Number every photo and dictate short, tagged notes on site, then give the AI the survey template, the photo list and the notes, and ask it to sort them into report sections and list what is missing. It structures and words the report; you verify every measurement and every safety-relevant observation yourself.
A blank checklist tells you what to photograph. It does not tell an AI what a photo shows. OpenAI's help page on image inputs says ChatGPT struggles with precise spatial tasks, may give approximate counts, and does not read file names or metadata. So the numbering, the tags and the measurements have to be yours, and the checklist below is built around that.
Tag every photo so the AI cannot misfile it
The tool sees pixels and the words you give it. It does not see the file name, the time stamp or the order in which you stood in the loft. That is why the capture system matters more than the prompt. Use three habits and the rest of the process becomes routine.
- Section markers. Before each group of photos, photograph a sheet of paper with the section written on it in marker pen: "SITE", "ROOF R1 EAST", "SHADE", "ELECTRICAL", "CABLE ROUTE". Each marker doubles as a divider for you and as a label the AI can read.
- A photo log written as you go. One line per photo in a notes app or a voice note: number, tag, what it shows, any measurement and where it came from. The number is the one your phone's gallery shows, so you can match them later.
- Tag letters. R for roof, S for shade, E for electrical, C for cable route, A for access and site, X for hazards and anything else. A note with no tag is a note nobody can sort.
Here is a short stretch of an illustrative log. Each line states its measurement source, so the report can say "measured" or "estimated" and not guess.
P5 | R | R1 east face overview from ground | pitch 35 deg (digital level, on roof)
P6 | R | R1 tile close-up | concrete tiles, 2 slipped at eaves
P11 | R | loft, R1 rafters | 100 x 50 mm at 400 mm centres (tape)
P13 | S | tree at boundary | deciduous, shades R2 low sun, winter (observed)
P15 | E | meter and cut-out | single-phase (label read)
P17 | E | consumer unit open | main switch 100 A (label read), 8 ways, 0 free
P20 | C | loft hatch to proposed inverter spot | 4.5 m from roof entry (laser)
NP | A | customer wants inverter out of sight | verbal, no photo"NP" stands for "no photo", which keeps spoken information in the same list as visual evidence. A survey note that exists only in your head will not reach the report.
The worked survey used from here on is an illustrative semi-detached house with two roof faces, photographed in 22 shots. The numbering below is the one the later examples quote.
| Group | Photos | What they record |
|---|---|---|
| Site (A) | 1 to 3 | Section marker, front elevation, side access with the gate width measured |
| Roof R1, east face (R) | 4 to 7 | Marker, face overview, tile close-up, eaves and gutter |
| Roof R2, west face (R) | 8 to 10 | Marker, face overview, chimney stack close-up |
| Loft (R) | 11 | Rafter size and spacing with a tape across |
| Shade (S) | 12 to 13 | Shade tool reading from the proposed array position, tree at the boundary |
| Electrical (E) | 14 to 18 | Marker, meter and cut-out, consumer unit closed, consumer unit open, free ways and labels |
| Cable route (C) | 19 to 22 | Marker, loft hatch to the inverter spot, route to the consumer unit cupboard, Wi-Fi signal screenshot |
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 nowThe photo checklist: what to shoot, why, and how to verify it
Work through the groups in the order of a typical visit. Each item is a prompt for a photo and a note; the last column is what you do on site so the report does not rest on the AI's reading of a picture. The consumer unit is the main distribution board of the home, which some call a breaker panel or fuse box.
Roof
| Capture | Why the report needs it | Verify on site |
|---|---|---|
| Each roof face from the ground, full face in frame | Fixes which face is R1, R2 and so on for every later note | Name the faces in your log; add the compass bearing |
| Covering close-up, edge of the face, ridge and eaves | Type and condition of the covering drive the fixing method and any repairs | Note type, any slipped, cracked or missing pieces, and age if known |
| Obstructions: chimneys, vents, skylights, dishes, dormers | They reduce usable area and cast shade | Measure each one's position from the nearest edge with tape or laser |
| Usable dimensions of each face, with the tape visible | The designer lays out panels from these numbers | Measure width and slope length yourself; note the setbacks you assumed |
| Pitch, with the level on the roof surface | Affects output modelling and fixing choice | Read the digital level; do not estimate from a photo |
| Loft: rafters or trusses, spacing, timber condition, any damp | Structural suitability is judged on this evidence | Tape across spacing, note section sizes, flag anything unusual for a structural opinion |
| Access: ladder or scaffold position, fragile sections | Sets the install method and the safety plan | Mark fragile areas in the log; stop if you suspect asbestos-containing material and follow your procedure |
Shading
Shade is where AI is least trustworthy, because it depends on sun position, time and season that a still photo does not show. Use photos to record what is there and a proper tool to judge what it does.
- Standard photos, not panoramas. OpenAI lists panoramic and fisheye images among those the model struggles with. From each proposed array position, take three ordinary photos: towards the midday sun direction, then about 45 degrees to either side. Add a marker photo noting the time, date and sky.
- Every object that could shade. Trees, neighbouring buildings, chimneys and roof obstructions get their own photo. Write whether a tree is deciduous, since winter shade differs from summer.
- A shade reading or a model. Use the shade tool or the design software's shading analysis, and note the reading and its source next to the photo. The AI must never turn a photo into a shade percentage.
Consumer unit and supply
| Capture | Why the report needs it | Verify on site |
|---|---|---|
| Meter and incoming cut-out or main fuse, labels legible | Shows supply type and rating, and where the incoming supply enters | Read the labels yourself; note single-phase or three-phase as printed |
| Consumer unit closed, with its location and surroundings | Space for isolators and whether the position is practical | Note room dimensions, cupboard depth and any obstructions |
| Consumer unit open, only if you are competent and it is safe to do so | Number of ways, free ways, protective devices and labelling | Read the main switch rating; count ways and free ways; never remove covers on live equipment unless trained |
| Visible condition: heat marks, loose cables, signs of earlier DIY work | Anything unsafe changes the next step | Record what you see in words; flag it to the office the same day |
| Existing solar, battery or EV charger | Affects connection and capacity | Photograph the equipment labels and the isolators |
Cable routes and inverter position
Photograph every leg of the proposed route with a tape or laser reading visible or noted: roof entry, loft run, proposed inverter spot, run to the consumer unit, and the battery position if one is planned. Each leg gets a measured length in the log and a note of the wall or floor type it passes through. Add a photo of the proposed inverter wall, and one of anything the route must avoid or cross, such as other services. Ask the customer where they want the inverter and whether visible cabling matters, and record the answer as an "NP" line.
Site, access and the customer
Photograph the front of the property, the side access with its width measured, parking and any scaffold position. Note pets, working-hours limits and the customer's own goals in their words. Screenshot the Wi-Fi signal at the proposed inverter position, because a weak connection is a monitoring problem found now or on the install day.
Prompt one: describe each photo, then correct it
Do not ask for the report yet. First ask for a plain description of each photo against your log, and read it. The AI's reading is a check on your log and your photos, not a source of new facts. Upload the photos in order, and put your log in the prompt, because the AI cannot see the numbers in your gallery.
I am attaching 22 survey photos in order. Photo 1 is a section
marker. My photo log is below, with my numbers and tags.
For each photo give: the number; one line on what is visible; the
checklist item it supports. Write UNCLEAR or UNREADABLE where you
cannot tell, instead of guessing. Do not measure, count panels,
estimate heights or distances, or judge safety or suitability.
List anything visible that my log does not mention.
PHOTO LOG:
[paste]Illustrative output often contains one miss that pays for the step.
| AI says | Your log says | What you do |
|---|---|---|
| Photo 17: consumer unit open, main switch reads 63 A | Main switch 100 A, read from the label | Your reading stands. Crop and zoom the label, check it, and record "100 A (surveyor, label)". The AI misread an angled label |
| Photo 9: west face with a chimney stack and a satellite dish | Chimney only | The AI found an obstruction you missed. Check on the roof, measure the dish's position and add it to the log |
| Photo 13: deciduous tree, about 10 m tall, south-west of the house | Tree at boundary, deciduous | Delete the height. Estimating heights is excluded in the prompt, and a tree height needs measuring on site |
| Photo 3: UNCLEAR, gate width | Gate 0.9 m (tape) | The tape is not readable in the photo. Retake or keep the log figure and mark it "tape, not photographed" |
Prompt two: draft the report from corrected captions and notes
Now paste the corrected log and a report template. Ask for your headings, a list of conflicts and a list of gaps. The two lists are the point: a survey that arrives with its own open questions saves a phone call.
Draft a solar survey report using ONLY my corrected photo log and
notes below. Use these sections: 1 Summary and open items; 2 Property
and access; 3 Roof faces (a table: face, covering, pitch, usable
width x slope length, bearing, obstructions, source of each figure);
4 Shading; 5 Electrical supply and consumer unit; 6 Cable routes (a
table of legs and lengths); 7 Hazards; 8 Questions for the customer
and for the office; 9 Photo appendix (number and caption).
Rules: never infer a measurement. If a figure has no stated source,
write "not recorded". Keep my units; do not convert. Do not write
suitable, compliant, safe, adequate or approved about anything.
Describe what was observed. List every conflict between notes and
photo captions. Put hazards at the top of section 1.
CORRECTED LOG AND NOTES:
[paste]Back to the semi-detached house in the photo table: concrete-tiled roof, an east face (R1) and a west face (R2). The notes record usable areas of 5.8 m by 3.1 m on R1 and 5.8 m by 2.4 m on R2 after setbacks and the chimney clearance, both at 35 degrees. They also record a single-phase supply, a main switch of 100 A, an eight-way consumer unit with no free ways, a loft inverter position 4.5 m from the roof entry, and an 11.0 m run from the inverter to the consumer unit cupboard. Section 1 of the draft should read something like this (illustrative):
Hazards: none recorded. Open items: (1) consumer unit has no free
way (photo 17); an electrician must assess before any new circuit is
specified. (2) Deciduous tree at the boundary shades R2 in low winter
sun (observed, photo 13; shade tool reading, photo 12); designer to
model. (3) Satellite dish on R2 added after photo review; position
measured at 1.2 m from the ridge.
(4) Customer prefers the inverter out of sight; loft position
proposed. Not recorded: loft boarding, scaffold position.That is what you want: open items first, each with its photo number and no recommendation dressed up as a fact. Compare it with what a first draft often does on its own. It writes "the consumer unit appears suitable for the new circuit", which no photograph can support. It writes "room for 16 to 18 panels", which you can test in seconds. If your panels are about 1.7 m by 1.1 m (check the datasheet), each takes close to 1.9 square metres. The two faces give 17.98 plus 13.92, so 31.9 square metres. Eighteen panels need about 34 square metres, which is more than exists, and sixteen need about 30.4, which would take near-perfect packing of two separate faces. The layout belongs to the designer anyway. Delete the sentence and write "layout to be confirmed in design software". The tutorial on what solar design software does and does not do explains why.
What you must verify yourself, and how
AI has structured the notes. It has not checked anything. Keep this table clipped to your tablet so the verification does not slide into the AI's column.
| Item | Why AI cannot do it | Your check |
|---|---|---|
| Roof dimensions and setbacks | Images are resized, and spatial tasks are a stated weakness | Tape or laser on site; record the source beside each figure |
| Pitch and bearing | A photo cannot give angle or direction reliably | Digital level on the roof surface; compass bearing, noting the tolerance of your tool |
| Shade impact | Needs sun path, date and time | Shade tool or design software model, with the reading noted |
| Roof structure and covering condition | Condition needs hands-on judgement | Loft inspection, rafter sizes and spacing, a structural opinion where in doubt |
| Electrical capacity and safety | A label in a photo is not an assessment | Qualified electrician decides on the consumer unit, protective devices and the new circuit |
| Hazards, including suspected asbestos or fragile areas | Absence from a photo is not absence on site | Stop work, follow your procedure, record the observation in words |
| Grid or supplier notification needs | Depends on the connection and the approving body | Check with the body that approves connections, before the proposal |
When the layout needs the designer, the hand-over is the report plus the photo appendix, which is why the open-items section comes first. If you also survey for EV chargers, the EV charger site survey checklist shares the consumer unit and cable-run photos, so one set of shots can serve both visits.
Upload limits, formats and privacy before the first survey
Check the practical limits of the tool you will use before you are in the loft. For ChatGPT, OpenAI's help pages give a limit of 20 MB per image and list PNG, JPEG and non-animated GIF as supported formats, so convert any HEIC photos to JPEG, or set the phone camera to its most compatible format. They also say paid plans can upload up to 80 files every 3 hours, while free accounts are limited to 3 file uploads a day, and that OpenAI may lower the limits at peak times. A 22-photo survey will not fit a free account.
If you are on a plan with a tight cap, try putting four numbered photos on each page of a PDF and attaching the PDF directly to the chat, not as a saved project file, because OpenAI notes that images embedded in project files may not be included automatically. Test it first: ask for the main switch rating from the consumer unit page and check the AI reads it correctly. If it cannot, use individual images for the electrical group and the PDF for the rest.
The photos show a customer's home, so treat them as personal information. Keep faces, letters, vehicle plates and screens out of frame, and remove location data from photos if you share the images beyond your own systems. OpenAI's upload FAQ says how uploaded images may be used depends on the service and your data settings, so check the plan's data terms and tell the customer in your survey terms that software tools process photos. Rules differ between markets, so ask an adviser once.
How to tell the report worked: a reconciliation and a call-back count
Two checks run in under ten minutes. The first is a number reconciliation: ask the AI, in a fresh chat, to list every number in the report with the log line that supports it. Anything it cannot match goes in the "not recorded" list, not into the report. Then read the list against your own notes. The second check takes longer to show results: count the call-backs from the designer or the office per survey. If the designer keeps ringing to ask about the same missing item, add a line to the template, because the report is missing a field and not the surveyor a skill.
For a sense of scale, suppose a surveyor used to spend 70 minutes on each report after a visit. If the routine above brings that to about 25, it is because the sorting and the first draft are fast, while the corrections and checks stay with you. Treat the figures as an illustration to compare against your own stopwatch. If your first week is slower than your old method, the cause is usually the log: a thin log gives the AI nothing to sort.
Where survey reports drift: photo numbers, reassurance and units
- Photo numbers drift. Photos are re-uploaded in a different order after one fails, and every caption shifts by one. You notice when photo 17 is described as a loft rafter. The describe-first pass catches it, and the section markers give you a quick way to re-anchor.
- The reassuring sentence. The draft says the unit "appears suitable" or the roof "looks sound". Nobody inspected either to that standard. Keep the banned-word rule in the prompt, and search the final text for suitable, compliant, safe, adequate and approved.
- Units and conversions. A survey that mixes millimetres, metres and feet is easy to misread, and an AI that helpfully converts can introduce errors. The rule "keep my units; do not convert" prevents it. If the office needs another unit, convert in the spreadsheet where the formula is visible.
Once the report is solid, it feeds the proposal and the permit paperwork. The solar installation quality checklist uses the same photo discipline after the install, and the tutorial on how surveyors turn site notes into reports with AI covers the wider habit for any survey type. When enquiries arrive and you want the pre-survey questions asked before you drive out, see the tutorial on solar chatbots that qualify leads by bill, roof and shade.
Questions surveyors ask about AI survey reports
Can AI measure the roof from my survey photos?
No. OpenAI's help page for ChatGPT image inputs says the model struggles with precise spatial tasks, may give approximate counts, and resizes images before analysis. Measure with a tape, laser and digital level on site, record the source of each figure in your notes, and leave layout fitting to the design software.
How many photos should a solar survey include?
Enough to cover every checklist group with a clear photo and a tag. The worked example in this tutorial uses 22 photos for a two-face roof, including section markers. A complicated roof, a second building or a battery location will need more. Missing photos cost a return visit, so over-shoot rather than under-shoot.
Should I tell customers their photos go into an AI tool?
Yes, in your survey terms or on the day. Photos of a home are personal information, and how uploads are used depends on the service and your data settings. Keep faces, documents and vehicle plates out of frame, check the plan's data terms, and ask an adviser about the rules in your market.
Should the report say the consumer unit is suitable?
No. Record what you saw: location, main switch rating from the label, number of ways, free ways, visible condition. Whether the unit can take the new circuit and protective device is a decision for a qualified electrician. Ban words such as suitable, compliant and safe from the prompt.
Which AI tool works for survey photos?
Any assistant that accepts images and lets you set rules. For ChatGPT, OpenAI's help pages give a 20 MB limit per image, JPEG and PNG formats, and a free-plan cap of 3 file uploads a day, so batch photos into a numbered PDF if you are on a free plan and check it stays legible.
Further reads
- Solar Permit Application: Can AI Draft It? What Installers Must Verify — The permit paperwork that follows once the survey is in.
- Solar Lead Generation: Cost per Lead, Cost per Survey and AI Follow-Up — How surveys get booked and what each one costs to win.
- Air Conditioning Installation Business: Survey Notes to Quote With AI — Another trade's route from survey notes to a priced quote.
- Damp Survey Report: Write Findings and Quote Lines With AI — Findings to quote lines in a different survey type.
- Thermal Imaging Survey Report: Turn Scans Into Plain-English Findings — Plain-English findings from scan images and notes.
- Solar Battery Quote Template: Options Explained With AI, No Backup Promises — What the survey must capture before a battery is quoted.
- 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.
- EPC Assessor Software: What to Check First, and Where AI Helps — Check scheme approval, approved-list status and update speed before you pay, then use AI for client emails and wording, never for lodgement.
- 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.
- Heat Pump Installer Software: Surveys, Heat Loss and Quotes, AI Compared — What a heat pump installer's software must do at survey, heat loss and quote stage, which tools cover which job, where AI helps and what vendors charge.
- Solar Installer Software: CRM, Design and Tracking, and Where AI Helps — Buy order, vendor prices and three priced stacks for a five-person solar installer, with what AI does in design, CRM and tracking, and what it does not.
- Heat Pump Quote Template: Explain Options With AI, No Savings Promises — Write a heat pump quote with AI that explains options, assumptions and exclusions, leaves design to the designer and never promises savings or grants.
- Can AI Write a Solar Sales Script? Discovery Questions and Objections — How to have AI draft solar discovery questions and objection handlers that stay honest, with role-play prompts and a claims checklist.
- AI for Solar Installation: Five Jobs a Small Installer Can Automate — Five installer jobs with illustrative hours saved each: lead qualification, survey notes, proposal wording, permit paperwork drafts and customer updates.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI Help pages (ChatGPT image inputs FAQ; uploading files and audio to ChatGPT), read November 2026.