Five car service tasks where a model removes routine
- A request in the customer's words. A voice note or a messy chat message becomes a record: make, model, year, mileage, what is wrong, when it is convenient to come. The mistake here is a missed question that makes the customer arrive without the right documents.
- A draft work order. The customer's complaint goes into the 'as reported by the customer' field. The mechanic adds the list of works after inspection.
- A parts request. The model turns 'knocks at the front' into a list of components worth checking and reminds staff to ask for the VIN. It does not name the part number itself; more on that below.
- An explanation to the customer. The mechanic writes a short note ('pads 2 mm, disc scored, replace both'), and the model turns it into a clear message for the customer with the photo the mechanic took.
- A service reminder. From the date of the last visit, the mileage and the interval in the manufacturer's service documentation, the model writes a personal reminder.
The first two tasks are described in the Rules for Providing Maintenance and Repair Services for Wheeled Vehicles (Order of the Ministry of Infrastructure No. 615, status 'in force'): services are provided on request or on a prior booking recorded in a log, and based on the booking the provider sets the date and time to bring the vehicle in. So recording the request is not a formality but the first document in the chain.
Work order: which fields the model fills in and which the mechanic does
Article 25 of the Law 'On Road Transport' lists the essential terms of a maintenance and repair contract. A work order can serve as that contract. We split these terms by who can fill them in and from what.
Rows are the essential terms from Article 25 of the Law 'On Road Transport', grouped in pairs. ✓ fills in; ≈ prepares a draft a person checks; ✗ does not fill in.
The logic is simple. Whatever the customer said themselves, a model records more accurately and faster than a person holding the phone with their shoulder. Whatever is a promise of the service (works, price, parts, warranty) is an obligation under the contract. The same Article 25 forbids the provider to impose additional paid services on the customer, so model 'recommendations' such as 'replace the discs while you are at it' without an inspection do not go into the draft.
What it costs per month
A hypothetical four-bay service: 300 requests a month, all by voice, one minute each. A transcription model transcribes each request, then a text model makes a draft work order: 1,500 tokens in (transcript, rules, form fields) and 400 out. Plus 120 explanations to customers (800 in, 300 out) and 200 service reminders (600 and 150). The volumes are hypothetical; Ukrainian tokenises less efficiently than English, so recalculate on your own texts.
Prices from the OpenAI API pricing page on 24 October 2026 (standard processing, short context, per million tokens): gpt-6-luna, $0.10 input and $0.50 output; gpt-6.1-sol, $2 and $10; gpt-6-astra, $10 and $50. Transcription with gpt-transcribe costs $0.0045 per minute. The National Bank rate on 24 October 2026 is UAH 45.051 per dollar.
| Model | One draft work order | One explanation | One reminder | Month including transcription |
|---|---|---|---|---|
| gpt-6-luna | UAH 0.016 | UAH 0.010 | UAH 0.006 | $1.51 (UAH 68) |
| gpt-6.1-sol | UAH 0.32 | UAH 0.21 | UAH 0.12 | $4.54 (UAH 205) |
| gpt-6-astra | UAH 1.58 | UAH 1.04 | UAH 0.61 | $17.31 (UAH 780) |
Calculation: 300 × 1 min × $0.0045 + (300 × (1,500 × input + 400 × output) + 120 × (800 × input + 300 × output) + 200 × (600 × input + 150 × output)) ÷ 1,000,000, multiplied by 45.051. Volumes are hypothetical; OpenAI API prices and the NBU rate on 24.10.2026.
The unusual thing in these numbers is that on the cheap model UAH 61 of the UAH 68 (89%) is voice transcription, not text. If customers write instead, the sum falls to UAH 7. So the model choice here depends on quality, not price: check which one confuses component names and mileage figures less on your real voice notes. The sum does not include the app that connects the messenger, the model and your accounting software: take that from the pricing of the platform you choose.
Parts: a catalogue, not a model
A model knows a Skoda has front brake pads. But the specific part number depends on trim, year and disc diameter, and a mistake costs a returned part and an idle bay. A model can name a part number confidently and be wrong, and nobody can check it without the catalogue.
So the split is:
- The model turns the customer's description into a list of components for the mechanic to check and reminds staff to ask for a photo of the registration certificate with the VIN, or the VIN itself.
- The parts manager looks the part up in an electronic catalogue by VIN.
- The mechanic confirms it is the right part and enters it in the work order.
Rules No. 615 also have a pricing nuance. If parts are unavailable when the work order is drawn up, or their quantity cannot be determined, the quantities are agreed with the customer in advance and paid after the work at the prices on the day the parts arrive, unless the contract says otherwise (Section V, item 6). The model can mention this in a message to the customer, but does not name the price.
Extra work and explanations to the customer
The most common conflict at a car service is 'we also did this for you'. The Rules are clear: if work not covered by the work order becomes necessary, the provider informs the customer by phone, and only with their consent is an additional contract concluded (Section III, item 6).
The model helps with the text: from the mechanic's note and photo it writes to the customer what was found, why it matters and what it costs according to the price list you gave it. What it does not do:
- It does not diagnose from a photo. The photo illustrates the mechanic's conclusion; it is not input for the model's own conclusion.
- It does not record the customer's consent. Consent is a person's answer that the manager records in the work order or an additional contract.
- It does not scare. Phrases like 'unsafe to drive' are for the mechanic who has seen the component.
Service reminders work the same way: the interval comes from the manufacturer's service documentation for the specific car, not from the model's memory. The model only calculates when the next visit is due from your data (mileage at the last visit, date, average monthly mileage) and writes the message.
Documents and personal data
To receive a service, the customer shows a passport or driving licence and the vehicle registration document (Rules No. 615, Section III, item 2). These are the most sensitive documents at a car service, and they do not go into an assistant. A licence plate, VIN and mileage next to a name and phone number are personal data: the Law 'On Personal Data Protection' defines them as information about an individual who is identified or can be specifically identified.
In practice, the request to the model needs only make, model, year, mileage and complaint. Your software inserts the name, phone and plate into the finished message. How to choose a plan and settings so the data does not go into model training is covered in the article on what data to give an AI assistant. On choosing between a chatbot builder, a ready assistant and custom development, see the article on a chatbot for business.
What to give the model and what not to
- Is the result a record of what the customer said?Complaint, make, mileage, convenient time. Yes → model, checked by the mechanic at intake
- Is the result a promise by the service?List of works, price, deadline, warranty. Yes → mechanic only; the model provides a text template
- Do you need to name a specific part or part number?A mistake costs a returned part and an idle bay. Yes → VIN catalogue, not the model
- Do you need to decide what is broken?A diagnosis from a description or photo. Yes → mechanic; the model writes an explanation from their conclusion
- Will a passport, driving licence or registration document go into the request?They are not needed for a draft. Yes → remove them from the request
A first step for a week
- Collect 20 past requests from the messenger, including voice notes.
- Run them through transcription and two models with the same instruction: fill in only 'make, model, year, mileage, complaint, preferred time' and list what data is missing.
- Compare with what the mechanic recorded at intake. Count two things: how often the model invented a detail the customer did not say, and how often it missed something they did say.
- Put the model with the lower numbers to work. For a week, handle requests in parallel: model and person.
Our position
The best place for a model at a car service is between the customer and the request log: transcribe the voice note, fill in the fields the customer named, and ask them the questions that are missing. Everything that is the service's obligation (works, price, parts, warranty, consent to extra work) stays with the mechanic, because these are essential terms of the contract. A more expensive model is not needed while the cheap one does not confuse components and figures on your voice notes.
The position does not hold for warranty service at an official dealer: there the work order and list of works come from the manufacturer's system, and your own assistant may contradict it. It also does not hold if customers mostly phone rather than write: then you need call transcription and a different calculation, as in the article on call analysis. The cost estimate assumes short voice notes and the volumes named above.
Prices and the exchange rate were checked on 24 October 2026 on the OpenAI API pricing page and at the National Bank of Ukraine; the rules were checked against the texts of the Law 'On Road Transport', the Rules for Providing Maintenance and Repair Services for Wheeled Vehicles (Order No. 615) and the Law 'On Personal Data Protection' on zakon.rada.gov.ua. The arithmetic was recalculated in code. The number of requests, voice note length and tokens are hypothetical.