AI Buyers Will Grade Your Answers Before Your Checkout

There is a piece in CBT News from 1 September 2026 arguing that the industry is crossing from AI shopping to AI buying. The case is that assistants have graduated from comparing trims and estimating payments to actually completing purchases, and that a dealership whose website cannot support the whole transaction will find itself part of the research journey but not the buying journey. AI, the argument goes, will take the path of least resistance.

The direction is right. The timing is not, and the gap between those two is where a lot of next year's budget is about to go astray.

What an agent can actually do this quarter

An agent acting for a buyer cannot complete a credit application on a dealer website. It does not have the buyer's Social Security number, it does not have signing authority, and the platforms building these assistants are moving away from letting them type sensitive data into third-party forms rather than toward it.

The end-to-end checkout is a real destination. It is not the thing being exercised in the next twelve months.

What an agent can do, at a scale no human shopper has ever managed, is ask. It can put the same five questions to eight stores in the time it takes to read a paragraph, tabulate the answers, and hand its buyer three rows.

The five questions that decide a shortlist

They are always roughly the same five.

• Is that exact VIN physically on the lot today, or in transit, or sold on Friday and still live on the feed
• What is the real out-the-door number, including accessories already fitted
• Will the store commit to a trade range sight-unseen, or is that strictly an in-person conversation
• Can the vehicle be seen at 6:40 on Thursday, after the buyer finishes work
• How long will the store hold it

Trace each of those through a real dealership. Nearly all of them route to a phone number or a form. Which means the agent's entire experience of the store is its response layer, not its homepage, not its merchandising, and not its checkout.

Why latency behaves differently when the counterparty is a machine

The MIT Lead Response Management study analysed three years of data across six companies, more than 15,000 web leads and more than 100,000 call attempts. It found the odds of contacting a lead drop by roughly 100 times when the call goes out at thirty minutes instead of five. The study measured contact and qualification, not close rates, which is worth stating precisely because that figure is frequently misquoted.

Contact is exactly the mechanic in play here, and the shift is in what a slow response now costs.

For a human buyer, slow response is friction. They already liked the store, so they wait, they call back, they forgive it. The lead-response curve is a conversion curve.

For an agent, slow response is not friction. It is absence. The agent is not emotionally invested in any rooftop and is not coming back. It asked, it got nothing inside its window, it wrote no response in the cell and moved on.

The store does not receive a bad row in the comparison table. It receives a blank one. A blank cell loses to a filled cell every time, including to the store down the road whose answer was worse than the one the first store would have given.

The honest counter-case

Three objections deserve stating.

First, the original argument is right in the long run. When agents can genuinely transact, the store that can be transacted with wins, and the work of building that will not be wasted.

Second, and more seriously, an answer layer that is fast and wrong is worse than no answer at all. A human hearing a wrong price on the phone hesitates and asks again. An agent records it, repeats it verbatim to the buyer, and now there is a documented number the store never meant to give, at machine scale.

Third, some of those questions genuinely should not be answered without a human. Declining to give a trade range sight-unseen is a legitimate answer, as long as it is given inside the window rather than left as silence.

The work order

It is narrower and much cheaper than a new commerce stack.

• Write down the five questions and name what answers each one within ten minutes, including at seven in the evening and on a Sunday
• Check that every channel gives the same answer. An agent cross-references the website, the syndicated feed and the phone, and disagreement between them reads as unreliability in a way it never did to a human who only ever saw one
• Measure question answered within ten minutes as a metric in its own right, separate from lead response time. A lead form and a question are different objects and only one of them is currently instrumented
• Log the questions that could not be answered. That log is the roadmap, discovered rather than assumed

The uncomfortable version is short. The industry is preparing to spend heavily on the last mile of a transaction that machine buyers cannot complete yet, while the first mile, a fast and consistent answer to a specific question about a specific car, is still routed through a phone that rings out after six.

Build the checkout. Just do not build it first.