You have seen this movie before - it is just running faster

The last major distribution shift in automotive retail arrived slowly enough to feel optional. Newspaper classifieds weakened over years. Dealer websites improved in uneven steps. The early movers got better at digital while everyone else still had time to catch up.

This shift does not have that kind of runway.

Stanford's 2025 AI Index measured the cost of querying a model with GPT-3.5-level performance at $20 per million tokens in November 2022. By October 2024, the comparable cost was $0.07. That is a decline of more than 280 times in roughly eighteen months.

That number matters, but not for the reason most AI pitches use it. Cheaper intelligence does not automatically make a dealership better. It makes small, specific software economically possible.

The old economics favored generic software

Until recently, building a system for one dealer group was difficult to justify. The discovery, engineering, testing and maintenance costs needed thousands of customers on the other side. So a vendor built one product for the middle of the market, sold it broadly and exposed a settings page for the differences.

That model produced useful software. It also produced the familiar last-mile gap: the product handles the general process, while the store still carries a spreadsheet, a manual reconciliation and three exceptions nobody told the implementation team about.

The cost collapse changes the minimum useful market. A tool no longer needs a national customer base to pencil. It may only need to remove one recurring task, settle one disputed number or make one existing vendor work better for one group.

Do not confuse cheaper building with easy operating

The newspaper comparison has a limit. Lower compute cost removes one barrier. It does not remove the hard parts: clean data, adoption, permissions, exception handling and trust.

A cheap model pointed at bad CRM data produces fast nonsense. A prototype without duplicate-send guards can burn customer trust before lunch. A dashboard with a different definition of a sale creates another argument instead of ending one.

That is why the order matters. The group needs its transactions, customers, website activity and useful public data in one governed place. Then the cheaper intelligence can work inside definitions the operator controls.

What an operator should do this quarter

1. List five recurring decisions. Not five technologies. Write down the questions that still require a meeting, a spreadsheet or a call to a vendor.
2. Time the manual path. Include gathering, cleaning, reconciling and waiting - not only the final task.
3. Find the source of truth. For a delivered unit, that is normally the DMS. For customer intent, it may be the CRM. Name it before building.
4. Pick one narrow build. Make it produce a receipt, a timestamp or an answer that can be checked.
5. Earn the next build. If the first one is not used, do not add another. Fix the workflow or stop.

The wall between an operator's idea and working software came down quickly. Walking through it still takes judgment. That is the durable advantage now: knowing which problems are real, which numbers settle them and which small system should exist by next month.