About 19 % of car buyers, and 25 % of new-vehicle buyers, used an AI tool during their search, and those buyers report the highest satisfaction levels, citing real-time answers and personalised recommendations (Cox Automotive, Car Buyer Journey Study, 2026). Buyers are not afraid of automation. They are afraid of a reply that does not speak to them. The two fears look alike. They call for entirely different remedies.

A tone survives automation on one condition: treat it as an explicit setting, decided and written down, rather than a side effect of the software. Whoever does not decide on a tone inherits a default one, and a machine’s default tone is the tone of a form.

What the buyer reads first is not your sales pitch

They read a rhythm. A length. A level of familiarity. Before processing a single fact about the gearbox, they know whether the message in front of them comes from someone who read theirs.

The signals that give away a manufactured reply are always the same, and none of them concerns vocabulary. Four hundred words in answer to “is it still available?”. A follow-up that repeats the previous follow-up word for word. Three questions in a row when the buyer asked one. A meeting proposal glued to the end of a message that called for no action at all.

None of this is a style problem. These are problems of proportion and sequence.

Tone is not a coat of varnish, it is a set of rules

A brand voice is described better by what it forbids than by what it intends. “Warm and professional” steers nothing. “Never ask two questions in the same message” steers something.

Register

Formal or familiar, short sentences or developed ones, technical language assumed or explained. An independent dealer specialising in German premium cars does not write like a generalist selling 6,000 euro city cars. The first can mention a timing belt replaced at 120,000 km without explaining what a timing belt is. The second has to explain, without sounding like they are explaining.

Proportion

This is the most neglected setting, and the most visible. Reply length should track the length and density of the request. A short factual question calls for a short factual answer, with no context-setting and no commercial projection. A reply that overflows its frame tells the reader they are talking to a system that unrolls, not to a seller who listens.

Commercial pressure

Not every message has to close. Some exist only to inform or to maintain the relationship, and that is fine. A buyer who mentions where they live is giving information, not requesting delivery. Answering with a transport offer burns the exchange. Mentioning soberly that delivery exists, if it does, is another matter.

These three settings are configured: vehicle by vehicle for openness to negotiation, threshold by threshold for qualification, delay by delay for follow-ups. That configuration work is what produces a voice, not the raw quality of the model.

Should you tell buyers they are talking to an AI?

The “they never notice” argument circulates widely in the trade. It is a bad one.

Bad legally, first. France’s CNIL publishes guidelines on deploying AI systems that process personal data, and informing people about automation, along with the framing of automated scoring, is part of them. On that ground, an automotive professional takes advice from a qualified lawyer rather than from the dealer down the road.

Bad commercially, second. The first two French used-car players to launch an app inside ChatGPT, leboncoin and Aramisauto, hide nothing of how immature the exercise is: Le Journal de l’Automobile noted in February 2026 that some recommendations surfaced during its tests were puzzling, and Olivier Flavier, leboncoin’s automotive managing director, conceded that the platform needs to gather statistics to refine the service. These people sell AI. They state its limits. That is exactly why they are believed.

The right question is not whether it shows. It is whether it holds up. A buyer who realises afterwards that they were talking to an assistant, and whose exchange was accurate, precise and proportionate, does not feel deceived. A buyer who receives three identical follow-ups knows before they find out.

A good tone will not fix a 42-hour delay

The average response time across the firms audited by Harvard Business Review stood at 42 hours, and 23 % of them never replied at all; those replying within the hour were roughly 7 times more likely to hold a qualifying conversation (Oldroyd, McElheran, Elkington, 2011). An earlier study by MIT Sloan and InsideSales, led by Dr James Oldroyd in 2007, measured that contacting a lead at 5 minutes rather than 30 multiplied the odds of reaching them by around 100, and the odds of qualifying them by around 21.

A perfect tone delivered two days later is worth nothing. A decent tone delivered in minutes is worth a sale. Osmosia replies in under 15 minutes, seven days a week from 7am to 11pm, with the register, the proportion and the level of pressure the seller has chosen.

The voice stays yours. What changes is the hour at which it speaks.