volume grows is the symptom, and you already have the number

There is a signal you can read without consultants. When volume grows, does the headcount processing it grow too? If it does, a process is scaling with people and should not be.

Here is the part that usually loses the pitch. We have no saving figure to give you, and anybody who gives you one before seeing how you work has invented it. There is no industry percentage of back office that AI removes from a dealership. There is your number, and you already hold it:

• how many people it takes to process a month of registration paperwork;
• how many hours a warranty claim sits between being opened and being sent to the manufacturer;
• how often a quotation is retyped by hand into a second system;
• how many supplier invoices pass under a human eye for a check that almost always ends the same way.

Those are four counts you can run in house, without buying anything. Anyone who asks for them before proposing something is doing the job. Anyone arriving with the saving already calculated has not yet opened your dealer management system.

Agents where the work repeats: contracts, warranty claims, invoices, homologation packs

What survives the cut splits in two, and the halves are built differently.

The first half is work that repeats identically. Sales and finance contracts, warranty claims, parts invoices, homologation packs, ownership transfers. An agent here understands nothing new: it reads a document, pulls out the fields, puts them where they belong and raises its hand when something does not add up.

Two conditions have to hold, and they are where projects run aground. The data has to be reachable: if the chassis number exists only inside a PDF in a shared folder, the problem is not the model, it is the plumbing. And the output has to land inside a system somebody already uses. An agent that produces a spreadsheet somebody then retypes into the DMS has moved the work, not removed it.

Then it needs an owner and a measure. An agent with no name beside it and no number to judge it by stays a pilot indefinitely.

Models where a number has to move, and in this industry there are two of them

The other half is different. There is no repetitive work to remove, there is a forecast made by eye. In a dealership the numbers that deserve a model are few, and two matter more than all the rest.

The first is what a car will be worth in three years. It sets the lease rate, the finance instalment and the price you can take a part-exchange at today, and it is the number moving most at the moment. Get it wrong and you find out three years later, when the car comes back.

The second is what to order and how much of it to hold. A car standing on the forecourt ties up capital and depreciates every month, and the ordering decision is usually taken on last season plus instinct.

In both cases the part that counts is not the forecast: it is the error band declared beside it. Without a range, a forecast is trusted further than it deserves, and the first time it is wrong it discredits every time it was right.

Every merger brings in two back offices, and the aggregation has not slowed down

According to Quintegia's Dealer Network Study, Italy's independent dealer-entrepreneurs fell by roughly 40% over ten years, and multi-manufacturer operators have reached 44% of the network. Roughly 9% of European franchise points changed ownership between January 2022 and September 2025 (ICDP, reported by Automotive News Europe). We have told that story in full elsewhere: here they do one job: to say how many different sets of procedures sit inside it.

Because an acquisition does not merge procedures. It inherits two of them. Two ways of opening a file, two charts of accounts, two teams doing the same thing in two defensible ways. And it runs like that for a while, because nobody has time to rewrite them mid-integration.

That is where the process map stops being an exercise and becomes the only way to decide which of the two procedures to keep. That is not an AI question. It becomes one afterwards, and only for what is left.

For an investor, the question is how much of the claimed advantage is a subscription

On the other side of the table the question changes shape. An investment committee looking at a target that claims a technology advantage has to decide whether it is paying for it twice. There are three questions, and they are dull on purpose.

Is the model the target's, or somebody else's? How much of that advantage is a subscription the competitor can sign tomorrow, at the same price, having done nothing to earn it?

Does the data it runs on stay inside the perimeter you are buying? An advantage living in a supplier's system or a manufacturer's platform is not an asset of the target.

Who owns the number, and how long have they measured it? If the supplier measures it, you are not buying an advantage. You are buying their deck.

That is the kind of check we run inside a commercial due diligence when a target's plan rests on a technology promise. It can also be asked on its own, as a single question with no transaction behind it, when the committee only needs to know whether the promise holds.

The most useful answer we can give is that you do not need AI for this

It happens, more often than anyone selling projects finds it convenient to admit. You reach the end of the map, cut everything that can be cut, and what remains has neither the repetition that justifies an agent nor a number big enough to justify a model. The honest answer is that you are finished: the work was the process, the process is fixed, and putting technology on top of a healthy process adds a fixed cost and a supplier.

We write it down because it is one of the things we say we deliver, and it is the only line on that list that costs its author anything.

A supplier who cannot say no has only one possible answer, and you know it before you call: the value of an opinion lies entirely in its being able to come back negative. And there is a practical reason too. A project started when it was not needed does not fail immediately. It fails slowly, after it has been paid for, and it takes the internal credit of the next project with it, the one worth doing. Saying no now keeps the door open for the time it counts.

So before signing any artificial intelligence proposal for a dealership, there are two questions to ask. Are you showing me the process map as it runs today, with the time and the cost against every step? And is the saving the proposal is built on my number, or yours?

If the first answer is no and the second is yours, what you are holding is not an AI project. It is a quotation.