For shop use, no single AI is best at diagnosis, because none can measure the car. ChatGPT, Claude and Gemini are the stronger choices for explaining codes and reading a manual you upload; Perplexity is useful when you want sources to check. Score them on your own three codes before paying.
Most results for this keyword are consumer apps that tell a car owner what a warning light might mean. A garage needs something different: help ordering tests, reading a long service manual and explaining findings, without trusting a guess that nobody verified. This tutorial gives you a scorecard, a 40-minute test and the limits to write down before anyone uses it on a customer's car.
What a chatbot can and cannot do on a customer's car
A fault code is a starting point, not a diagnosis. The code says which circuit or monitor failed a check, not which part is at fault. A chatbot knows the general meaning of generic codes and the usual causes, which is useful. It cannot see freeze-frame data you did not paste, cannot test a sensor, and does not know that your van has a known intake leak on that engine unless you say so.
That split drives the whole scorecard. Good uses are the ones where a wrong answer costs a few minutes and gets caught: a test order, a plain-English explanation, a summary of a long procedure. Bad uses are the ones where a wrong answer goes straight onto a car: torque figures, airbag or high-voltage procedures, wiring colours and brake system specifications. The table sorts the common jobs.
| Job | Risk if the AI is wrong | Use AI? |
|---|---|---|
| Explain what a generic code means in plain words | Low: you know the code | Yes |
| Suggest a test order from codes and symptoms | Medium: costs time | Yes, as a checklist you review |
| Find the right section of an uploaded manual | Low if you read the page | Yes, ask for the page number |
| Quote a torque value, capacity or clearance | High: parts and safety | Never from memory; read it from the manual page |
| Airbag, high-voltage battery or brake procedures | Severe | No. Manufacturer procedure and training only |
| Write the customer explanation of your finding | Medium: reputation | Yes, after you check it |
The four assistants and what each brings to a shop
Prices are USD list prices from the vendors; confirm on each page, as plans change. All four have a free tier suitable for the test below.
| Assistant | Paid plan for one person | Shop strength | Watch-out |
|---|---|---|---|
| ChatGPT | Plus, $20 a month | Good at structured checklists and explaining in plain words; accepts large file uploads | Confident tone even when unsure; ask it to say what it cannot know |
| Claude | Pro, $20 a month ($17 billed annually) | Careful with long manuals; reads PDFs up to 1,000 pages | Visual layout (diagrams, tables as images) is analysed only up to 100 pages |
| Gemini | Google AI Pro, $19.99 a month | Convenient if the shop already runs on Google's mail, docs and drive | Check which model your plan uses, since app models change often |
| Perplexity | Pro, $20 a month | Shows linked sources for each claim, so you can open the forum thread or bulletin it used | Sources include forums and blogs; a link is not the same as a reliable procedure |
The honest ranking for a small garage: use one of ChatGPT, Claude or Gemini as the main assistant, and keep Perplexity's free tier for "what is the known issue with this engine" questions where seeing the sources matters. Paying for all four is not sensible. I am not publishing a league table of which one got a particular code right, because answers vary by vehicle, by model version and from day to day. What follows is a test you can run yourself in your own shop, which tells you more.
A 40-minute test on the same three fault codes
Choose three codes you have seen recently and already solved, so you know the true answer. A useful mix is one common, one ambiguous, one that needs context. For illustration, here are three generic codes with their standard titles:
- P0171, system too lean (bank 1). Common, many possible causes.
- P0420, catalyst system efficiency below threshold (bank 1). Often blamed on the catalytic converter when something upstream is the cause.
- P0301, cylinder 1 misfire detected. Needs the engine type and mileage to say anything useful.
Give each assistant the same prompt, with a vehicle description and symptoms in the same words. Here is a template.
You are assisting an experienced mechanic. Be concise and cautious.
Vehicle: [year, make, model, engine, mileage].
Code: P0171. Freeze frame: coolant 88 C, short-term fuel trim +18%, long-term +14%, idle.
Symptoms: rough idle when cold, no warning lights other than the engine light.
1. List the three most likely causes in order, with a one-line reason each.
2. Give a test order for a shop with a scan tool, smoke machine and multimeter.
3. State what you cannot know from this information.
4. Do not give any torque values, capacities or wiring colours.Score each reply on five criteria, with 0 to 2 points each, so the maximum is 10 per code. Do the scoring blind if someone else can paste the replies without names.
| Criterion | 0 points | 1 point | 2 points |
|---|---|---|---|
| Cause list | Misses the real cause | Real cause present but low down | Real cause in the top three |
| Test order | Starts with replacing parts | Mixed tests and part swaps | Tests before parts, cheapest first |
| Honesty about limits | States things it cannot know | Generic disclaimer only | Says what data is missing |
| No invented figures | Quotes a spec anyway | Quotes with a caveat | Declines and points to the manual |
| Plain-English summary for the customer | Jargon heavy | Usable with edits | Ready after a quick read |
An accuracy table from your own run might look like the layout below. The figures here are placeholders to show the shape, not results.
| Code | ChatGPT | Claude | Gemini | Perplexity |
|---|---|---|---|---|
| P0171 | [score /10] | [score /10] | [score /10] | [score /10] |
| P0420 | [score /10] | [score /10] | [score /10] | [score /10] |
| P0301 | [score /10] | [score /10] | [score /10] | [score /10] |
| Total /30 |
Keep the sheet. When a vendor releases a new model, run the same three codes again and compare. After a few rounds you will have your own, shop-specific accuracy record, which is far more useful than anyone's generic ranking.
What a good answer and a weak answer look like on P0420
This is the main worked example, because P0420 is where lazy diagnosis is expensive: a catalytic converter is a costly part, and the code does not prove it has failed. Say a four-bay garage has a ten-year-old hatchback with the code, no other codes, and an owner who has been told by a friend to "just fit a new cat".
A weak reply (illustrative) reads: "P0420 means your catalytic converter is failing. Replace the catalytic converter. This typically costs between $X and $Y." It jumps to the part, states a price it cannot know, and offers no tests.
A good reply (illustrative) reads: "The code means the downstream oxygen sensor signal suggests the converter is storing less oxygen than expected. The converter is one cause, but check first for exhaust leaks before the downstream sensor, a lazy or contaminated downstream sensor, and misfires or rich running that damaged it. Suggested order: check for other stored codes and fuel trims; inspect the exhaust for leaks; compare upstream and downstream sensor waveforms on a scope or scan tool; only then consider the converter. I cannot tell from this information whether the sensors have been replaced recently."
The second reply scores 8 or 9 out of 10 on the rubric: real cause in the list, tests before parts, and an honest statement of what is missing. You would still read it critically. Your own experience might add "on this model, check for the known exhaust manifold crack first". That local knowledge is the point: the assistant gives you a clean checklist, and you decide what the car needs.
Feeding it your service manual without trusting it blindly
The biggest practical gain is not the code explanation, it is search through documents. Upload the relevant manual PDF, or a section of it, and ask for the procedure with the page number. Then open the page and read the figure yourself. A useful prompt is below.
I have uploaded the service manual section for the cooling system.
Find the procedure for bleeding the system after a thermostat replacement.
Give me: the step numbers in order, the page number of each step, and any
warning boxes verbatim. If the manual does not cover it, say so.
Do not add steps from general knowledge.Three limits to respect. First, file size and page limits depend on the plan and the product; Claude, for example, reads PDFs up to 1,000 pages but analyses visual layout only up to 100, so a wiring diagram on page 400 may not be read as a picture. Second, scanned manuals may extract poorly, so spot-check any figure against the page. Third, check the terms of your information subscription before uploading its content, because some licences restrict how material can be copied or shared.
Here is a before-and-after of the time saved (illustrative). A technician searching a 600-page manual for a specific bleeding procedure might spend ten minutes finding it; with a page-numbered answer and a check against the page it is two to three. Across fifteen such lookups a week that is over an hour back, which the subscription cost easily covers.
P0171 and P0301: what the follow-up question should look like
A first reply is rarely the useful one. The value comes from the second round, where you paste what the tests showed and ask for the next step. Take the lean code. Say the first reply suggested checking for an intake leak, a weak fuel pump or a dirty airflow sensor. You run a smoke test and find nothing, then read the trims: long-term fuel trim is +14% at idle but drops to +3% at 2,500 rpm. Paste that back:
Smoke test found no intake leak. Trims: LTFT +14% at idle, +3% at 2,500 rpm.
MAF reading at idle looks plausible. What does the idle-only correction point to,
and what would you test next? Keep it to three steps.A trims pattern that fixes itself as airflow rises is the classic sign of a small unmetered air leak, because the leak matters most when the engine pulls little air. A good assistant should say so and suggest, for example, checking the brake booster hose, the crankcase ventilation hose and any vacuum lines the smoke test could not reach. If it instead tells you to replace the airflow sensor, that is a sign it is guessing, and a good reason to score that reply down.
Now the misfire. P0301 with no context gives you a generic list: spark plug, ignition coil, injector, compression. The assistant becomes useful once you add what you did: "swapped the coil to cylinder 3 and the misfire followed to P0303". That one line of evidence points at the coil, and a strong reply says so rather than repeating the generic list. The scoring question is whether the assistant uses the evidence you gave it. Some replies ignore it and restate the list; mark those down.
Here is how the three codes might look as a short sheet to pin up (illustrative):
| Code | Context the assistant needs | Evidence that changes the answer |
|---|---|---|
| P0171 | Engine, mileage, fuel trims at idle and at 2,500 rpm | Trims that normalise with revs suggest an air leak |
| P0420 | Other codes, upstream and downstream sensor behaviour | A lazy downstream sensor or an exhaust leak changes the diagnosis |
| P0301 | Engine type, mileage, plug age, swap test results | A misfire that follows the coil points at the coil |
Turning a finding into the sentence the customer reads
When the diagnosis is done, the same assistant can write the explanation, which is where many garages lose time. Paste your technical note and ask for two or three short sentences without jargon. A before and after (illustrative):
Before (technician note): "P0301. Coil swap test, misfire followed to cyl 3. Coil 1 OK on bench. Replace coil on cyl 1 and plug, clear, road test."
After (draft for the customer): "Your engine was misfiring on one cylinder. We moved the ignition part between cylinders to test it, and the fault followed the part, so that is what needs replacing. We will also fit a new spark plug, clear the warning and test drive the car."Notice what you check in the draft: that it does not say "diagnosed by AI", that it keeps to the work you actually carried out, and that it does not promise a result the test drive has not confirmed. The same tool can draft the estimate wording, but you set the prices.
When the code is not the problem
Some jobs never start with a code. A customer says "it makes a noise when I turn left and the car is slow to start". Here the assistant helps with questions to ask: when does the noise happen, at what speed, hot or cold, does it change with steering input? Ask it for a symptom-gathering list, not a diagnosis, and print it for the front desk. A chatbot that interrogates the customer well saves a technician five minutes of guessing and sets a better first inspection.
One more edge case: electric and hybrid vehicles. Anything involving the high-voltage system needs manufacturer procedure and trained staff. Use AI for the explanation of how the system works in general terms, never to say that a component is safe to touch.
Limits to write on the workshop wall
A short written policy protects you and the customer. Keep it to a page that every technician has read.
- AI never decides a repair. It suggests tests; the technician decides after testing.
- No safety-critical figures from AI. Torque values, brake and steering specifications, airbag and high-voltage procedures come only from the manufacturer or your professional information system.
- Verify every number on the page. If the AI cites a manual, open that page.
- No customer personal data in prompts. Vehicle details and codes are fine; names, phone numbers and registration plates are not needed. Use a business plan or switch off model training on personal plans.
- Log AI use on the job card. A note such as "checklist from AI, tests confirmed by technician" keeps the record honest.
A realistic mistake shows why rule two exists. A technician asked a chatbot for the torque on a wheel hub nut, got a confident figure, and used it without checking. The figure was for a different variant of the same model. Nothing failed that day, but a loose or over-torqued hub can fail later. The fix is the rule: figures only from the manual page for that exact vehicle.
Using the same tools on the parts of the job that are not diagnosis
A shop gets more safe value from the writing side than from the diagnosis side. Customer explanations, repair status messages and service reminders are low-risk and repeatable. Ask the assistant to turn "rear pads at 2 mm, discs scored, fluid dark" into three sentences a customer will understand, then edit. How garages use AI to send repair updates customers appreciate covers the wording, and AI service reminders that bring garage customers back covers the follow-up.
If your question is broader than diagnosis, the roundup of AI for auto mechanics: estimates, parts research and messages maps the other jobs. For the deeper look at where ChatGPT fault-finding stops being reliable, read ChatGPT for car mechanics: fault-finding help and where it stops.
Choosing after the test
When the scores are in, apply a simple rule. If one assistant leads by three or more points over the three codes, take it. If they are within two points, choose on practical grounds: the one that handles your manual PDFs better, the one that fits your existing accounts, or the cheapest paid plan. Perplexity earns a place only if its sources were actually useful to you, not just present.
Start on the free plan of your top pick for a month and keep the score sheet. Upgrade only if you hit the limits within the week, which tends to happen once a few technicians share the habit. If more than one technician will use it, the business plans from the main vendors need at least two seats; check each vendor's current pricing page before you buy, and set the written limits above before anyone relies on it.
Questions mechanics ask about AI diagnosis
Can AI replace a scan tool or a diagnostic subscription?
No. A chatbot cannot read the vehicle, run actuator tests or pull live data, and it has no access to your manufacturer procedures unless you upload them. Treat it as a second pair of eyes that explains codes, suggests a test order and reads a long manual for you, with the scan tool still doing the measuring.
Which AI is best at reading a service manual PDF?
Any of the big assistants can answer from an uploaded PDF, but check the page limits on your plan, and confirm every number it quotes against the page. Claude reads PDFs up to 1,000 pages, with visual layout analysed only for the first 100 pages. Ask for the page number each time.
Is it safe to trust AI torque specs and wiring colours?
No. Torque values, fluid capacities and wire colours must come from the manufacturer or your professional information system for that exact vehicle. AI models can produce plausible-looking numbers that are wrong. Use the assistant to find the right section of a manual you supplied, then read the figure from the page itself.
Should a customer see the AI's diagnosis?
Show the customer your findings and the tests you ran, not a chatbot answer. A diagnosis is your professional judgement, and an AI suggestion that was never tested is a guess. Use the assistant to help write the plain-English explanation of what you found, then check every claim before it goes on the estimate.
Further reads
- AI Receptionist for Mechanics: What It Costs Against Missed Calls — What an AI receptionist costs against missed calls.
- Are AI Glasses Worth It for Mechanics? A Small-Garage Test — A small-garage test of AI glasses.
- Mobile Mechanic Software for One Van: Jobber vs ServiceM8 With AI — Software for a one-van mobile mechanic, with AI.
- Quoting Custom Jobs Faster With AI in a Small Workshop — Quoting jobs faster with AI in a small workshop.
- Auto Repair Intake Form: Let AI Capture the Complaint Clearly — A 14-field digital intake form, symptom follow-ups by complaint type, an AI summary prompt that never diagnoses, and a read-back the customer confirms.
- Auto Repair Shop SOPs: AI-Drafted Intake, Handover and Quality Steps — Three complete shop SOPs for intake, quality check and handover, drafted with AI from your own notes, with photo steps and a way to train new technicians.
- 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.
- AI for HVAC Contractors: Getting Answers From Manuals and Spec Sheets — How an HVAC contractor can search manuals and spec sheets with AI: what to load, a prompt that forces page numbers, a ten-question test and safety limits.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI, Anthropic, Google and Perplexity pricing pages and help centres, checked October 2026; Anthropic documentation on PDF support. Diagnostic trouble code titles are the standard generic OBD-II definitions.