Give the AI your numbered photo checklist and a batch of job photos. It matches each photo to a checklist line, flags blurred or unreadable shots, and lists the lines with no photo. The installer then retakes what is missing on site and signs the commissioning tests, which stay manual.
The AI judges whether a photo exists and can be read, not whether the installation is right. It cannot check a torque setting, a polarity or an earth connection from a picture, and vision models misread small print. Used as a photo audit before the crew leaves, it catches missing evidence while a retake still costs ten minutes instead of a return visit.
The 32-shot checklist, grouped by where the camera points
This is the list for a typical residential rooftop system with a string inverter and no battery, written as a general illustration. Your installation standard, your certification body and your grid operator add to it or change it, so treat the IDs as a template to edit, not a rulebook. Each line says what the photo must show, because "photo of the inverter" is not a standard anyone can check.
Group A: roof and mounting (8 shots)
Almost all of this is hidden once the modules go on, so the photos taken before and during installation are the only evidence you will ever have.
| ID | Shot | The photo must show |
|---|---|---|
| A1 | Roof before work | Wide view of the roof surface, its condition and the access arrangements |
| A2 | Roof attachments | Each row of fixings with flashing in place, before the rails go on |
| A3 | Rails | Layout and spacing before modules are fitted, with a tape or marker for scale |
| A4 | Rail joints | Close-up of each splice with its fixings in place |
| A5 | Module clamps | Close-up of mid and end clamps, fully seated, at the end of each row |
| A6 | Bonding | The earth or bonding connection on the rails and on a module frame, where the design uses them |
| A7 | Array edge | Wide shot showing the distance from the array to the roof edge and ridge |
| A8 | Finished array | Overview from the ground with every module visible |
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowGroup B: cabling on the roof and into the building (6 shots)
Cable faults are slow to appear and awkward to trace, so the evidence here is about routing, protection and labels.
| ID | Shot | The photo must show |
|---|---|---|
| B1 | Cable routing | Wide shot with cables clipped up off the roof surface |
| B2 | Clips and ties | Close-up of UV-rated fixings at a representative spacing |
| B3 | String labels | One close-up per string at the array end, with the string number and polarity readable |
| B4 | Connectors | A mated pair, locked together and not under strain |
| B5 | Roof entry | The cable penetration sealed and flashed |
| B6 | Inside entry | Cables entering the building through conduit or a gland, secured |
Group C: DC side and inverter (7 shots)
| ID | Shot | The photo must show |
|---|---|---|
| C1 | DC isolator | Cover closed, labelled, with clear working space around it |
| C2 | Inverter location | Wide shot showing the mounting, the clearances and the ventilation space |
| C3 | Inverter rating plate | Model and serial number readable at full size |
| C4 | DC terminations | Strings landed at the inverter, labelled to match B3 |
| C5 | AC terminations | The AC connection at the inverter or its AC isolator |
| C6 | Inverter display | Status after start-up with no fault showing, job reference in frame |
| C7 | Monitoring portal | Screenshot showing the system online, with the job reference |
Group D: AC side, meter and protection (6 shots)
| ID | Shot | The photo must show |
|---|---|---|
| D1 | AC isolator | Labelled and accessible |
| D2 | Main distribution board | Wide shot (the breaker panel) with the new protective device in place |
| D3 | Protective device and circuit label | Close-up with the label readable |
| D4 | Generation meter | Serial number and starting reading readable, if one is fitted |
| D5 | Earthing | Any main earth or bonding connection touched by the work |
| D6 | Warning labels | Signs fixed at the board, the isolators and the meter position, as required |
Group E: records and handover (5 shots)
| ID | Shot | The photo must show |
|---|---|---|
| E1 | Test record | The completed commissioning sheet, every box filled, job reference at the top |
| E2 | Instrument label | The calibration date on the tester used |
| E3 | Settings screen | Export limit or grid settings, where they apply |
| E4 | Schematic | The single-line diagram as left on site |
| E5 | Handover sheet | The signed customer handover sheet |
If a battery is fitted, add shots for its location and clearances, its rating plate and its isolation, and write the additional lines into the list before the job starts. A survey checklist for the stage before this one, with the photos that decide whether the job can proceed, is in the solar site survey tutorial.
Shooting rules that make the sorting reliable
The AI can only sort what the crew captures, so the rules start on the roof, not in the office. Five habits do most of the work.
- Put the job reference in frame. A small whiteboard or printed card in the corner of the shot ties the picture to the job even if the file name is lost.
- Name each file with the job number, the checklist ID and a subject, such as J2417_B3_string1-label.jpg. The name is a claim, not evidence, but it makes mismatches easy to spot.
- Shoot labels twice. One wide shot shows where it is, and one close-up makes it readable. Images are resized before analysis, so small print on a wide shot becomes mush.
- One subject per photo. A single frame that tries to show the isolator, the meter and the board will be matched to one line and counted for three.
- Retake in the moment. If a label is glared out, move and shoot again before leaving the spot. Do not leave it for the audit to catch later.
Tools that capture the job reference, time and checklist in one step help here. A comparison of photo-report apps for a neighbouring trade is in the roofing photo report tutorial.
The sorting prompt, and what the first run returned
Paste the checklist (IDs and descriptions) into the prompt, then attach photos in batches. This version fixes a weakness found in the first run, described below.
You are checking commissioning photos for one residential rooftop solar job
(job J2417: 16 modules, 2 strings, one string inverter, no battery).
The checklist below has IDs A1 to E5. I will attach photos in batches of 10,
named with the job number, a checklist ID and a subject.
For each photo:
1. Say which checklist ID the image CONTENT shows, whatever the file name says.
If the content and the name disagree, write MISMATCH.
2. Say READABLE, BLURRED, TOO DARK, TOO FAR or GLARE.
3. Describe in one line what is visible. Do not say whether the work complies.
Do not treat any serial number, reading or test value you read as fact.
After I say DONE, give a table of every checklist ID with a status (OK,
PARTIAL, UNREADABLE, MISSING), the supporting file names, and a list of
retakes. Where a line needs one photo per string, say which strings are covered.
[checklist A1 to E5 pasted here]The first version of the prompt lacked the line about file content and returned the summary below. It is an illustrative extract.
ID Status Supporting photos
A1 OK J2417_A1_roof-before.jpg
...
B3 OK J2417_B3_string1-label.jpg, J2417_B3_string2-label.jpg
C3 UNREADABLE J2417_C3_inverter-plate.jpg (serial digits blurred)
D3 MISSING none
...
Summary: 30 OK, 1 UNREADABLE, 1 MISSING. Retakes: C3, D3.Opening the five riskiest shots by hand (plate, string labels, board label, isolator labels, test sheet) turns up the problem. The two B3 photos both show a label reading "String 1". The crew took the string 1 label twice and named the second file "string2" by mistake, and the assistant trusted the file name. B3 is therefore only partly covered, and string 2 has no label photo at all. The C3 and D3 flags were correct. So the fixes are: tell the model to judge by content and to write MISMATCH, as the prompt above now does; require the strings covered to be named, as it also does; and always open the riskiest five yourself.
A worked job: 38 photos in, 3 gaps out
Now the numbers. The job is 16 modules of 400 W each, which is 6.4 kW, wired as two strings of eight into one string inverter, with a generation meter and no battery. The checklist gives 32 lines, 33 photos at minimum because B3 needs one photo per string. The crew uploaded 38: the required shots plus second-distance label photos and a retake.
Sorted by hand in the office, 38 photos take about 25 minutes, opening each one, renaming it and ticking the list. With the prompt, four batches (10, 10, 10 and 8), a read of the table and the five-shot spot check, say 12 minutes. The saving in minutes is modest. What matters is when the audit runs. Done in the van before the crew leaves, each gap is a ten-minute retake.
| Line | After the first run | After the human check | Action |
|---|---|---|---|
| 29 lines (all but three) | OK | OK | None |
| B3 string labels | OK | PARTIAL: string 1 only | Photograph the string 2 label |
| C3 inverter rating plate | UNREADABLE | UNREADABLE | Retake at close range without glare |
| D3 device and circuit label | MISSING | MISSING | Take the close-up |
Three gaps in a 32-line list is easy to produce on a busy day, and a quick scroll through the gallery would not reveal them, because each photo looks fine at a glance. If any of the three had been found a week later, the cost is a return visit: say two people and a van for half a day, which this illustration prices at $300. Three gaps could mean up to three trips, though a sensible planner would batch them into one. The point is not the dollars. It is that the evidence file is complete at handover, which matters if there is a warranty claim, an insurer's question or a dispute later. The same photo-sorting idea for a heating installer is covered in the boiler installation photo checklist.
Where vision models go wrong on solar photos
Know the failure modes before you trust a "complete" result. These are the ones to test for, each with how it shows up.
- Trusting file names. Shown above. The model sees "string2" in the name and agrees with it. Cure: instruct it to judge by content, and open the label photos yourself.
- Confusing similar boxes. A DC isolator and an AC isolator are both grey boxes with a handle. The model may match either to either. Cure: the checklist asks for the label text in frame, and you check that the text matches the line.
- Misreading digits and letters. Serial numbers and string labels are small print, and similar characters get swapped. Cure: use the photo to confirm legibility, then type values from the plate or portal. Never let the assistant fill a record from a photo.
- Counting. Ask "are there clamps on every row end?" and the answer will be confident and unreliable. Cure: a count belongs to a person; the checklist asks for a close-up of one clamp per row end and the human counts rows.
- Glare and reflections. Glass and glossy labels turn into white patches. A good model says GLARE; a poor one describes what it expects to be underneath. Cure: the readability status, plus a person opening any photo marked GLARE.
- Judging compliance. Left unrestricted, it will say a cable clip spacing "looks compliant". It cannot measure from an angle, and it does not know your standard. Cure: the prompt forbids compliance statements and asks only for description.
- Hidden detail. It cannot see a torque mark under a cap, a cable behind a panel or a bond that was never made. The photo shows what the crew pointed at. Cure: that is what the checklist wording is for.
The general limits of assistants reading photos and files, including what breaks, are set out in the tutorial on ChatGPT and photos.
Check that the sorting works before you rely on it
A tool that has not been tested against known gaps is a tool that will say "complete" about an incomplete job. Test it by planting problems.
- Take five finished jobs with complete photo sets.
- From each, remove three photos at random and swap one more for a blurred copy, so each job has four known problems and 20 in all.
- Run the prompt on each job and record which planted problems it found, which it missed and what extra items it flagged.
- Accept it only if it finds every planted problem, or if every one it missed is one you now guard with a manual spot check. A single miss on a missing photo, as opposed to a blurred one, deserves a prompt change.
- Tolerate false alarms. A couple per job is acceptable: flagging a good photo costs a minute, and missing a bad one costs a visit.
Repeat the test after any change to the prompt, the checklist or the AI model you use, since model updates can shift behaviour without notice. Keep the five riskiest shots on a manual spot check for every job, whatever the test says.
Upload limits and customer privacy
Photos are bulkier than text, and each assistant caps them differently. As of 1 November 2026, from the vendors' own documentation: Claude on its apps takes 20 images per message at 10 MB each; Gemini's apps take 10 files per prompt; and ChatGPT accepts images up to 20 MB each, with paid plans allowing up to 80 files every three hours and the free plan 3 file uploads a day. Batches of ten fit within every one of these, which is why the prompt above works in tens. A 38-photo job is four batches.
Check the privacy side before the first upload. Photos of a house can show the address, vehicles, people and sometimes documents on a table. Business plans of ChatGPT, Claude and Gemini do not train on your content by default, while consumer plans carry a data-controls setting that you can use to stop chats being used for training. Check the one you are on, crop out people and number plates where you can, and put photo consent into your customer terms.
Missing-photo alerts: three ways to wire it up
- By hand. Copy the retake list from the summary into the job note and message the crew lead. This is enough for a small firm.
- A checklist sheet. Put the 32 IDs down column A of a spreadsheet, paste the AI's status list beside it, and use a COUNTIF formula or conditional formatting to highlight any ID that is not OK. One glance at red cells before the job is closed.
- A job app with required photos. If you already pay for one, check what its own checklists and AI do before building anything. CompanyCam's pricing page says its AI features are included on Crew and above (Crew is listed at $129 a month billed annually), and that Crew includes standardised checklists. Jobber lists checklists from Connect and Housecall Pro from Essentials. The wider picture of installer software is in the solar installer software tutorial.
Commissioning tests stay manual
The photo audit is a check on evidence. The tests that prove the system is safe and working are a separate job, done by a qualified person with calibrated instruments, and no photo or AI summary replaces them. The international standard for this, IEC 62446-1, covers the documentation to hand over after installation of a grid-connected PV system, together with the commissioning tests and inspection used to verify safe installation and correct operation. Summaries of it cite polarity checks, string short-circuit current tests and earth and insulation checks, and it expects verification by a skilled person. The exact tests, limits and order come from the standard and from the installation rules and grid operator requirements that apply to you, so take them from there, not from a chat.
Three rules keep the line clear.
- The assistant never records a test value. Type each reading into the test sheet from the instrument. The photo of the instrument display, or of the completed sheet (E1), is backup. A misread 0.5 as 0.6 on a transcription is the kind of error you only find in a dispute.
- The assistant never says pass or fail. The prompt forbids compliance statements, and the person who ran the test signs it.
- A missing test is not a missing photo. If E1 shows a test sheet with a blank box, the line is incomplete, whatever the photo quality. Put "any blank box on the sheet" into the E1 wording so the model reports it.
From sorted photos to the handover pack
Once the gaps are closed, the same assistant can do the dull part of the pack: captions and an index. Each photo gets its checklist ID, its location and a plain description, and the index goes at the front of the file the customer receives. A before and after on one photo (illustrative):
Before: IMG_4412.jpg
After: B3 String 1 label at the array end. String number and polarity
readable. 14:07, job J2417.Ask for captions in the same prompt run, tell it to describe only what is visible, and read a sample before accepting the lot. The pack itself, the paperwork it should hold and how to gather it, is covered in the construction handover pack checklist, and the customer-facing items such as the signed handover sheet (E5) sit alongside the index.
Finally, keep the audit summary with the job file: the table, the retakes and the date they were closed. If a claim arrives in two years' time, the file shows a complete evidence set and the date it was complete. That record is the real product of the exercise, and it only exists if the check runs on every job, not just the ones that feel risky.
Questions installers ask about AI photo checks
Can AI replace the final inspection or the commissioning tests?
No. It can check that photos exist and are legible. Commissioning tests and inspection need a qualified person with calibrated instruments, following your installation standard and grid operator rules. Treat the photo audit as a check on the evidence, never a sign-off on the work.
How many photos should a residential job produce?
Enough to cover every checklist line, plus a second distance for labels and plates. In the worked example, 38 photos covered a 32-line list, because strings need one label photo each and some shots were retaken. Count lines first, then add duplicates for anything you would have to prove.
Can the AI read inverter serial numbers reliably?
Not reliably. Similar digits get confused, and images are resized before analysis, so small print blurs. Use it to check that a serial photo is legible, then type the number from the plate or the portal and keep the photo as backup.
Is it safe to upload customer photos to a chat assistant?
Photos can show house numbers, vehicles and people. Business plans do not train on your content by default, and consumer plans have a data-controls setting; check yours. Crop where you can and keep photo consent in your customer terms.
Which assistant is best at sorting photos?
Test them. ChatGPT, Claude and Gemini all accept image batches but differ in limits. Run each on a job where you have removed three photos on purpose, and choose by how many of the planted gaps it finds, not by reputation.
Further reads
- Snagging Report Template: Sections, Severity Order and AI Captions — Turn what you find into a snagging report with severity order.
- Water Heater Installation Checklist: Photo Proof and Handover Pack With AI — The same photo-proof method for another install trade.
- Solar Permit Application: Can AI Draft It? What Installers Must Verify — What AI can and cannot draft for permit applications.
- AI for Solar Installation: Five Jobs a Small Installer Can Automate — Five other jobs a small installer can hand to AI.
- CCTV Installation Checklist: Commissioning, Signage and Handover With AI — Commissioning, signage and handover for CCTV installs.
- AI for Electrical Project Management: Diaries, RFIs, Snag Lists — Site diaries and snag lists for electrical contractors.
- AI Solar Design Software Compared: Cost per Proposal for Small Installers — Aurora Basic and Premium, OpenSolar and manual design priced per proposal at 5, 10, 20 and 40 proposals a month, with the break-even for each.
- Construction Photo Documentation App: What AI Adds, and Which Plan — CompanyCam's AI credits change with the plan. Compare Core, Crew and Scale with cheaper options, set naming rules, then test one live job.
- EV Charger Installer App or Software: AI and Features Beyond Quotes — A neutral checklist for EV charger installers: survey forms, certificates, handover packs, scheduling, what a general job app can do and where AI helps.
- HVAC Warranty Claims: Document the Failed Part and File With AI — A contractor's claim pack for manufacturer warranty claims: plate photos, serial numbers, dates and failure notes, with AI to draft and check.
- Solar Installer Marketing: Local Pages and Case Studies Written With AI — Fact-sheet templates, prompts and claim checks for AI-drafted town pages and install case studies a solar installer can publish with confidence.
- Water Damage Restoration Software: Moisture Logs and Scope Notes Checklist — A checklist of what a water damage job file must record, what to ask restoration software, and where AI helps with moisture logs and scope notes.
- Can AI Design a Solar Array? What Design Software Does and Doesn't Do — What Aurora-style AI roof models and auto-layout tools produce, and the checks a qualified designer still has to make on a small roof.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: IEC 62446-1 (documentation, commissioning tests and inspection for grid-connected PV systems; scope as summarised in standards listings), Anthropic and OpenAI vision and file-upload documentation, Google Gemini help, CompanyCam, Jobber and Housecall Pro pricing pages, checked 1 November 2026. The checklist is a general illustration; apply your installation standard and grid operator rules. Examples are illustrative.