Most dealership AI projects do not fail because the model is weak. They fail because nobody defines the process around it. The store buys a tool, watches a polished demo, assigns no operating owner, and discovers 60 days later that usage is sporadic and the old process is still running beside the new one.

A better 90-day plan starts with one measurable workflow. It gives that workflow an owner, a baseline, clear human handoffs, and a weekly review. The goal is not to “implement AI” across the store. The goal is to make one important job work better and leave behind a repeatable method for the next job.

Days 1-15: Diagnose the work before selecting the tool

Start where Dealer AI Guy’s implementation framework starts: the dealership’s leads, phones, CRM habits, website, service lane, and manager handoffs. Its consulting guide recommends diagnosing and prioritizing before implementation, training, and measurement. That order matters.

Interview the manager who owns the outcome and two employees who perform the work. Observe the actual process instead of relying only on the written SOP. Capture:

  1. The event that starts the workflow.
  2. Every system and person it touches.
  3. Where work waits, gets duplicated, or disappears.
  4. Which decisions require a manager.
  5. What the customer experiences when the process fails.
  6. The current baseline for speed, volume, completion, and outcome.

Good first candidates are narrow and frequent: missed-call recovery, unsold lead reactivation, declined-service follow-up, daily manager reporting, or inventory-description quality control. Avoid a first project that requires replacing the CRM, changing every department, or resolving years of inconsistent data.

Days 16-30: Design a bounded pilot

Write a one-page pilot charter. It should name the business problem, the scope, the owner, the users, the systems involved, the customer or employee data involved, the human approval points, and the first KPI.

Then decide what AI is allowed to do. There is a meaningful difference between drafting, recommending, and acting.

Level AI role Example
Assist Drafts for human review Produce a suggested follow-up text
Recommend Ranks or flags work Identify leads most likely to need manager attention
Act Performs a bounded task Schedule within approved rules and open time slots

For the first pilot, favor assist or recommend. Autonomous action should come only after the store has tested data quality, exceptions, consent rules, and escalation behavior.

The NIST AI Risk Management Framework organizes responsible AI work around governing, mapping, measuring, and managing risk. A dealership does not need a federal-size program, but it should borrow the discipline: document the use case, identify possible harm, test before launch, and retain a person who can stop the system.

Days 31-45: Build the workflow and the exception path

Configure the smallest version that can produce evidence. Do not automate every branch. Build the normal path and the three most common exceptions.

For each step, specify:

  • Input: what approved data enters the process.
  • Output: what the AI produces or changes.
  • Reviewer: who checks the result, if required.
  • Escalation: what conditions route the work to a person.
  • Record: where the action and outcome are logged.

Test with real but appropriately protected examples. Include difficult cases: an angry customer, a missing price, a duplicate lead, an unavailable vehicle, a Spanish-language request, a financing question, and a request to stop messages. The happy path proves the demo works. The exception set proves the process can survive the store.

Days 46-60: Train by role, not by feature

Managers need to know the operating rules and scorecard. Frontline users need to know when to use the workflow, how to review output, and when to escalate. Administrators need to know permissions, logs, and shutdown procedures.

Run training on the store’s own scenarios. Give each role a one-page SOP. End every session with a live task and a manager sign-off. Training should answer four questions:

  1. What job is this system helping us do?
  2. What does a good output look like?
  3. What may never be sent or decided without a person?
  4. Who owns the problem when the output is wrong?

Days 61-75: Run in controlled production

Launch to one team, one rooftop, or one shift. Review a sample of outputs every day during the first week. Track failures as categories, not anecdotes: wrong information, wrong tone, missed opt-out, failed integration, duplicate action, weak escalation, or user bypass.

Car Dealership Guy’s reporting on dealers building their own AI tools emphasizes starting from operational pain and maintaining the system after launch, not simply building a clever prototype. The same principle applies to vendor tools. The store must own the operating standard even when it does not own the software.

Days 76-90: Decide to scale, repair, or stop

Compare the pilot period with the baseline. Use a small scorecard:

  • Adoption: percentage of eligible work processed through the workflow.
  • Speed: time from trigger to first completed action.
  • Quality: percentage passing manager review without correction.
  • Outcome: appointments, completed follow-ups, recovered opportunities, or time saved.
  • Risk: exceptions, complaints, consent failures, or incorrect claims.

Do not scale because employees “like it.” Scale when adoption is durable, quality is acceptable, the outcome moved, and the risk log is controlled. If the result is weak, determine whether the problem is the tool, the data, the process, or management follow-through.

Ninety days should end with evidence, an owner, an SOP, and a decision. Anything less is still a demo.

The deliverables leadership should receive

At day 90, the dealer principal or GM should have the pilot charter, current process map, approved-tool and data rules, SOP, training record, exception log, scorecard, vendor decision, and recommendation for the next workflow. That package makes the work transferable across managers and rooftops.

Sources and further reading

This article is general operational guidance, not legal or compliance advice. Review customer communications, data use, advertising, credit, employment, and privacy requirements with qualified counsel.