Yes, as a second opinion that proposes what to test next, not as a diagnostician. ChatGPT turns fault codes and symptoms into a ranked test list and puts findings into plain customer wording. Don't trust it for torque figures, wiring colours, part numbers, safety-critical procedures or the final cause: only tests on the vehicle settle those.
A fault code names the circuit or condition that a control unit objected to. It doesn't name the part to replace. ChatGPT has never driven, smelled, probed or scoped your customer's car, so it ranks suspects from patterns in text, and it states them in the same assured voice whether it is right or guessing. OpenAI's own help page warns that it can sound confident even when it is wrong. The shop's job is to turn its list into tests, and to keep the decision about what gets replaced.
What the model sees, and what it never will
ChatGPT works only from what you type, paste or upload. It has no link to your scan tool, no memory of the car from last year's service, and no idea how the engine sounded on the test drive. Unless web search is switched on, it answers from training data that stops at a cut-off date. With search or deep research on, OpenAI says it can pull in current web sources and cite them, but those sources are often forum threads and parts-shop pages of uneven quality.
The code itself tells you how far to trust it. A fault code is five characters: a letter for the system (P for powertrain, B for body, C for chassis, U for network) and four more characters. When the second character is 0, the code is a generic one with the same definition on every make, so the model has seen it explained thousands of times. When it is 1, the code is manufacturer-specific and the meaning depends on the vehicle. Codes whose second character is 2 or 3 are split between generic and manufacturer ranges in ways that differ by system, so look those up rather than assume. Manufacturer codes are where ChatGPT is most likely to borrow the meaning of a similar-looking code from another make and present it as fact.
It can also look at photographs, which tempts people to upload a picture of a leaking hose or a worn tyre. OpenAI's image-input page accepts PNG, JPEG and non-animated GIF files up to 20 MB each, says results may be less accurate when an image is ambiguous or unclear, warns that descriptions can be incorrect and counts approximate, and notes that images are resized before analysis. Treat a photo identification as a lead, never a finding.
Six jobs where it saves a technician's time
None of these asks it to decide anything. Each produces something a technician or service adviser can read in under a minute and either accept or bin.
- A ranked test list from codes and symptoms. Give it the vehicle (year, make, model, engine), the complaint in the customer's own words, the codes, the freeze-frame values and any readings. It returns checks in order of cost and likelihood, with what a pass and a fail would mean. The value is in the ordering and in the "what would a fail look like" lines.
- The questions nobody asked at the counter. "Only when it's cold", "only after rain", "only when the heater is on". Ask for ten questions a service adviser should put to a customer who reports hesitation under load. Better answers at the counter shorten the diagnosis more than any clever prompt, and an intake form that captures the complaint clearly makes that routine.
- A plain-English explanation for an apprentice. "Explain how long-term fuel trim is calculated, what a high positive number tells us, and give three questions to check understanding." Explanations of how a system works are well covered in text, and mistakes in them are easier to spot than mistakes in a diagnosis.
- A bay checklist from a manual excerpt you paste. The model turns a page of procedure into numbered steps and keeps the manual's figures word for word; you then compare the checklist against the page. Check your manual subscription's licence terms before pasting anything into an outside tool.
- Job-card notes in full sentences. Technician shorthand such as "cyl 3 coil, plug ok, 41 misfires" becomes a note the adviser can read to the customer, with no new facts added.
- A devil's-advocate pass. Once you have a theory, ask: "What three other explanations fit these readings that I haven't tested?" If only one use survived, I'd keep this one. It widens the list when you're anchored, and the model has no pride invested in your first theory.
Worked triage: three network codes after a battery change
The Monday job card reads: five-door diesel hatchback, 112,000 miles, warning lights on and a stall at junctions, starting the day after the owner fitted a new battery. The scan shows three stored network codes and no engine codes. Everything below is illustrative. The vehicle is generic, the figures are invented, and the manual excerpt was written for this tutorial, not copied from any manufacturer. Use your own manual's limits.
The prompt and the pasted excerpt
The prompt gives the model the readings and the only source it may use for limits. Rule 3 in the task is the one that does the most work.
You are helping a workshop technician plan tests. You have not seen the car.
Vehicle: five-door hatchback, 1.6 diesel, 112,000 miles. No customer details.
Complaint: warning lights on and a stall at junctions, starting the day
after the owner fitted a new battery.
Scan: U0100, U0121, U0140 stored. No P-codes. None pending.
Readings: battery at rest 12.4 V; cranking 8.7 V; voltage drop from the
negative post to the engine block while cranking 0.6 V; charging at
2,000 rpm 14.1 V.
Manual excerpt (invented for this example; it is the only source of limits):
"Multiple communication codes
1. Check battery condition and state of charge before testing any module.
2. Cranking voltage below 9.0 V: test battery and cables first.
3. Voltage drop, negative post to engine block, while cranking: limit 0.2 V.
4. Charging voltage at 2,000 rpm: 13.5 V to 14.8 V.
5. Only if steps 1 to 4 pass, inspect module connectors and data-line
wiring. Replace a module only after a confirmed internal fault."
Task:
1. For each reading say pass, fail or "not in excerpt", and quote the
line you used.
2. List up to five tests, cheapest first, with what a pass and a fail
would tell me.
3. Where the excerpt gives no figure, write "not in excerpt". Do not
supply a number from memory.
4. Do not tell me what to replace. Tell me what evidence would justify it.The answer that came back
What follows is an illustrative reply, trimmed for space, of the sort that prompt tends to produce.
Readings against the excerpt: cranking voltage 8.7 V is a fail (step 2 sets a floor of 9.0 V). Voltage drop of 0.6 V is a fail (step 3 sets a limit of 0.2 V). Charging voltage of 14.1 V is a pass (step 4 allows 13.5 V to 14.8 V). Resting voltage of 12.4 V is not in the excerpt.
Tests, cheapest first: (1) Look at and wiggle the negative clamp, the body earth strap and the engine earth for movement and corrosion; a pass is nothing moving and bright metal, a fail is visible movement or powdery deposits. (2) Repeat the drop measurement with the probe on the cable end instead of the post, which separates the clamp from the cable. (3) Load-test the new battery; a pass means the battery is not the cause. (4) Clean the connections, run the car for 20 minutes with the lights and heater on, and rescan. (5) Check the resistance across the data pair at the diagnostic socket, which should read about 60 ohms.
Evidence that would justify replacing a module: codes that return after steps 1 to 4 pass, plus a confirmed internal fault under step 5 of the excerpt.
What I'd change before it reaches the bay
- Delete the figure in test 5. The model was told not to supply numbers from memory and did anyway. A resistance value like that can be right for one network layout and beside the point for another. Keep the test, replace the number with "limit: see wiring manual", and take the test conditions from the manual too, since resistance checks on a vehicle network are normally made with the battery disconnected, which the reply never mentioned. Note also that the excerpt allows data-line checks only after steps 1 to 4 pass.
- Add the steps only your shop can add. Record all four readings on the job card, and photograph the clamp before cleaning it. The model can't know your habits.
- Reorder by what's on your bench. If the load tester is already out, the battery test goes first. "Cheapest first" is the model's guess about your workshop.
- Keep the last paragraph. Tying a module replacement to evidence is exactly what you want written down if the customer later asks why you didn't simply swap the module.
What the vehicle said
The negative clamp had a film of corrosion underneath, hidden by the terminal cover, and it was loose enough to turn by hand. The drop across the clamp alone read 0.55 V. After cleaning and tightening, the drop fell to 0.08 V, cranking voltage rose to 10.1 V, charging stayed at 14.1 V, the codes were cleared and none returned after a 25-minute drive.
The sums, at an assumed $95 an hour: 0.9 hours of diagnosis is $85.50, plus a $12 clamp, so $97.50. The route the three network codes invited, replacing a module, would have been an assumed $640 part plus 1.5 hours of fitting and coding ($142.50): $782.50, with the codes still coming back.
The same codes without the readings
Paste only "U0100, U0121, U0140, new battery, warning lights" and the usual reply is a flat list: faulty engine control unit, wiring fault, bad earth, low battery voltage, software update needed. All five are true of almost any set of network codes, none is ranked and none comes with a limit. The model can be only as specific as your measurements, which is why the habit that matters is measuring first and asking second.
Six confident wrong answers and the check that catches each
The first column describes patterns, written for illustration; they aren't quotes from one session. Every row ends in a check against the vehicle, the manual or a bulletin portal, not another prompt.
| What it says | Why it slips | Check before acting |
|---|---|---|
| "The catalytic converter has failed; replace it" (from a catalyst-efficiency code alone) | The textbook cause is the one written about most, so it comes first. | Look for exhaust leaks, misfires and fuel-mixture faults first, and compare the readings from the sensors either side of the converter. The same code can come from several causes. |
| "Tighten the bracket bolts to 85 Nm" (no source given) | Numbers are the easiest thing to invent, and they look right on screen. | Use the manual's figure for that vehicle and year. A number that isn't in your pasted excerpt should be treated as invented. |
| "Code P1xxx means..." (a manufacturer-specific code) | Meanings differ between makes; the model borrows from a similar code. | Look the code up in the manufacturer's service information. Its reading of a P1 code is vocabulary, not authority. |
| "This is covered by service bulletin 12-345" | Reference numbers look real and are easy to make up. | Search your bulletin portal for the number. If it isn't there, treat it as invented and ask the model for the topic instead. |
| "Bank 1 is the driver's side" | Engine layouts differ, and the rule gets stated as universal. | Bank 1 is normally the bank that holds cylinder 1. Find it on the engine and in the manual's diagram before swapping a sensor. |
| "It's most likely the alternator" (echoing the customer's theory in your prompt) | Models tend to go along with whatever theory the prompt contains. | Leave the customer's theory out of the prompt and ask for three alternatives that fit the readings. Then let a test, such as charging output or a voltage drop, decide between them. |
Notice what the checks have in common: each ends at something physical or something written by the manufacturer. None of them is "ask ChatGPT again". A second answer from the same model is not a second source.
Jobs that never go through a chatbot
The test for this list is simple. If a wrong answer could hurt someone, or can't be seen until it's too late, the job is out.
- Torque figures and tightening sequences on wheels, brakes, steering, suspension and driveline parts. The figure comes from the manual for that vehicle, every time.
- Airbag and seat-belt pretensioner work. Follow the manufacturer's procedure, not a summary of it.
- Hybrid and electric high-voltage systems. Isolation and testing procedures come from the manufacturer and are carried out by staff trained for them. An answer on a screen is not a basis for working near an orange cable.
- Fuel-system depressurising and refrigerant handling. Same rule: the manufacturer's procedure, carried out by staff holding whatever certification the work requires, not a chat.
- Key programming, immobiliser work, module coding and flashing. Use the manufacturer's or approved tool's route. A wrong step can leave a car that won't start.
- Wiring colours and connector pin-outs. They change with model year and market, so use the wiring diagram for the exact vehicle.
- Camera and radar calibration distances, targets and tolerances. The procedure is specific and the tolerances are tight.
- Anything signed. Roadworthiness statements, warranty claims and safety advice on an invoice carry a technician's name. The name belongs to a person, so the content does too.
Customer wording: three rewrites that keep the facts
The front counter is where ChatGPT pays its way. The rule: the facts in the message come from the job card, and the model changes tone and order, not content. Every number and part name in the draft must also appear in your notes. A short rewrite prompt does the job:
Rewrite the technician's note below as a short text to the customer.
Keep every fact, number and part name exactly as written. Do not add
causes, prices or promises. Do not soften any safety advice. Plain,
polite, no jargon. Use [first name] and [your name] as placeholders.
Note: [paste the note]Three notes, three results (all illustrative).
A code is not a diagnosis. The technician wrote: "P0455 evap large leak, cap seal looks OK, smoke test needed, 1h diag to confirm". The rewrite:
Hi [first name], we've scanned your car and it has recorded a large leak in the fuel-vapour system, the pipes and valves that stop fuel fumes escaping. The code tells us which system failed its self-check, not which part has failed. We've checked the fuel cap seal and it looks fine, so we'd like to run an hour of diagnostic testing at $[rate], including a smoke test that shows where the vapour escapes, then send you a firm quote before any repair. Reply "yes" and we'll start. [your name], [workshop]
It works because it says what we know and what we don't, asks for approval of one defined step, and promises a quote rather than an outcome.
Fixed, but intermittent. After the earth-clamp job above, the note says the clamp was loose and corroded, readings are in limits, no codes after a 25-minute drive:
Hi [first name], your car is ready. We found a loose, corroded battery earth clamp and corrected it. The readings are now within the manufacturer's limits and the warning lights haven't returned after a 25-minute drive. Because the fault came and went, we can't promise it was the only cause. If a warning light comes back, call us and we'll start from the readings we recorded today.
Compare that with the version a model produces if you don't forbid promises: "Your car is now fully fixed and won't give any more trouble." The honest version costs nothing and avoids a comeback argument.
Declined safety work. The note says the front pads measured [x] mm against a minimum of [y] mm in the manual, the discs are lipped, the customer declined work today:
Hi [first name], thanks for collecting your car. As we discussed, the front brake pads measured [x] mm against a minimum of [y] mm, and the discs are worn at the edge. We advise replacing both before further driving. You've chosen to wait, and we've noted that on your job card. We can fit them any day this week; just reply with a day that suits you.
Here the model's instinct is to soften: "you may wish to consider". Tell it to leave the advice sentence untouched, or write that sentence yourself. Ask your insurer or adviser how declined safety work should be recorded. For messages during a repair rather than after it, how garages use AI to send repair updates covers timing and wording.
Nine lines for the bay and the front counter
Print this, adjust it, and put it where the technicians and advisers will see it.
- The manual beats the chatbot. If they disagree, the manual wins and the disagreement goes on the job card.
- Describe the vehicle by year, make, model, engine and mileage. No names, phone numbers, registration plates or VINs. See whether customer data is safe in ChatGPT for the reasoning.
- Paste only the manual lines you need, and check your subscription's licence terms first.
- Chatbot suggestions are recorded as "suspect", never as "cause". A cause needs a test result.
- No part is ordered on a chatbot's say-so alone.
- Jobs on the never list (torque figures on safety parts, airbags, high voltage, coding and flashing) don't go through it.
- Everyone who uses it has their own login. OpenAI's terms say you may not share account credentials.
- Customer messages polished by ChatGPT are read through by the sender, and every number and part name is matched to the job card before sending.
- A wrong answer that reached the bay is a finding, not a confession. Write it on the tally sheet below so the rule can change.
Plan prices, and a four-week tally of hits and misses
An individual Plus plan lists at $20 a month. ChatGPT Business lists at $20 per user per month on annual billing ($25 monthly), needs at least two users, and doesn't train on business content by default. A free account is enough for wording jobs, but it allows only 3 file uploads a day, where paid plans allow up to 80 files every 3 hours, which matters if you paste manual pages as images or PDFs. On personal plans you can switch off use of your chats for training under Settings, then Data controls ("Improve the model for everyone"); that doesn't delete chats already saved.
At an assumed labour rate of $95 an hour, $20 is about 12.6 minutes of paid time. If the tool saves one technician 13 minutes in a month, the plan has paid for itself. The cost that matters is the checking: each answer needs a minute or two against the manual.
To find out whether it earns its place, keep a tally for four weeks. Every use gets one mark: saved time, no difference, misled but caught at the check, or misled and reached the bay or the customer. Illustrative figures:
| Week | Uses | Saved time | No difference | Misled, caught | Misled, reached bay or customer |
|---|---|---|---|---|---|
| 1 | 9 | 3 | 4 | 2 | 0 |
| 2 | 7 | 4 | 2 | 1 | 0 |
| 3 | 6 | 4 | 1 | 1 | 0 |
| 4 | 6 | 3 | 2 | 1 | 0 |
| Total | 28 | 14 | 9 | 5 | 0 |
Half the uses saved time, five were misleading answers (about 18%), and the check caught all five. That is the routine working as designed, because the check exists to catch exactly those. Any number in the last column means the rule changes that week. If more than a third of uses made no difference, narrow the tool to customer wording only. If you want to compare ChatGPT against other assistants on the same codes, this fault-code test of four assistants does it, and the estimate template tutorial is the natural next job to hand over once the rule is in place.
Questions garage owners ask about ChatGPT
Can I paste scan-tool output straight into ChatGPT?
Yes, as text. Copy the codes, freeze-frame values and live readings into the prompt, and describe the vehicle by year, make, model, engine and mileage. Leave out the customer's name, phone number, registration plate and VIN, because a plate or VIN plus a name identifies a person. The model can't connect to your scan tool, so it sees only what you paste.
Can we paste pages from our paid service-manual subscription?
Check the licence terms of that subscription first. Some providers restrict copying content or sending it to outside services, and the terms are the only authority on that. If they're silent or unclear, ask the provider. Even where it is allowed, paste only the lines the job needs rather than whole pages, and keep the manual open as the source of every limit.
Which ChatGPT plan suits a small garage?
A one-person shop can use an individual plan: Plus lists at $20 a month, and a free account covers wording jobs. With two or more users, ChatGPT Business lists at $20 a user a month on annual billing ($25 monthly), has a two-seat minimum and doesn't train on business content by default. Each person needs their own login, because OpenAI's terms bar sharing credentials.
Does it work for hybrids and electric cars?
It can explain how a system works and draft customer wording, but its knowledge of recent models is thinner and the stakes are higher. Isolation, testing and repair procedures for high-voltage systems come from the manufacturer and are carried out by trained staff. Use ChatGPT for vocabulary and explanations around that work, never for the procedure itself.
Can ChatGPT read a photo of the scan-tool screen?
Usually, but check every character. OpenAI's image page says images are resized before analysis and that descriptions can be wrong, and a photo of a small screen is exactly where a 0 becomes an 8 or a freeze-frame value lands on the wrong line. Retype the codes and readings, or tick each one against the screen before the answer goes on the job card.
Further reads
- Best AI for Auto Mechanics: Estimates, Parts Research and Messages — Estimates, parts research and messages, sorted by risk level.
- Auto Repair Shop SOPs: AI-Drafted Intake, Handover and Quality Steps — Turn the shop rule into written intake and handover steps.
- Should a Small Auto Repair Shop Use AI? A Decision Checklist — A scored checklist before paying for any AI tool in the garage.
- AI Service Reminders That Bring Garage Customers Back — Bring customers back with reminders that don't overpromise.
- How to Anonymise Client Data Before You Paste It Into AI — A fuller method for stripping identifiers before you paste.
- AI Receptionist for Mechanics: What It Costs Against Missed Calls — What a phone agent costs against the calls you miss.
- Are AI Glasses Worth It for Mechanics? A Small-Garage Test — A one-week test plan for AI glasses in a small garage: what hands-free photos and voice notes save, what they cost, and where they must not be trusted.
- Body Shop Quality Control Checklist: Post-Repair Checks With AI — Filled-in post-repair checks for panels, paint, electrics and photos, with sign-off rules and prompts that turn your own comebacks into a checklist.
- Mobile Mechanic Software for One Van: Jobber vs ServiceM8 With AI — One van, about 60 jobs a month: what Jobber and ServiceM8 charge, how each handles vehicles and parts, and what their AI does for a mechanic.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help articles (Does ChatGPT tell the truth?, ChatGPT Image Inputs FAQ, File uploads FAQ, Data controls FAQ), OpenAI Terms of Use, OpenAI ChatGPT Business pricing page, and published OBD-II fault-code format and code-definition references.