Start with listings, then lead follow-up, then phone answering. Listings come first because the AI only drafts, you check every line against the car, and the work touches all 20 vehicles. Follow-up is second: a quick first reply wins conversations. Phones come last, because a voice agent speaks to buyers live and nobody reads its words first.
The order only changes if your own numbers say so. If most enquiries reach you by phone and you already miss a third of the calls, a plain missed-call text goes in before anything else (many business phone systems can send one automatically, so check yours). One rule holds at every stage: the AI never discusses finance terms, trade-in values or whether a car can be had for less. Those stay with a person.
Why a 20-car lot should not copy a dealer group's order
Larger dealerships often begin with phones and a business development centre (a BDC: the team that answers calls and chases leads), because they have hundreds of enquiries a month and staff to supervise a new tool. A lot of 20 cars has the opposite shape. There is one owner, perhaps two salespeople, a few dozen enquiries a month, and a stock list in which every car is 5% of the inventory.
That shape changes the sums in three ways. Hours are the scarce resource, so the first job should save owner time without creating something new to supervise. Every public word carries the owner's name, with no compliance team behind it. And one badly described or late-listed car is a twentieth of the lot, so errors in the stock data hurt more than they would on a lot ten times the size.
Three questions that put the jobs in order
Ask the same three questions of each job. Who reads the words before a customer does: you, a salesperson, or nobody? How often does the job come round in an ordinary week? And if it goes wrong, what does it cost and can you take it back? A job that you read first, that recurs every week and whose mistakes can be corrected scores well. Phone answering fails the first and third questions.
| Job | Who reads it first | How often (20-car lot) | A mistake costs |
|---|---|---|---|
| Listings | You, before it goes live | About two new cars a week, plus price changes | A correction, unless a false claim has already reached a buyer |
| Lead follow-up | You or a salesperson, before it is sent | 10 to 20 enquiries a week | A sent message cannot be recalled, and a finance promise is hard to walk back |
| Phone answering | Nobody, in real time | 30 or so calls a week | A promise made aloud to a caller, with no draft to catch it |
Running cost follows the same order. ChatGPT Plus and Claude Pro both list at $20 a month for individuals and Google AI Pro at $19.99 (list prices, so confirm on each vendor's page), and each has a free tier, which is enough for listings and follow-up. On the phone side, a general AI receptionist such as Smith.ai includes 25 calls a month on its free plan and then charges $3 a call, with a Pro plan from $150 a month. Services built for dealerships, usually sold as an AI BDC, publish prices of roughly $500 to $800 a month as of October 2026. That is a big fixed cost to attach to the one job nobody reviews.
A week at a 20-car lot, timed
Here is the scoring applied to one imagined week. The lot holds 20 cars, sells about nine a month and has the owner plus two salespeople. This week two cars arrive and need listing, 15 enquiries land (website form, marketplace messages, email) and 30 calls come in, eight of which go unanswered because everyone is with a customer. These figures are illustrative, and a one-week tally of your own beats them.
| Job | Minutes now | With AI drafting and your check | Saved |
|---|---|---|---|
| Listings, 2 cars | 80 (40 each) | 40 (20 each) | 40 |
| Follow-up, 15 first replies and 12 chasers | 123 (5 per first reply, 4 per chaser) | 48 (2 per first reply, 1.5 per chaser) | 75 |
| Phones, 30 calls | 90 (3 per call) | 84 (the agent takes 12 calls, plus 30 minutes reading its call log) | 6 |
Listings save 40 minutes, follow-up 75 and phones six, after the half hour you spend checking what the voice agent did. Time is only part of the picture, though. Follow-up changes how soon a buyer who has messaged several lots hears from you, listings change how soon a car is live, and phones change the least while carrying the most risk. Until a number says otherwise, that is the order.
Test the phone case with a sum before you buy anything. Eight missed calls a week is about 32 a month. Suppose you guess that one caller in ten would have come to view the car and one viewer in four would have bought: that is 3.2 viewings and 0.8 of a sale, or about $1,600 of gross profit at $2,000 a car. Both guesses are yours to replace with a four-week tally. A prize that size justifies a $3-a-call plan or a missed-call text. A service at $500 to $800 a month needs one extra sale every four months at the low end, and one every two and a half months at the high end, just to cover its fee, before counting the time you spend supervising it. The AI BDC tutorial runs that break-even sum for a bigger enquiry count.
Listings first: give the AI facts, not adjectives
The quickest way to a bad listing is to ask for "a great description of a 2017 hatchback". The assistant fills the gaps with plausible words: one owner, full history, immaculate. A listing is a public statement about a car you are selling, so each claim has to come from your own records. The fix is a fact sheet filled in on the day the car arrives, and a prompt that forbids anything not on it.
Fill one sheet per car from the paperwork and your own inspection. I'd always include the stock number, year, model and trim, fuel and gearbox, the odometer reading with the date it was taken, number of keys, number of owners, the service history as it really is (stamps, invoices or neither), work you have done, the safety inspection result with any advisories, every known blemish with its photo number, whether a vehicle history report has been run, the price and your written warranty terms. Here is a filled-in sheet and the prompt that goes with it.
CAR FACT SHEET (our own records; add nothing else)
Stock no: L-014
Vehicle: 2017 five-door hatchback, 1.2 petrol, manual, 5 seats
Odometer: 62,418 miles (read on [date])
Colour: dark blue. Keys: 2
Owners: 2 (from the registration document)
Service history: 4 stamps, last at 55,900 miles; no invoices held
Work done by us: new front pads and discs, 2 rear tyres, full valet
Safety inspection: passed on [date]; one advisory, light surface rust on rear subframe
Known blemishes: scuff on rear bumper (photo 7), small chip in windscreen (photo 9)
History report: not run yet
Price: $9,495
Warranty: 3 months, parts and labour, per our written termsYou write used-car listings for a small independent lot. Use only the facts on the sheet below.
- Never add equipment, history or condition claims that are not on the sheet.
- If a fact is missing, write "ask us" for that item.
- Quote the odometer reading exactly as given.
- Name every blemish and advisory.
- Do not mention finance, payments, trade-ins or negotiating.
Write three things: a headline of 15 words or fewer, a description of 120 to 150 words in plain English, and a two-line social post.
[paste the fact sheet here]Headline: Two-owner 2017 hatchback, 62,418 miles, new brakes fitted
Description: This dark blue five-door hatchback has a 1.2 petrol engine and a manual gearbox. It has had two owners and shows 62,418 miles. The service book has four stamps, the last at 55,900 miles, and we hold no invoices. We have fitted new front pads and discs and two rear tyres, and the car has been valeted. It passed its safety inspection with one advisory: light surface rust on the rear subframe. Cosmetic marks are a scuff on the rear bumper (photo 7) and a small chip in the windscreen (photo 9). Two keys are included. Vehicle history report: ask us. It comes with a three-month parts and labour warranty under our written terms. Price $9,495. Message us to arrange a viewing.
Social post: Two-owner 2017 hatchback, 62,418 miles, new front brakes and two new rear tyres. $9,495. Message us to book a viewing.That output is illustrative, and deliberately plain. Every sentence traces to a line on the sheet, "ask us" fills the one gap (the history report), and the blemishes are named. Run the lazy version of the same request and the failure is easy to see. Asked loosely for "a friendly listing", an assistant is likely to produce lines such as "lovely one-owner hatchback in immaculate condition with full service history, accident free, finance available". Against the sheet, "one-owner" is wrong because there were two, "full service history" overstates four stamps and no invoices, "accident free" claims something nobody has checked because the history report has not been run, and "finance available" is a finance statement nobody approved. A buyer who asks to see the service invoices has caught you out before the viewing starts.
Checking takes about five minutes a car, and a sentence with no source goes. On this lot that moves a listing from roughly 40 minutes to roughly 20, and the habit matters more than the prompt wording. This is the list I would print and keep at the desk.
- Every sentence has a source line on the sheet.
- The odometer reading in the text matches the dashboard photo, digit for digit.
- The history, service and warranty wording matches the paperwork, not the adjectives.
- Every blemish and advisory on the sheet appears in the text and has a photo.
- The price in the text matches the website and the windscreen sticker.
- Nothing mentions finance, trade-ins or negotiating.
Ask for three versions in one go and run the same check on each. The listings and photos tutorial covers source tags and photo edits in more depth; at this size the fact sheet is the piece that pays back first.
Follow-up second: drafts you approve from your phone
Follow-up comes second because it reuses what the listing stage built: the fact sheet is the source of truth for the reply too. The routine is short. Send every enquiry (website form, marketplace messages, email) to one inbox or one sheet with the date and time, first name, channel, car and message. Write a half-page of reply rules once: opening hours, how test drives work, what ID you need, how deposits work and the lines you never cross. Then, for each enquiry, paste the message, the car's fact sheet and the rules into the assistant and ask for a reply of under 70 words that proposes two viewing times and ends with a question.
ENQUIRY (website form, 9.12 pm): "Hi, is the dark blue hatchback (L-014) still for sale? Would you take 8k? I'd need finance as well. [first name]"
DRAFT REPLY, before checking:
Hi [first name], thanks for asking about the dark blue hatchback. It is still here as of this morning and I'll confirm again before you travel. The advertised price is $9,495, and I'd rather talk price in person once you've seen the car. Finance is no problem, we work with several lenders and most people get approved. Would Thursday at 5.30 pm or Saturday at 11 am suit for a viewing? Bring your driving licence if you'd like a test drive. [your name]Three things in that draft need a person. The finance sentence goes, because "most people get approved" promises an outcome that belongs to a lender. "Still here as of this morning" stays only if someone has looked at the car today, since the buyer will drive over to find it. And "I'll confirm again before you travel" is a commitment you have just made on your own behalf, so keep it only if you will really do it. I'd keep a short list of phrases like these next to the reply rules and add to it whenever a draft slips one in.
Chasers follow the same pattern. Draft a second message two days later and a third a week after the first, each with something new: a better photo, an inspection detail, a viewing slot that has just opened. Never invent urgency. "Another buyer is viewing it on Saturday" belongs in a message only if it is true. A sensible target is a first reply within 15 minutes during opening hours, which a phone alert on the shared inbox makes realistic. When volume justifies it, the first-reply automation tutorial shows how to move these drafts into a form-to-CRM flow. At 15 enquiries a week, copying and pasting is fine.
Phones last: the calls a machine may take
A phone agent differs from the first two jobs in the way that matters most: it speaks to the customer in real time and nobody reads the words first, so everything it says is final. That is why it comes last, and why the first version should be modest. Decide call by call what it may do.
| The caller asks | Who handles it | Why |
|---|---|---|
| Opening hours, address, directions | The machine alone | Facts that never need judgement |
| "Is the blue hatchback still there?" | The machine only if it reads a live stock list; otherwise it takes a message | A wrong "yes" sends someone on a wasted trip |
| Book a viewing or test drive | The machine takes details, a person confirms | Someone has to check the car, the keys and the diary |
| "What's your best price?" | A person | Anything that moves the price is a commitment |
| Trade-in value | A person, after seeing the car | Value depends on condition and paperwork |
| Finance, payments, approval | A person, or a lender | Regulated wording and a lender's decision |
| A problem after a sale | A person, the same day | The cost of getting it wrong is the customer |
The risk is not hypothetical. In December 2023 a visitor told a dealership's website chatbot to agree with everything he said, and it "agreed" to sell a new car for $1. The story was widely reported, the dealer did not honour the "deal" and the chatbot was switched off afterwards. The lesson holds for phones as well as chat: an AI that talks to customers can be pushed off script, so the safest design gives it nothing it can promise. The chatbot rules tutorial lists the limits to set before anything goes live.
Run the phone stage in four steps. Tally missed calls for four weeks, with the date, time and reason if you know it. Start with a missed-call text, which costs almost nothing. If the tally still justifies a voice agent, limit its brief to the first three rows of the table, set it to transfer on everything else, and listen to every call for the first two weeks. Finally, have it say at the start that the caller is speaking to an automated assistant. The EU AI Act's transparency rule for such systems has applied since August 2026 if you have customers in the EU, and it is the honest thing to do wherever you trade. The receptionist cost tutorial prices this at three call volumes using vendor list prices.
What stays with a person from the first day
Some wording never goes out under the AI's hand, whichever stage you are at. Keep a standing list beside the prompts, and treat any line on it that appears in a draft as a reason to cut it, not to soften it.
| If a draft says | Why it comes out | Say instead |
|---|---|---|
| "You'll be approved easily" | Approval is a lender's decision | "We can talk through payment options when you visit" |
| "From $189 a month" | Any payment figure needs a person responsible for finance wording | Nothing, unless that person supplies and approves it |
| "We'll give you $4,000 for your trade-in" | Value depends on inspection | "Bring it in and we'll value it properly" |
| "Accident free" | Nobody has checked | What the history report says, if one has been run |
| "I can do $8,000 for you" | A commitment made in writing | "Come and see it first and we'll talk price in person" |
| "Still available" | Not checked today | "Let me confirm and come back within the hour" (only if you will) |
Advertising rules for finance, vehicle history and warranty claims differ by place and change over time. Ask your finance provider or a qualified adviser what the wording must carry, and keep AI drafts away from it until you have an answer.
When the order flips: three lots that should not follow it
The order above is a default for a lot with mixed enquiry sources and no service department. Three common shapes change it, and each shows the three questions at work.
A roadside lot where most enquiries are phone calls and about four in ten go unanswered should set up the missed-call text in its first week, alongside listings, because the loss there is measurable on day one. It still leaves the voice agent for later: a text and a faster call-back cost almost nothing and carry no spoken promises.
An online-first lot that sells through marketplaces and its own website, and rarely takes a call, can start with follow-up. Its listings may already be decent, and the 60 or so web enquiries a month are where a slow reply loses the sale. Listings come second and the phone stage may never be needed.
A lot with a service bay changes the phone sum. Booking an oil change or an inspection is a routine, bounded call, so the voice-agent case is stronger there than it ever is for sales calls. The voice agent tutorial for service bookings gives the call flow and a test script. Even then, keep sales calls on the human side of the hand-off.
Three gates between one job and the next
Moving on should be a decision with a test, not a mood. These are the gates I'd use on a lot of this size.
- Listings to follow-up: four listings in a row needed no factual corrections, and every car went live within one working day of arriving.
- Follow-up to phones: four weeks with a median first reply under 15 minutes in opening hours, and a missed-call tally that is still above about five a week (my rule of thumb, so adjust it to your own sums).
- Pause rule at any stage: a finance promise, a wrong fact or an invented deadline that reaches a customer pauses that stage for a week while you tighten the prompt and the reply rules.
One sheet with four numbers is enough to run the gates. Here is how a lot might fill it in over a month; the figures are illustrative.
| Measure | Week 1 | Week 4 |
|---|---|---|
| Cars live within one working day of arriving | 1 of 2 | 2 of 2 |
| Factual corrections per listing | 3 | 0 |
| Median minutes to first reply in opening hours | 190 | 12 |
| Calls missed | 8 | 6 |
On those numbers the lot has passed the first gate and the second, but the missed-call count has barely moved, which makes it the next candidate. Whether it is worth a $3-a-call plan, a $150-a-month plan or nothing at all is the sum from earlier, run with the lot's own gross profit per car. If it comes out as a fraction of a sale a month, a missed-call text and a better voicemail greeting may be all that phones ever need.
Questions a small lot owner asks before starting
Should a small lot buy an AI BDC service first?
Usually not. Published prices for dealership AI BDC services run roughly $500 to $800 a month, and the fee only pays when enquiries and calls are already going unanswered. Start with listings and follow-up using an assistant you can already afford, tally missed calls for four weeks, then run the break-even sum with your own gross profit per car. If the sum is close, trial a cheaper pay-per-call plan first.
Can AI quote monthly payments or say a buyer will be approved?
No. Keep every payment figure, rate and approval statement out of AI drafts and out of any phone agent's script. Approval is a lender's decision, and finance advertising is regulated in many places, so the wording needs a person who is responsible for it. The AI can say that payment options are discussed in person, and nothing more specific than that.
Which assistant should the lot use for listings and replies?
Any mainstream assistant will do, because the quality comes from your fact sheet and rules rather than the brand. Individual paid plans of ChatGPT and Claude list at about $20 a month and Google AI Pro at $19.99, and each has a free tier. Before pasting anything, find the model-training switch in the privacy settings, and keep customers' phone numbers and addresses out of prompts while you test.
Do I have to tell customers they are talking to an AI?
Whenever a customer is talking to an automated assistant, such as a phone agent or website chat, say so at the start. The EU AI Act's transparency rule for such systems has applied since August 2026 if you have customers in the EU, and openness is the better habit anyway. Replies you draft with AI, read and send yourself are different, because you are the sender. If you are unsure how a rule applies to you, ask a qualified adviser.
Further reads
- How Can AI Help Car Dealerships? A 30-Day Plan for a Small Lot — A 30-day schedule once you have chosen the first job.
- ChatGPT for Car Sales: Follow-Ups, Listings and What Never to Promise — Prompt wording for follow-ups, and what never to promise.
- ChatGPT for Car Dealerships: What Customer Data Never to Paste In — Which customer details to keep out of any prompt.
- How to Use AI in Car Sales: Trade-In Enquiries Before the Appraisal — Gather trade-in details without the AI quoting a value.
- AI for Used Car Sales: Price Each Car Against Local Listings — Price each car against comparable adverts before you list it.
- AI Marketing for Car Dealerships: A Four-Week Plan for One Lot — A four-week marketing plan for one lot.
- Responding to Dealership Reviews With AI: Samples and Rules — Sample replies to reviews, with hard rules on faults and finance.
- AI Tools for Car Sales: Test-Drive Follow-Ups, Quotes and Call Notes — A salesperson's routine for AI follow-ups after a test drive and tidy call notes, with prompts, samples and the lines AI must never write.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: ChatGPT, Claude, Google AI and Smith.ai pricing pages; published pricing pages of dealership AI BDC services; press coverage of the December 2023 dealership chatbot case; EU AI Act, Article 50.