Machine Learning for Car Dealers and Workshops
Last updated:
A forecourt is a stock problem with number plates
Every used car on the forecourt is cash that is losing value each day. Dealer groups know this, and the ones who manage it best treat used cars like a trader treats inventory: buy at the right price, price to a target sale date, and move ageing stock before it becomes a problem.
Consider a dealer group with 12 sites holding 900 used cars. If average stock days fall from 55 to 45, the cash released and the depreciation avoided are both substantial, and the arithmetic comes straight from your own stock book. Machine learning for car dealers is mostly about making those buy, price and move decisions with better information than the national average valuation.
Dealer and workshop machine learning use cases
- Days-to-sell prediction. Estimating how long a specific car will take to sell at a given price at a given site, using your own history of similar cars.
- Part-exchange and buying valuation. Adjusting market valuations for condition, spec, colour and how that model has actually sold for you.
- Stock allocation between sites. Moving a car to the site where similar cars sell fastest.
- Service retention. Predicting which customers will not return for their next service or MOT, so the reminder and offer reach them in time.
- Parts demand forecasting. Stocking the parts your workshops will need, and not the ones gathering dust.
- Workshop job time estimation. Predicting real labour time from vehicle, job type and technician history, so bookings fit the day.
- Lead scoring. Ranking online enquiries by likelihood to buy, so sales staff call the right people first.
Buy, build or both
Market pricing platforms already give dealers retail valuations, market days supply and competitor pricing. They are good, and building your own national valuation model would be foolish. The question is what your own data adds on top.
| Question | Bought-in pricing tool | Model on your own data |
|---|---|---|
| What is this car worth nationally? | Strong | Unnecessary |
| How fast will it sell at my site? | General market view | Learns your locations and customers |
| What should I pay for this part-exchange? | Market guide | Adds your reconditioning costs and sale history |
| Which service customers are about to lapse? | Not covered | Needs your DMS and service history |
| What parts should each site hold? | Not covered | Needs your parts sales and job data |
In practice the best dealer projects combine the two: market data as an input, your history for the local adjustments. Our post on AI pricing optimisation explains how price and sale speed models are usually tied together.
Service retention is the quiet money
Sales get the attention, but aftersales often earns most of the gross profit. A customer who bought a car three years ago and stops servicing with you is a loss that rarely appears on any report.
A retention model uses service history, mileage patterns, vehicle age, warranty status, distance from the site and past response to reminders to rank customers by risk of lapsing. The model is modest. The value comes from what you do with it: a timely reminder, a fixed-price service offer for out-of-warranty cars, or a collection service for customers who have moved further away. The same principles appear in our guide to churn prediction models.
- Score customers monthly, not once a year
- Reserve discounts for high-value customers at genuine risk, not everyone
- Hold back a small group from any offer so you can measure whether it worked
- Respect marketing consent in the DMS before contacting anyone
The data inside a dealer management system
Most dealer groups have years of useful data in their DMS, locked behind limited reporting and awkward exports. Common problems we meet:
- Reconditioning costs posted weeks after sale, or not linked to the vehicle
- Duplicate customer records across sites, so retention looks worse than it is
- Stock records that lose the original advertised price after reductions
- Workshop times recorded as sold hours rather than actual clocked hours
None of this blocks a project, but it has to be sorted before a model is trusted. Budget for extraction and cleaning to take as long as the modelling.
What it costs, and who should not bother
Indicative ranges: a days-to-sell and pricing guidance model on your sales history, eight to twelve weeks including a buyer-facing screen; a service retention model with campaign integration, six to ten weeks; parts demand forecasting across sites, eight to twelve weeks.
An independent garage with two ramps does not need any of this. Good booking software and automatic MOT reminders will do more. Groups with under about five sites should usually start with a market pricing tool and better reporting, and add a model only when the reporting shows a clear, repeated gap.
A good first project
Start with aged stock. Pull two years of used car sales with purchase price, advertised prices over time, reconditioning cost and sale date. Build a model predicting days to sell, run it against current stock every Monday, and give buyers and site managers a short list of cars that need action. It is measurable within a quarter.
At SpiderHunts, we build these on top of the systems you already run, using data science methods that buyers can interrogate, because a pricing tool the used car manager does not believe will be ignored by Wednesday.
Frequently asked questions
Can machine learning price used cars better than market valuation tools?
How does a dealer predict which customers will stop servicing?
Do we need to replace our DMS?
Is parts forecasting worth doing?
Too much capital sitting on the forecourt?
Tell us how many sites you run and what your dealer management system exports. We will say which prediction would help most and whether a bought-in tool already covers it.
Related services
What we build for problems like this one