AI can draft the paperwork wording around a repair job: photo captions, supplement requests, insurer replies, repair summaries and daily job notes. A technician or estimator supplies the facts, the AI structures them, and a person signs off every estimate, price and repair decision. It does not write the estimate, set labour hours or judge whether a repair is safe.
The biggest saving is often the supplement, the request to the insurer to add operations or parts that were not visible at the first estimate and turned up after teardown. It is paperwork-heavy (photos, reasons, references), it holds a car and a bay while it waits, and it starts life as a technician's rough notes. The catch is that an assistant left to itself fills gaps with confident detail: a procedure reference, a part number, a labour time. So every prompt here keeps those as blanks for the estimator, and the AI only arranges what the technician actually said.
Where the paperwork sits on one collision job
Names and order differ between shops, insurers and estimating systems, so read this as a pattern rather than a standard. The useful part of the table is the last two columns: for every document there is a part the AI can draft and a part a person has to own.
| Stage | The paperwork | What the AI drafts | What a person owns |
|---|---|---|---|
| Drive-in | Photos, vehicle details, the customer's account of the impact | Numbered photo captions from voice notes; a check of the shot list for gaps | Taking the photos and confirming the vehicle details |
| Write-up | The estimate and the authorisation to repair | Nothing on figures | The estimator writes and signs the estimate in the estimating system |
| Teardown | Hidden-damage findings and photos | Captions from the technician's dictation | Finding and photographing the damage |
| Supplement | The request to add operations, parts or sublet work | The narrative, and a cross-check of notes, photos and lines | Lines, prices, procedure references and approval of the wording |
| Repair | Daily notes, parts status, customer updates | Tidied notes and draft status texts | Recording what was actually done |
| Completion | Repair summary, quality check, invoice, insurer close-out | The customer summary; a list of authorisations against the invoice | The quality-check sign-off and the invoice |
Three signatures are never delegated: the estimate, the supplement lines and the completion quality check. Everything else on the page is wording, and wording is where AI is genuinely quick. Insurer portals vary, so the practical pattern is for the AI to prepare the paragraphs in your standard blocks and for a person to paste them into the portal, rather than letting any tool log in on your behalf.
Photo captions: write from the technician's words, not the pixels
Photos are the evidence behind every supplement, and an unlabelled folder of 30 images is hard evidence to use. Practitioner guidance on supplements keeps returning to one point: photograph hidden damage straight after teardown, with a wide shot and a close-up, because a supplement is judged on its evidence and an unclear photo invites a query or a re-inspection. Captions turn the folder into a file someone else can read in two minutes.
The method is to caption from the technician's own words. While walking round the car, or at the end of teardown, the technician dictates a line per photo into a phone, saying the photo number, the area and what the camera shows. The assistant tidies that into one caption per photo in a fixed format. Here is an illustrative example, first the dictation and then the result.
Photo 14, rear bumper cover off, wide shot from the back. Photo 15, close up right end of the reinforcement bar, bent about four centimetres, tape measure in shot. Photo 16 is the left end, no damage. Photo 17, tail panel crease, whole width, with the straight edge across it. Photo 18, close up of the paint cracked along the crease. Photo 19, right parking sensor bracket, snapped at the clip. Photo 20, left radar module bracket pushed in a bit. Photo 21, odometer.14 | Rear, bumper cover removed | Wide view of rear structure after teardown | Mon
15 | Rear reinforcement bar, right end | Bent about 4 cm out of line; tape measure in shot | Mon
16 | Rear reinforcement bar, left end | No damage (technician) | Mon
17 | Tail panel | Crease across full width; straight edge across panel | Mon
18 | Tail panel, close-up | Paint cracked along the crease | Mon
19 | Right parking-sensor bracket | Snapped at the clip | Mon
20 | Left radar module bracket | Pushed in slightly (technician's description) | Mon
21 | Odometer | Reading to be added by technician | MonTwo details in that output are deliberate. Photo 21 says "reading to be added", because the rule is that the AI never reads a figure off a picture and never fills in a number the technician did not say. And photo 16 keeps the technician's own "no damage" rather than a more confident "undamaged". Captions are evidence, so they should be exactly as strong as the observation.
The second use is a shot-list check. Keep a standard list (four corners, VIN plate, odometer, each damaged panel with a wide shot and a close-up, and every hidden-damage finding with a measurement where there is one) and ask the assistant which items have no caption. On the job above it would come back with "no VIN plate photo, no front three-quarter views, and the radar bracket has only one view". That is a useful second pair of eyes, and I would stop there. Assistants can look at images, but asking one to find damage or judge its severity from a photo invites confident mistakes, and the finding has to come from a technician who has had the panel off. The same principle runs through the roof damage reports tutorial: document, don't adjust. For the intake photos themselves, the shop SOP tutorial has a photo step you can copy.
One rear-end job, start to finish, with the sums
Take an illustrative job: a four-year-old hatchback hit from behind at low speed. The estimator writes the first estimate in the estimating system at $3,150, made up of parts $1,420, labour $1,380 (23 hours at $60) and paint and materials $350. The customer authorises it and the car goes to teardown on Monday.
By the afternoon the technician has the findings and the photos 14 to 21 shown above. The estimator's supplement adds four items, priced from the estimating system and the supplier, never by the AI.
| Line | Operation | Part | Labour | Other |
|---|---|---|---|---|
| 1 | Replace rear reinforcement bar | $215 | 1.0 h | |
| 2 | Pull and refinish tail panel crease | 2.5 h | Paint and materials $40 | |
| 3 | Replace right parking-sensor bracket | $38 | 0.5 h | |
| 4 | Radar calibration after repair (sublet) | Sublet $180 |
Parts come to $253, four labour hours at $60 is $240, paint and materials add $40 and the sublet calibration is $180, which makes $713 and takes the job from $3,150 to $3,863. Writing that request the usual way takes this estimator about 35 minutes: rereading the notes, matching photo numbers, finding the words. With a structured draft and the cross-check described below it takes about 15. If the shop completes 40 jobs a month and 22 of them need a supplement, 22 times 20 minutes saved is 440 minutes, a little over seven hours a month back to the estimator. The bigger gain is that the request leaves on Monday afternoon instead of Tuesday, and every day a supplement waits is a day the car holds a bay and a rental runs.
The supplement template and the missing-item check
A supplement narrative has six jobs: say what was found, where it is, why it was not visible at write-up, what operation it needs, what supports that (a manufacturer procedure, where one applies) and what it does to the schedule. The AI can arrange all of this from the technician's notes. It must not supply the support: a repair procedure or labour time quoted from memory reads just as convincingly when it is wrong, so the prompt keeps references, part numbers and labour times as placeholders for a person.
You prepare supplement requests for a collision repair shop. Use only the technician's notes and the photo captions below.
- One numbered finding per damaged item: where it is, what was found, what operation is needed, photo numbers.
- Say why an item was not visible at write-up only if the notes say so; otherwise write [estimator to confirm].
- Never add an operation, a part number, a labour time or a manufacturer procedure. Where a reference is needed write [procedure reference: estimator to add].
- Use the technician's words for severity. Do not make damage sound better or worse than the notes.
Output: a one-line reason, the numbered findings, the effect on the schedule, and the photo list.
TECHNICIAN NOTES: [paste]
PHOTO CAPTIONS: [paste]SUPPLEMENT REQUEST 1 | Job [ref] | four-year-old hatchback | teardown Monday
Reason: additional rear damage found at teardown, not visible at write-up with the bumper cover fitted.
Findings
1. Rear reinforcement bar: bent about 4 cm out of line at the right end; no damage found at the left end (photos 15, 16). Operation: replace bar. [procedure reference: estimator to add]
2. Tail panel: crease across the full width with cracked paint (photos 17, 18). Operation: pull and refinish.
3. Right parking-sensor bracket: snapped at the clip (photo 19). Operation: replace bracket.
4. Left radar module bracket: pushed in slightly (photo 20). Operation: calibration after repair. [procedure reference: estimator to add]
Schedule: bar and bracket not in stock; supplier ETA Thursday. [days added: estimator to confirm]
Photos attached: 14 to 20.The sample shows the placeholders doing their job. The estimator fills in the procedure references and the days added, and nothing in the text claims more than the notes did. A first draft from a looser prompt often says something like "tail panel requires replacement" when the technician said pull and refinish, which is a labour line rather than a parts line. A narrative that disagrees with its own estimate lines gets queried, and that is why the second step matters.
The second step is the missing-item check. Paste the technician's notes, the photo captions and the priced supplement lines, and ask three questions: which items in the notes have no line, which lines have no matching note or photo, and where does the wording disagree with a line? On this job it would flag that the notes mention a calibration while the draft lines had none, which is how line 4 got added. It does not add or price anything; it lists, and the estimator decides. If a draft ever makes the damage sound worse than the photos show, treat that as a risk, not help. A supplement works because it is accurate, and an insurer's view of your shop builds up over many of them.
One more habit: watch the running total. Insurers also compare a job's cost with the vehicle's value, so a supplement that pushes the figure towards the point where they may write the car off deserves a phone call as well as an email. The AI cannot know that threshold, so keep it on the estimator's checklist.
When the insurer queries a line: a reply built from your own notes
Not every supplement is approved as sent. A typical query asks why a part is being replaced rather than repaired, or why a refinish carries the hours it does. The reply is the same material again: the technician's findings, the photos and the procedure reference. Paste the query, the relevant notes and the photo captions, and ask for a reply that answers only what was asked, cites the evidence by photo number, and ends by saying what you need from the insurer, whether that is approval or an inspection date.
INSURER QUERY: "Line 1: please confirm why the reinforcement bar needs replacing rather than straightening."
DRAFT REPLY:
Thanks for your query on line 1. The right end of the rear reinforcement bar is bent about 4 cm out of line (photo 15, tape measure in shot); the technician found no damage at the left end (photo 16). The technician's assessment is that the bar should be replaced rather than straightened. [procedure reference: estimator to add] The part is on order for Thursday, so a decision today keeps the job on schedule. Please confirm approval, or tell us when your inspector can attend.Four checks before that goes out. The assessment is the technician's, not the AI's, so the estimator should be able to stand behind it. The sentence about Thursday is true only if the supplier has confirmed it. The procedure reference still waits for a person to supply it. And the tone offers an inspection instead of an argument, which is the right default. If the insurer disputes the finding itself, or the price, a person takes over the conversation. The AI should not argue procedure or money on the shop's behalf.
Daily job notes: dictate, tidy, keep the log honest
A job record is only useful if it is written on the day, and technicians are better at saying things than typing them. The routine is a two-minute voice note per car at the end of the shift, tidied by the assistant into four fixed headings: done, waiting on, customer contact and flags for tomorrow. The rule is that the AI may reorder and reword what was said, but it may not add an operation, and it may not harden a hedge.
| Dictated at the end of Wednesday (raw) | Tidied entry for the job record |
|---|---|
| Bar and bracket not here yet, supplier said tomorrow now. Pulled the tail panel, filler on it, ready for primer tomorrow morning. Customer rang about the courtesy car, told her Friday maybe. Don't forget the sensor wiring clip on the left, it's loose, check when the bumper's back on. | Wed. Done: tail panel pulled and filled, ready for primer. Waiting on: reinforcement bar and sensor bracket (supplier now says Thursday). Customer contact: asked about courtesy car; told "Friday, maybe". Flag: left sensor wiring clip loose; check when bumper is refitted. |
The second rule comes from a realistic slip: a draft that turns "told her Friday, maybe" into "completion expected Friday", which is one short step from a customer text. Keep hedges in the log. A related slip is claiming work that is not done. Picture a completion summary that says "rear radar calibrated and verified" because the supplement contained a calibration line, while the sublet was booked for the next day. The quality check should catch that, but it is better if the draft never says it. The body shop quality control checklist is the place to make "every line on the job has a record of being done" a standing check, and the repair updates tutorial shows how to turn the same notes into customer texts without over-promising.
The drafts are only as good as the dictation, so agree a few habits with the team and pin them next to the workstation.
- Say the photo number first, then the area, then what the camera shows.
- Name each panel the same way every time, from a short list the whole shop shares.
- Give a measurement whenever you have one, and say what you measured against.
- Say "not done" or "not checked" out loud. An omission in a note becomes an assumption in a draft.
- Keep the hedges in: "supplier says Thursday" is not "arriving Thursday".
- Leave out names, phone numbers and claim numbers, because the notes get pasted into a prompt.
Repair summaries and the insurer close-out
At the end of the job two documents get written: the summary the customer reads and the closing paperwork the insurer needs. The customer summary should say, in plain English, what was damaged, what was repaired or replaced, which parts were fitted as recorded on the job, and what to do and watch for afterwards. The aftercare and warranty paragraphs come from your own written terms and your paint supplier's guidance. Put them in the prompt as fixed text, because an assistant asked to write aftercare advice from memory will invent a cure time.
For the insurer close-out, the useful AI job is reconciliation: list every authorisation with its date and amount, and set the list against the invoice. In the illustrative job the authorisations are $3,150 on Monday and the $713 supplement on Tuesday, a total of $3,863. If the invoice says $3,901, the difference is $38, which happens to be the price of the sensor bracket: the bracket was added to the invoice by hand when it arrived and was also on the supplement. The assistant spots the match, the estimator checks the invoice and corrects it, and the insurer never has to query it. The assistant only compares the figures it has been given, so a person still confirms the totals in the system before anything is sent.
Redacting claim details before they reach an assistant
Claim paperwork is full of personal data: the policyholder's name, address and phone number, policy and claim numbers, sometimes injuries, and always the vehicle. Most of what the AI needs is none of that. It needs the damage description and the job. A simple redaction habit before anything goes into a prompt does most of the work.
| Leave out of the prompt | Use instead |
|---|---|
| Policyholder's name, address, phone, email | "The customer", or nothing |
| Policy and claim numbers | Your own job reference |
| Full VIN and number plate | Year, make and model, plus the job reference |
| The insurer contact's name and details | "The insurer contact" |
| Injury or medical details from the claim | Nothing; they do not belong in repair notes |
| Photos showing faces or plates | Crop them, or caption without uploading |
Plan choice matters too. ChatGPT Business and Claude Team both need at least two seats. A standard seat on either lists at $25 a month, or $20 a month billed annually, and neither trains on business content by default. Individual ChatGPT and Claude plans list at about $20 a month and have a model-training switch in the privacy settings that you should check rather than assume. Setting names change, so look at the current screen. Then read your agreements with insurers and any repair network: some may set rules on where claim data can go, and it is better to know before pasting than after. Data-protection rules can apply to this kind of information, so a line about AI tools in your privacy notice is worth adding; a data-protection adviser can give you the wording.
Five numbers that show the paperwork is faster and cleaner
Pick five numbers and write them down before you start, so the comparison is honest. These are the ones I would use, with illustrative before and after values for a shop doing 40 jobs a month.
| Measure | How to count it | Before | After |
|---|---|---|---|
| Hours from teardown to supplement sent | Teardown photo time against the sent message | 26 | 6 |
| Supplements returned for more information | Count per month | 7 of 22 | 2 of 22 |
| Estimator minutes per supplement | Stopwatch for one week | 35 | 15 |
| Photos captioned when the request is sent | Share of photos on each request | 40% | 100% |
| Notes with an operation missing from the job record | Spot-check ten notes a week | Not measured | 0 |
If the numbers do not move after a month, the usual cause is the dictation rather than the AI. Short, specific notes with photo numbers make good drafts, and mumbled summaries make guesses. The last row is the one to watch: it should be zero, and a single hit means the prompt or the habit needs tightening before anything else gets rolled out.
Questions body shop owners ask about AI paperwork
Can AI write the estimate itself?
No, and I would not let it. An estimate carries money and safety consequences, so the figures, labour times and parts prices must come from your estimating system and a trained estimator who signs it. AI is useful for the words around the estimate: captions, summaries, supplement narratives and job notes. Treat anything that looks like a price or a procedure in an AI draft as a placeholder until a person confirms it.
Will an insurer accept a supplement written with AI?
Insurers judge the evidence: the photos, the operations requested and the references that support them. A supplement stands or falls on those, not on who typed the narrative. Agreements and portals differ, so ask your insurer contact whether they have rules about AI-assisted paperwork, or about where claim data may be processed, before you build a routine around it.
Can AI spot hidden damage from the photos?
I would not rely on it. Assistants can describe what is visible and check a shot list for gaps, but judging severity or finding concealed damage needs a technician with the panel off, and a confident wrong answer is worse than none. Use the AI to check that the photos exist and are captioned, and leave the diagnosis to the people who repair the car.
Do I need new software to start?
No. Copy the technician's dictation into an assistant you already have, check the draft, and paste it into your estimating system or the insurer's portal. Your estimating-system vendor may offer AI features of its own, and they differ by product and plan, so ask what exists, what it costs and where the data goes before you buy anything.
Further reads
- Auto Repair Intake Form: Let AI Capture the Complaint Clearly — Capture the customer's account of the impact clearly at drive-in.
- Auto Repair Price Estimate Template: Build One With ChatGPT — Where the sums live: a sheet that checks the figures twice.
- Best AI for Auto Mechanics: Estimates, Parts Research and Messages — Which assistant suits garage wording, and what stays manual.
- Following Up Unapproved Repair Estimates With AI — Chase unapproved estimates honestly, with partial approval allowed.
- AI Receptionist for Mechanics: What It Costs Against Missed Calls — What an AI receptionist costs against missed calls.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI's ChatGPT Business pricing page and Claude's pricing page for seat prices, seat minimums and data-training defaults; practitioner and trade-press guidance on supplement and teardown documentation.