Start with the DMS, not the AI

The fastest way to waste an AI budget is to start with the assistant.

A dealership does not have an intelligence shortage. It has a context problem. The delivered unit lives in the DMS, the shopping intent lives in the CRM, the current behavior lives on the website and useful enrichment lives in outside APIs. Until those sources meet, the assistant is forced to guess.

1. DMS - where the transaction settles

Start where money changes hands. The FTC describes the DMS as touching accounting, payroll, parts, inventory, service, vehicle financing and manufacturer data flow. It is not always elegant, but it is the closest thing the store has to a transaction ledger.

Pull stable, documented data through an approved API or controlled syndication path. Land it in a warehouse the group controls. Preserve source timestamps and keys. Do not begin by cleaning everything. Begin by making the raw history reproducible.

2. CRM - intent, activity and identity

The CRM explains what happened before the transaction: leads, calls, appointments, notes and ownership. It is also full of duplicates and conflicting outcomes. That is expected. The job is to connect it to the DMS without letting a CRM status redefine a delivery.

Identity matching should be explicit and reviewable. Start with deterministic matches such as known IDs, normalized email, phone and address. Keep confidence levels. Do not silently merge two people because their names look similar.

3. Website - first-party behavior

The website shows what the customer is doing now. Store event-level activity that matters to an action: vehicle views, form starts, completed leads, service interest and consent. A giant clickstream is not automatically useful. Collect the events the group can define and govern.

4. Public and third-party APIs - owned enrichment

Only after the internal sources agree should the group add enrichment. NHTSA publishes recall datasets and APIs. Market values, incentives and other feeds can add context, but each needs an owner, refresh schedule and license review.

Security is part of the architecture

The FTC Safeguards Rule requires covered dealers to inventory data, control access, encrypt customer information, use multifactor authentication, monitor systems and oversee service providers. Those are not compliance decorations added after the build. They shape the build.

Separate customer information from operational aggregates where practical. Give a reporting tool a governed view instead of raw finance tables. Log access. Keep the smallest useful field set. The goal is more usable context with less unnecessary exposure.

The first useful outputs

Do not wait for a perfect model. Ship outputs that prove the foundation:

• a daily delivery count reconciled to the DMS;
• a lead-to-sale view where the CRM cannot grade itself;
• a duplicate-customer review queue;
• a recall match preview with no outreach attached;
• a source-health report that stops when data is late or malformed.

Once those are trusted, the assistant becomes useful. It can call governed tools that already know what a sale is, what data it may see and where the answer came from. Start with the DMS. Earn the AI.