A pricing model hands you a number for a three-year-old estate with 62,000 km on the clock. It looks roughly right, but you cannot use "roughly right" to defend a trade-in figure to a customer, or to justify a markdown to your principal. The question that actually matters is not what the model said, but why it said it. Explainable AI car pricing is the discipline of answering that second question: turning a single recommended price into a readable breakdown of the factors that pushed it up or down.
This article is for dealers and pricing managers who are being asked to trust AI valuations and want to understand the mechanics underneath. We will cover what feature attribution means in plain terms, how a SHAP-style breakdown reads on a real car, where these explanations can mislead you, and what to look for when you evaluate a tool. The goal is not to make you a data scientist. It is to give you enough grip on the why-this-price that you can challenge a number when your instinct says it is off, and back it with confidence when it is right.
Why a single price is not enough
A bare number gives you no way to disagree intelligently. If the model says EUR 21,400 and your gut says EUR 19,500, you have a standoff with no evidence on either side. You cannot tell whether the model is right and your instinct is anchored on an old sale, or whether the model has over-valued a feature that does not actually move metal in your region.
The cost of an unexplained price is not abstract. Over-value a part-exchange and the margin is gone before the car reaches the forecourt. Under-value it and you lose the deal, or the customer walks to a competitor who priced the same car with more confidence. Repeat that across a month of appraisals and the drift adds up. An explanation does not remove the judgement; it gives the judgement something to work with.
How feature attribution works in plain terms
Most modern pricing models are not simple formulas you can read off a page. They learn patterns from large numbers of past transactions, and those patterns interact in ways that are hard to inspect directly. Feature attribution is the bridge: a way to ask the model, "for this specific car, how much did each factor contribute to the final price?"
Start from a baseline
The method begins with a baseline, sometimes called the expected value. Think of it as the average price the model would predict for a car in this broad segment before it knows anything specific about your vehicle. From there, each known fact about the car nudges the price away from that baseline.
Add and subtract contributions
Every feature gets a contribution: a direction (up or down) and a size. The figures below are an invented illustration to show the shape of the output, not numbers from any real valuation. Suppose the baseline for the segment is EUR 19,000:
| Factor | Direction | Contribution | Reading |
|---|---|---|---|
| Baseline (segment average) | - | EUR 19,000 | Starting point before this car's specifics |
| Mileage below segment average | Up | +EUR 1,400 | Lower-than-typical km adds value |
| Desirable trim and options | Up | +EUR 900 | Spec the local market pays for |
| Age (older than ideal) | Down | -EUR 700 | Standard depreciation pull |
| Minor cosmetic condition flags | Down | -EUR 300 | Reconditioning the model expects |
| Local demand for body style | Up | +EUR 500 | Regional appetite right now |
| Recommended price | - | EUR 20,800 | Baseline plus all contributions |
The contributions add up to the final number, which is the property that makes this kind of explanation trustworthy: nothing is hidden in the gap between the baseline and the price. This additive style of breakdown is what people mean when they refer to a SHAP-style explanation, after a well-known method for fairly distributing a prediction across its inputs. You do not need the maths to use the output. You need to read it like a receipt.
Local versus global explanations
There is an important distinction. A global explanation tells you which factors matter most across the whole model, on average, mileage and age, perhaps, dominating everywhere. A local explanation tells you why this car got this price. For day-to-day pricing decisions, the local explanation is the one you want, because the same factor can pull in different directions depending on the rest of the vehicle. High mileage on a workhorse diesel reads differently from high mileage on a city runabout, and only a local breakdown shows you that nuance.
What a good explanation lets you do
Once you can read the contributions, the explanation becomes a working tool rather than a curiosity.
- Challenge the outliers. If the model adds a large amount for a feature you know the local market is indifferent to, you have a specific thing to question rather than a vague unease.
- Defend the number. When a customer pushes back on a trade-in figure, you can point to the mileage advantage and the demand signal instead of saying "the system decided".
- Catch silent errors. Explanations surface a model leaning on something stale, such as a regional demand signal that has not refreshed, or treating an optional extra as standard fit.
- Train your team. New appraisers learn faster when they can see the reasoning behind a price, not just the price.
This is the same logic behind why an AI valuation should show its working in the first place. If you want the broader picture of how a valuation is produced end to end, see how AI car valuation works, which covers the process from data ingestion to recommendation.
Where explanations can mislead
Explainability is powerful, but it is not magic, and treating it as a guarantee is its own kind of risk.
First, an explanation describes what the model did, not what is true. If the model learned a spurious correlation, say, a particular colour happening to coincide with a strong period of sales, the attribution will faithfully report that colour added value, even though the link is noise. A clear explanation of a flawed pattern is still a flawed pattern.
Second, attributions can be unstable when features are correlated. Mileage and age tend to move together, so the split of credit between them can shift in ways that look arbitrary. Read them as a combined story rather than fixating on the exact euro split between two intertwined factors.
Third, an explanation tells you about the model's reasoning but not about the quality of the inputs. If the mileage was entered wrong or the trim was mis-decoded, the attribution will explain a confidently wrong price. That is why explainability has to sit alongside trustworthy source data, the subject of the trust layer that shows where automotive AI gets its sources.
How to evaluate explainability in a pricing tool
When you assess a valuation product, the explanation is as important as the headline accuracy. Use these criteria.
- Local, per-car breakdowns. Can you see why this vehicle got this price, with named factors and directions, or only a generic "factors we consider" list?
- Additive and complete. Do the contributions actually reconcile to the recommended price, or is there an unexplained gap?
- Plain language. Are factors labelled in dealer terms (mileage, trim, demand) rather than opaque model internals?
- Tied to the inputs. Can you trace each factor back to the underlying data field that produced it?
- Paired with uncertainty. Does the tool also give you a range, not just a point estimate?
That last point deserves weight. An explanation tells you why a price was set; a confidence interval tells you how sure the model is. A narrow range on a common, well-evidenced car and a wide range on a rare spec are both useful signals. The combination of the two is what turns a number into a decision you can stand behind. For more on reading that range, see confidence intervals in car valuation.
Where VehIQ fits
VehIQ is being built so that a price is never just a number. The platform is designed to pair every AI valuation with the factors that shaped it and a confidence interval, drawn from canonical European vehicle data with field-level lineage, so you can trace a contribution back to the input that produced it. The aim is to give dealers and pricing managers the why-this-price directly in the workflow, rather than asking them to trust a figure on faith.
VehIQ is pre-seed and still being built, so this describes what the system is designed to do, not deployed results. It is intended to run alongside the tools you already use rather than replace them, and to keep the underlying data in open formats you own. If you are weighing up AI pricing, the practical test is simple: ask any tool to show its working, on this car, today. An explanation you can read is the difference between using a model and merely obeying it.