Your Data Does Not Need to Be Perfect Before AI Can Help

The phrase "our data is too messy" has delayed more dealership data work than almost any technical limitation. The concern is understandable. Customer relationship management (CRM) records contain duplicates. Dealer management system (DMS) fields are used differently across rooftops. Inventory feeds arrive late. Repair-order notes are inconsistent. Managers maintain local spreadsheets because the official report does not answer the question they face at 8 a.m.

None of that should be ignored. It also does not mean the dealership must complete a multiyear cleanup before artificial intelligence (AI) can contribute. Perfect data is not a practical operating standard. Traceable data is: the test is whether a manager can get from a recommendation back to the deal, the unit, the customer or the repair order that produced it.

This chapter is a guide rather than an argument. The general manager picks the decision. The controller owns the contract. Thirty days is enough, and the checklist above is the thirty days, written for a fictional store whose first decision is the morning aged-inventory review. Substitute the decision; the structure holds.

Days 1 to 3: begin with the decision, not the database

A broad instruction to clean the CRM has no natural stopping point. A decision such as which active units require an aging review this morning is much easier to define. The output of the first three days is one sentence: the decision, the person who makes it, and the meeting where it is made. In the fictional store it reads: the used-car manager decides, in the Tuesday aging meeting, which units get a price move, a merchandising fix, a transfer or a hold.

That scope makes data quality operational. A missing acquisition date is no longer an abstract defect in a large table. It is a specific reason a unit could not be evaluated. A stale market feed is not a general integration concern. It is evidence that the pricing recommendation should be withheld until the source is current. The same logic applies to negative-gross review, service recovery, lead follow-up, or a weekly executive briefing. Each workflow should begin with a clear decision and the minimum evidence needed to support it.

Days 4 to 14: write the minimum viable data contract

A useful data contract does not require every field in every system to be pristine. It establishes a reliable boundary for one operating use case, and it has six parts. The controller writes it with the department head who owns the decision.

• Source. Name the system and record that supplied each important fact. A manager should be able to return to the original deal, vehicle, customer or repair order. For the aging review that is seven fields: stock number, acquisition date, current price, recon status, days in stock, market comparison and last manager action.
• Definition. Agree on the business meaning of each measure. Days in stock, sold and appointment shown can differ across systems and stores unless leadership defines them. Days in stock from acquisition or from retail-ready is the first argument the fictional store had, and it was worth having on day 9 rather than in the meeting.
• Freshness. State how current the data must be for the decision. A weekly benchmark can tolerate a different delay than a same-day price review.
• Identity. Specify how records are matched across systems and how uncertain matches are handled. A confident unknown is better than a silent false match.
• Completeness. Identify which missing fields block a recommendation and which merely reduce confidence.
• Ownership. Assign responsibility for correcting a recurring source issue or accepting a documented limitation.

This contract gives managers something concrete to audit. It also gives technical teams a priority list grounded in dealership operations rather than a generic cleanup backlog.

Days 15 to 18: run it by hand once

Before any software touches the decision, the owner runs it from the sources, by hand, one time. The used-car manager pulls the seven fields for every unit over 30 days and tries to make the Tuesday call. Every field that is missing, stale or disputed goes on a list, and every item on the list is assigned to the person who owns the source.

This is the step stores skip, and it is the one that produces the backlog. In the fictional store the hand run found acquisition dates missing on 9 percent of units because transfers from a sister rooftop arrived without them, a market feed that refreshed on Monday night for a Tuesday morning meeting, and two definitions of recon-complete between the shop and the lot. None of those were discoverable from a dashboard. All three had an owner by day 18.

Days 19 to 21: do the compliance work once

Here the guide earns its keep for the controller. The Federal Trade Commission's Safeguards Rule requires a dealership, as a non-bank financial institution, to maintain a written information security program covering the customer information it holds, which means knowing what that information is, where it lives and who can reach it. Since May 2024 the Rule also requires notifying the FTC within 30 days of discovering a security event affecting 500 or more consumers.

That inventory is the data contract's source column with an access rule added. The three days are spent reconciling the two: mark which fields in the workflow are customer information, who may see them, and for what purpose. For an aged-inventory review the honest answer is that almost none of it is customer information, and the workflow should not carry any that it does not need. For a declined-service recovery workflow the answer is different, and the inventory says so before the workflow is built. The second figure above lays the overlap out, including the mapping and governance functions of the NIST AI Risk Management Framework, which ask the same questions in different words.

Days 22 to 25: make uncertainty visible

Data systems often create false confidence by presenting a precise answer without showing the quality of the inputs. A responsible operating workflow should expose uncertainty. If recon status is missing, the manager should see that limitation. If two customer records may refer to the same person, the system should not quietly merge them. If a source feed is older than the agreed window, the output should say so.

The general manager writes three confidence rules: when the workflow recommends, when it requests human review, and when it declines to answer. Traceability lets the user inspect the path from source to conclusion. Shared definitions make that path understandable across departments. Confidence rules determine which of the three happens. This is especially important for natural-language analysis. A clear response should distinguish a verified fact, a calculation based on defined fields, and an interpretation that depends on incomplete context. The goal is not to make every answer sound certain. It is to make the evidence and limits clear enough for a manager to decide.

Days 26 to 30: let the workflow improve the data

Data quality is not a one-time project. It is a feedback loop. When a workflow repeatedly encounters the same missing field, conflicting definition or late feed, the business impact becomes visible. Leadership can then decide whether the source process deserves attention. The last five days name an owner and a cadence for each recurring defect, and then run the decision from the contract in the real meeting.

This approach prevents cleanup work from becoming detached from value. The store fixes the problems that obstruct important decisions first. It can also measure progress in operational terms: fewer blocked reviews, fewer manual reconciliations, more source records available at decision time, and clearer ownership of exceptions.

One more reason to know the sources this well: continuity. When CDK Global's systems went down in June 2024, Anderson Economic Group estimated the direct losses to affected dealers at $1.02 billion. A store that knows which decisions depend on which feed knows, on the first morning of an outage, which meetings it can still run and which it cannot. A store that does not finds out by trying.

AI readiness is therefore less about reaching a universal cleanliness score and more about establishing disciplined evidence for a specific decision. One auditable workflow creates a foundation the dealership can extend. The data improves because the operating process reveals exactly where improvement matters.

Questions for the next operating review

• What recurring decision could the team define with a small, traceable set of source fields in the next thirty days?
• Which business terms produce different answers across systems or rooftops, and who owns the definition?
• When required data is missing or stale, does the workflow expose the limitation or quietly produce an answer anyway?