Dictate notes on site element by element with a location, observation, severity and photo number, transcribe them, then have AI map the notes into your report template's headings in plain English. Check every defect, measurement and recommendation against your photos before signing. AI drafts the wording; the findings and the liability stay yours.
The time goes in the write-up, not on site: turning "NE corner, damp, 18%, pic 42" into a paragraph a homeowner or client understands. That's the part AI does well. The danger is that it does it too well, adding a likely cause you didn't observe, or softening "further investigation required" into something a buyer or lender reads as reassurance.
Capture on site so the AI has something reliable
AI can only be as accurate as the notes. The single biggest improvement most surveyors can make is dictating to a fixed pattern, so every note carries the same five pieces of information:
- Element: roof coverings, chimney, rainwater goods, external walls, windows, and so on, in your template's order.
- Location: elevation and position, stated the same way every time ("rear elevation, left side, ground floor").
- Observation: what you saw or measured, not what you think caused it.
- Severity: your own rating in your template's terms.
- Photo reference: the photo number, said aloud.
Then, separately and only when you mean it, say "cause" and your opinion. Keeping observation and opinion apart in the dictation keeps them apart in the report. A before and after of the same note:
Before: "back wall's a bit damp by the kitchen door, probably the path, pic 42,
not too bad I think, check the gutter as well"
After: "External walls. Rear elevation, left of kitchen door, ground floor.
Damp meter readings up to 18 percent at low level over about one metre.
Severity: repair soon. Photo 42.
Cause: likely the path, which is level with the damp course.
Also: gutter over the kitchen door, see rainwater goods, photo 43."The second version takes ten more seconds on site and saves minutes at the desk, because the AI no longer has to guess what "a bit damp" means or which gutter you meant.
Decide in advance how you correct yourself out loud. Mid-sentence corrections ("no, damage to...") are where transcription goes wrong, because the software drops or moves the pause that carried the meaning. A simple convention works: say "scrap that", pause, then repeat the whole note from the element name. The prompt further down has a line telling the AI that only the repeated version counts. On a roof or beside a busy road, record the note again away from the noise rather than shouting over it; a transcript full of guessed words costs more time at the desk than the minute it takes to walk round the corner.
Getting speech and scribbles into text
You need a transcript before AI can draft anything. Pick one of these and use it consistently:
| Tool | How it works | Cost | Watch out for |
|---|---|---|---|
| Voice Memos on iPhone | Records and transcribes on the phone; transcripts on iPhone 12 or later, in a range of languages | Free with the phone | No speaker labels; check the transcript before you leave site |
| Transcribe in Word | Home, Dictate, Transcribe, then upload .wav, .mp4, .m4a or .mp3 audio; works in the new Edge and Chrome | Included with Microsoft 365; Microsoft's help page lists 300 minutes of uploaded audio a month on a standard subscription | A heavy survey month can exceed the cap |
| Otter | Records or imports audio, transcribes with timestamps, searchable | Pro $8.33 per user per month billed annually, or $16.99 monthly | Audio is processed in the cloud; check its data terms against your client obligations |
| Handwritten notes | Photograph pages with the OneDrive app's scan feature (Microsoft Lens has been retired) or Google Drive's scanner, then have AI read the scan | Included with those apps | Handwriting recognition misreads numbers; check every figure |
Whichever you choose, read the transcript once before it goes near the AI. Transcription errors in numbers ("18" becoming "80", "1.5 m" becoming "15 m") are the ones that survive into reports. The trade-off between machine and human transcription is covered in AI vs human transcription, and turning scanned pages into usable text in how to turn handwritten forms into data.
Handwritten pages need one extra step, because AI reading a scan mixes up 1 and 7, 5 and S, and decimal points that are really pen marks. Before asking for any drafting, ask for the figures on their own:
Read this scanned page of site notes. Do not draft anything yet.
List every number on the page with the words either side of it, one per
line, in the order they appear. Put a ? after any figure you are unsure of.An illustrative reply for a page covering the rear elevation: "moisture 18% at low level", "crack approx 3? mm wide, left of window", "photo 51", "photo 57?", "DPC approx 150 mm above ground". The two question marks are the useful part: the crack was 5 mm on site, and the second photo was 51 again, not 57. Correct the list, paste it back, and only then ask for the draft. It takes about two minutes a page and removes the error that is hardest to spot later, because a wrong number reads as fluently as a right one.
A report prompt built around your template
Give the AI your template's section headings, your severity scale and a firm set of rules. Keep the prompt in a Project (ChatGPT, Claude) or saved prompt so every report in the practice uses the same one:
You are helping a building surveyor turn transcribed site notes into draft
report sections. Use ONLY what is in the notes. Write in plain English for a
non-technical client, in the third person ("the surveyor noted...").
For each note:
- Put it under the correct template heading: [list your headings in order].
- Keep the location, measurements and photo reference exactly as dictated.
- Keep the severity rating exactly as dictated. Never upgrade or downgrade it.
- Only state a cause if the note says "Cause:". Otherwise do not suggest one.
- Where the note recommends further investigation, use the standard paragraph
[FURTHER-INVESTIGATION] word for word.
- If anything is unclear or missing (no photo, no location), write [CHECK: ...].
- If a note is followed by "scrap that", ignore it and use only the repeat.
Do not add general advice, maintenance tips or defects not in the notes.
NOTES:
[transcript]Run against the "after" note above, an illustrative draft looks like this:
EXTERNAL WALLS
At the rear elevation, to the left of the kitchen door at ground floor level,
the surveyor recorded damp meter readings of up to 18% at low level over
approximately one metre (photo 42). Rating: repair soon. The surveyor considers
the likely cause to be the adjoining path, which is level with the damp-proof
course. Lowering the path or forming a gap between the path and the wall is
usually a straightforward and inexpensive repair.
RAINWATER GOODS
[CHECK: note refers to the gutter over the kitchen door, photo 43, but no
observation was dictated for it.]The edit it needs: the last sentence of the first section, "usually a straightforward and inexpensive repair", appeared from nowhere. It breaks the "no general advice" rule and it's exactly the kind of reassurance a client quotes back if the repair turns out to involve more. Delete it. The [CHECK] flag on the gutter is doing its job: you meant to dictate an observation and didn't, so go to photo 43 and write it.
A standard-paragraph library the AI picks from
Most practices already reuse wording for common findings and recommendations. Put those paragraphs into a short library the prompt references by label, so the AI inserts your approved wording instead of paraphrasing it. Paraphrase is where liability language drifts. A few filled-in entries:
[FURTHER-INVESTIGATION]
The surveyor recommends that this is investigated further by a suitably
qualified specialist before you commit to the purchase or enter into a
contract, so that the cause, extent and cost of any repair can be established.
[MONITOR]
No repair is needed at present, but this should be checked periodically
and after severe weather. If it worsens, seek advice.
[SERVICES-NOT-TESTED]
The services were inspected visually only and were not tested. The surveyor
recommends a test by a qualified engineer before exchange of contracts.Twenty to thirty entries cover most reports. The rule in the prompt, "use it word for word", matters more than the library itself.
The library has a weak spot on unusual buildings. On a timber-framed house or one with a flat-roofed rear extension, the AI will reach for the nearest library paragraph even when it doesn't fit, because the prompt tells it to use the library. In one illustrative draft, a note reading "felt roof covering to rear extension, ponding over approx 1 square metre, severity: repair soon, photo 88" came back with the [MONITOR] paragraph attached, so a repair-soon item was followed by "no repair is needed at present". Add a rule: "If no library paragraph clearly matches, write [NO LIBRARY MATCH] and do not choose one." Write that recommendation yourself, and if the same finding turns up on later jobs, it becomes a new library entry.
A 1930s house, start to finish, at a two-surveyor practice
The practice in this illustration has two surveyors and a part-time administrator and produces mainly residential condition reports. A typical inspection of a 1930s three-bedroom house takes about two and a half hours on site; the write-up used to take another two and a half to three hours at the desk, working from handwritten notes and around 120 photos.
With the new routine:
- On site: dictation in the five-part pattern adds perhaps ten minutes. Photos are numbered in the camera and said aloud.
- Transcription: Voice Memos transcribes during the visit. The surveyor reads the transcript in the car for five minutes and corrects three numbers.
- Drafting: the transcript goes into the practice's Claude Team project with the prompt and paragraph library. A draft of all sections comes back in a couple of minutes.
- Checking: 45 to 60 minutes, going through the draft section by section with the photos open (the next section describes what to check).
- Final edit: the summary and overall opinion are written by the surveyor from scratch, because that's the part the client reads first and relies on most.
Desk time drops to roughly 75 to 90 minutes a report. Across eight reports a week between them, that's well over ten hours back, most of which the practice uses to take on more inspections rather than work later.
Two more jobs AI does well once the draft exists
Cross-checking your summary against the body
You write the summary yourself, but it's easy to leave out an item when the body runs to 30 pages. Once the checked sections are final, ask the AI to do the tedious comparison:
Below are my report's summary and its full body. List every item rated
"urgent" or "repair soon" in the body, with its heading and photo number.
Then list any of those items that the summary does not mention.
Do not rewrite anything.An illustrative reply: "Items rated repair soon: External walls, rear elevation, damp at low level (photo 42); Windows, first-floor rear bedroom, failed sealed unit (photo 67); Roof coverings, slipped tiles above bay (photo 18). Not in the summary: Windows, failed sealed unit (photo 67)." That's a real omission caught in seconds. Check the list against your ratings table once, because the model can miss an item as easily as you can, but it rarely misses the same one you did.
Formatting schedules for commercial work
Commercial work follows the same pattern with a different output. A surveyor preparing a dilapidations schedule for a catering company leaving a leased kitchen unit dictates each item with the lease clause, the defect and the remedy. The AI turns the transcript into the schedule's columns (item number, lease clause, defect, remedy, photo) in minutes. The one column it must never fill is cost. Costs come from your own price data or a contractor's figures, entered by you, because a guessed rate in a schedule that may be negotiated line by line is worse than a blank.
What AI gets wrong in survey reports, and how to catch it
| Error | How it shows up | How to catch it |
|---|---|---|
| Invented cause | "Likely due to a defective gutter" when you only recorded staining | Search the draft for "due to", "caused by", "likely"; each needs a "Cause:" in the notes |
| Severity drift | "Repair soon" becomes "attention may be needed" | Compare every rating with the transcript; the prompt forbids changes but check anyway |
| Swapped locations | Left and right elevations reversed, or two rooms merged | Check each location against its photo |
| Altered figures | 18% becomes 1.8%, or 1.5 m becomes 15 m | Tick every number against the transcript and photos |
| Boilerplate that doesn't apply | A paragraph about chimney flashing on a house with no chimney | Read every section against what the property actually has |
| Softened warnings | Your [FURTHER-INVESTIGATION] wording paraphrased | Search for the library's key phrases; they must appear verbatim |
Why read the transcript before drafting? An illustrative near miss: a surveyor preparing a schedule of condition for a sports equipment shop's new leased unit dictated "Shopfront. No, damage to the lower frame, left side, photo 12", correcting herself mid-sentence. The transcription came out as "Shopfront: no damage to the lower frame", and the draft duly said the shopfront was undamaged. In a schedule of condition that records the state of a unit at the start of a lease, that sentence could cost the tenant at the end of it. The photo showed the damage clearly; comparing the draft with photo 12 caught it. Make the photo check non-negotiable for every item, not a sample. For the general discipline of reviewing AI work without slowing to a crawl, see how to set up human review for AI work.
A second check catches a different slip: photo references that don't line up. Ask for a reconciliation once the draft is complete:
Here are the transcript and the draft report. List (1) photo numbers in the
transcript that are missing from the draft, (2) photo numbers in the draft
that are not in the transcript, and (3) any photo number used under two
different headings. Output the three lists only.An illustrative result on the 1930s house: "(1) 43, 96. (2) 24. (3) 18, under Roof coverings and Chimney." Photo 43 was the gutter already flagged [CHECK]. Photo 24 was a mistranscription of "photo 42". Photo 96, an internal wall crack, had dropped out of the draft entirely. Photo 18 genuinely showed both the slipped tiles and the chimney flashing, so that one needed nothing. Three short lists, two real fixes, about a minute of work.
Logging the first ten reports to see whether the routine works
Don't judge the routine on the first report, which will be slow while you learn the prompt's habits. Log the first ten instead, with four columns: desk minutes, [CHECK] flags, sentences deleted as invented, and ratings you had to correct. An illustrative log from the practice above:
| Report | Desk minutes | [CHECK] flags | Invented sentences deleted | Ratings corrected |
|---|---|---|---|---|
| 1 | 150 | 14 | 9 | 2 |
| 4 | 110 | 8 | 5 | 1 |
| 7 | 90 | 5 | 2 | 0 |
| 10 | 80 | 3 | 2 | 0 |
Read the columns separately. Falling [CHECK] flags mean the dictation is getting more complete, which is the surveyor improving rather than the AI. Invented sentences that stop falling at around two a report usually come from one habit of the model, often a closing reassurance like the "inexpensive repair" line earlier; add a specific ban to the prompt ("never comment on the cost or ease of a repair") and watch the next three reports. A corrected rating after report five is a signal to re-read the prompt's severity rule, because that is the error with the most consequence for the client.
Costs and data handling for a small practice
The tools are cheap next to the time saved. For a two-surveyor practice: phone transcription is free, Word's Transcribe comes with Microsoft 365, and an AI assistant on a business plan costs about $25 per seat monthly (ChatGPT Business or Claude Team Standard, $20 billed annually, minimum two seats) or $21 per user per month on annual billing for Microsoft 365 Copilot Business. Budget $50 to $70 a month in total, plus a few hours to build the prompt and paragraph library.
Use business plans rather than free consumer tools: survey notes carry client names, addresses and details of people's homes, and business plans exclude your content from model training by default. Keep each job's material in the client file, not scattered across chat histories. Add a sentence about AI-assisted drafting to your terms of engagement, and tell your professional indemnity insurer how you use it. The same notes-to-document pattern works for other site-based professions; AI inventory and check-out reports shows how letting agents apply it, and voice notes into job sheets and quotes how trades do.
Surveyors' questions about AI-drafted reports
Can AI look at my site photos and identify defects?
Image-capable assistants can describe a photo and will often name what they think they see, such as cracking or staining, but they can't judge cause, extent or significance the way a site inspection can, and they're sometimes confidently wrong. Use them to caption and sort photos against your notes, or to check a photo reference matches the element described, not to find or diagnose defects.
Should I tell clients that AI helped draft the report?
Say so in your terms of engagement: that you use software, including AI tools, to help draft reports from your inspection notes, and that every finding and recommendation is yours and checked by you. Clients mostly care that the surveyor inspected, judged and signed. Keep the statement accurate, and make sure your process matches it.
What should I ask my professional indemnity insurer?
Ask whether the proposal form or policy says anything about AI-assisted work, whether they expect a particular checking process, and whether using AI changes how they view a claim arising from a drafting error. Describe your actual routine: structured notes, AI drafting, a line-by-line check against photos. Get the answer in writing and keep it with your procedures.
Further reads
- How to Write SOPs With AI: From Rough Notes to Clear Procedures — Turning your dictation routine into a written procedure for the practice.
- Can ChatGPT Read PDFs, Spreadsheets and Photos? What Breaks — What AI tools reliably read from photos and scanned pages.
- How Much Does an AI Note-Taker Cost per Seat? — Comparing transcription tool costs per surveyor.
- How to Draft Risk Assessments With AI and What a Human Must Check — A similar draft-then-check method for site risk assessments.
- Can AI Help a Small Architecture Practice? Where It Saves Time — Where AI saves time in a neighbouring profession's paperwork.
- AI Hallucinations Explained for Business Owners: Causes and Fixes — Why AI adds plausible detail you never gave it.
- Writing Client Updates and Snag Lists With AI for Builders — Voice note on Friday, client update by five; a room-by-room snag walk turned into a table your trades can work through, with the wording kept careful.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Apple Support, Voice Memos transcription; Microsoft Support, Transcribe your recordings in Word; Otter.ai pricing; Microsoft Lens retirement notice and OneDrive scanning help.