Landing inventory on the right rooftop cut cross-rooftop days-to-turn from 52 to 42, lowering floor-plan interest costs.
App share tripled in four quarters, giving marketing a first-party channel for offers and paid upsells.
Fresher inventory and fewer forced markdowns lifted per-unit used gross profits for the group.
Our client is a private franchised dealer group in the U.S. Southeast, operating 20 rooftops across nine OEM brands. The group retails about 14,000 new and used vehicles a year, writes roughly 290,000 repair orders, and employs 1,100 people. Every store runs its own DMS instance, CRM, and used-car pricing tool.
Growth by acquisition left the group with capable software but no shared view. Weekly reports arrived days late, pieced together by hand from twenty different logins. Decisions that belonged at the group level stayed trapped inside separate stores instead
Leaders pulled numbers from twenty DMS logins, then rebuilt the group picture by hand in a weekly spreadsheet.
A sedan sat for 90 days at one store while three sister locations had steady buyers waiting for that exact trim and mileage.
Off-the-shelf software handles the single-store job well, but it can’t consolidate data, guide cross-rooftop decisions, or let you own your data.
Management could not see real-time project profitability, including labor costs, materials, purchases, and subcontractor expenses.
Without service and parts demand forecasting, bays sat idle on Tuesdays, and parts arrived after the customer had already walked out.
Every model and score lived inside a per-seat SaaS subscription the group rented, so the data asset left the group with the vendor.
We replaced nothing in the existing stack. Our team built an integration and decision layer above the group's DMS, CRM, and pricing tools, so commodity work stayed on commodity software. Above that layer sit the models no vendor sells off the shelf: cross-rooftop stocking, fixed-ops forecasting, and group-wide lead routing.
We consolidated 20 DMS instances, every CRM, and market feeds into one cloud warehouse with modeled, versioned tables. Identity resolution joins customers, VINs, and repair orders across rooftops that never shared records. Governance, access, and lineage live with the group, so the asset compounds. A consolidated BI layer reads from the same models the ML services do.
Demand and pricing signals from all 20 markets feed a ranking service that spots a unit selling slowly in one market and in demand in another. It weighs local turn history, market days' supply, transport cost, and floor-plan carry before recommending a transfer or a markdown. Used-car managers see a ranked daily list, with the reasoning attached to every recommendation.
Gradient-boosted models forecast repair-order volume, labor hours by skill, and parts consumption per store at a weekly and daily grain. Inputs include RO history, vehicles in operation nearby, warranty and recall campaigns, seasonality, and appointment patterns. Service managers staff bays against projected load, and parts managers stock ahead of it.
Instead of locking leads to one lot, every inquiry is scored for close likelihood and matched against whole-group inventory. The system pushes exact steps, assigned reps, and vehicle matches right into existing store CRMs, avoiding new software. Smart routing respects brand, distance, and BDC capacity, while monthly performance updates continually refine the model.
Numbers measured across all six markets in the four quarters after full rollout.
Average used days-to-turn fell from 52 to 42 across the group as transfers moved units toward real demand. 10 fewer days of carry per unit pulled floor-plan interest and depreciation back out of the used-vehicle margin.
Service managers started staffing bays to the forecast, and parts managers stocked against it, boosting same-store service and parts revenue. Locations still waiting for rollout grew at less than half that pace.
Fresh inventory and fewer emergency markdowns lifted gross profit on every used unit sold. Aging stock cleared out much faster, cutting the painful wholesale losses that hit when vehicles sit on the lot for too long.
By scoring and routing leads group-wide, conversion set rates jumped from 39% to 50%, a 28% relative gain. Sales managers worked a prioritized queue rather than a random inbox, letting follow-ups run without manual triage.
Month-end reporting shifted from a 5-day spreadsheet grind to an automated dashboard updated daily. The group can now conduct its Monday leadership meeting using a single, live view that covers all 20 store locations.
Inside the Architecture of the Group Decision Layer
Every layer was chosen for group-scale volume, DMS reality, and client-owned data.
| Layer | Technologies | Rationale |
|---|---|---|
|
Ingestion & integration
|
Fivetran, custom Python connectors, DMS APIs (CDK, Reynolds, Dealertrack, Tekion), Apache Airflow
|
Handles per-store API quirks and nightly extracts from four separate DMS vendors without custom glue per rooftop.
|
|
Storage & compute
|
Snowflake on AWS, S3 staging zone
|
Separates storage from compute, so twenty rooftops of transaction history query fast without a fixed hardware bill.
|
|
Modeling & governance
|
dbt, Great Expectations, Snowflake RBAC
|
Version-controlled SQL models give finance one shared definition of gross, turn, and absorption across every store.
|
|
Forecasting & ML
|
Python, LightGBM, XGBoost, scikit-learn, Prophet baselines
|
Gradient boosting handles sparse, seasonal repair-order and parts data better than deep nets at this data volume.
|
|
MLOps
|
MLflow, Docker, GitHub Actions, Evidently AI
|
Model registry, drift monitoring, and scheduled retraining keep transfer and forecast logic honest as markets shift.
|
|
Data
|
DynamoDB, Aurora Serverless v2, Amazon S3
|
Key-value storage serves fast availability lookups; relational storage holds contracts, invoices, and audit history.
|
|
Decision services
|
FastAPI microservices, Redis, CRM and DMS write-back
|
Low-latency scoring pushes the recommended next action into the CRM the sales team already has open all day long.
|
|
BI & access
|
Power BI, row-level security, Okta SSO
|
Row-level security lets each general manager see their own store while the group office sees all twenty at once.
|
This engagement fits multi-rooftop dealer groups, RV and powersports networks, equipment dealers, and rental or fleet operators running 5 to 50 locations on per-store systems.
If your DMS and CRM treat every store as an island, and group reporting is still a late-arriving spreadsheet, this case study may speak to you It also fits groups that grew through acquisition and inherited a mix of DMS vendors that nobody wants to migrate off. We added value over the existing stack in one quarter without ripping anything out, retraining staff, or paying per-seat fees.
AI Transfer Calls
Fixed-Ops Forecasts
Client-Owned Data
No DMS Migration
One View
Live in 90 Days
Unify your store systems, automate forecasting, and make smarter cross-rooftop decisions without replacing a single tool or retraining staff.
Explore our broader portfolio of client work across industries to see what lines up with your goals. Every engagement brings its own technical demands, and we adapt our tools to match your exact setup.
