Use of AI in Building Surveying: Client Data and UK GDPR Basics

Often yes, but only through the right plan and with identifiers removed first. A building surveyor can use ChatGPT for drafting if the site notes are redacted, the account is a business plan with data processing terms, and the client's details stay out of the prompt. Raw client files on a free account are the risky case.

This is general information, not legal advice. If a client contract, a professional body rule or an insurer says something stricter, that wins, and a data-protection adviser can confirm your set-up. The plan you pay for matters less than people expect, though: even on a business plan, the safest prompt is one that never contained the client's name, address or circumstances in the first place.

Follow me on Instagram@sagnikteaches

What counts as personal data in a survey file

UK GDPR applies to information that identifies a living person, directly or in combination with other details. A building survey file holds more of it than most surveyors notice at first glance, because the report is about a property but the file is about people.

Connect on LinkedInSagnik Bhattacharya

• Obvious identifiers: client names, addresses, phone numbers, email addresses, solicitor and lender details, and instruction reference numbers that map to a person.
• Property details that point to a person: a postal address, a full address code, a title number, a plot number on a small development, or a very unusual building.
• Site photographs: people in frame, family photographs on a mantel, post on a doormat, number plates, a child's bedroom, a visible alarm code or a key safe.
• Notes about occupants: illness, disability or mobility adaptations noted during access, religious items, or comments about the vendor's circumstances. Health information is a special category of data and needs extra care.
• Financial details: mortgage valuation figures linked to a named client, purchase price, fee quotes in a named client's file.
• Third parties: neighbours named in a boundary or party-wall note, tenants, contractors.

A photograph of a roof void with no people and no address is probably not personal data on its own. The same photograph filed under a client's name, with the property address in the filename and metadata, becomes part of a personal file. The test is whether someone could work out who the information is about.

Subscribe on YouTube@sagnikteaches

Yes or no: what can go into a general AI tool

The table assumes the same two set-ups throughout. "Consumer plan" means a free or individual account. "Business plan" means a team or business account with a data processing addendum accepted, and with member settings controlled by an admin. Neither is a safe harbour for anything you are not allowed to share under your terms of engagement.

The pattern is simple. Anything that helps the AI write a better paragraph about a defect is fine once it is anonymous. Anything that only identifies the client or the occupants is not needed for drafting, so leave it out of the prompt.

The redaction routine, step by step

This takes under five minutes on a typical set of notes once it is a habit. Do it in a plain text editor or a document that is not synced to the AI tool, so the original stays untouched.

1. Copy the notes into a scratch document. Never edit your master file.
2. Replace the client. "Mr and Mrs [surname]" becomes "the client". Replace every first-name use as well.
3. Replace the address. Use a descriptor: "a 1930s semi-detached house, three bedrooms, brick with a tiled roof". Keep the era, type and construction, because those are what the AI needs.
4. Remove location clues in the text. Street names, estate names, neighbours' names, landmark references, and the name of a school or business next door.
5. Remove reference numbers. Your instruction number, the lender's reference and any title number.
6. Remove personal circumstances. "The vendor is elderly and unwell" tells the AI nothing about the roof. Replace with "access was limited" if relevant.
7. Check photos. Crop to the defect. Blur anything personal. Strip location and device data from the files, which most phones and photo apps can do on export.
8. Reread as a stranger. If a neighbour could say "that's the house at the end of our road", reduce detail further.
9. Keep a key. Note "Client A is Instruction 2026-114" in your own system, not in the prompt, so you can reinsert details in your own document afterwards.

A before and after helps. This is invented text for illustration.

Before:
Inspected 14 [street] for [client name] on 12 June. Vendor [name] (recovering from a hip operation, ground floor only). Rear gable damp reading 28% on timber near chimney breast. Lender: [lender] ref 77-3321. Neighbour at no. 16 has a lean-to against the party wall.

After:
Inspected a 1930s semi-detached house for the client. Access limited to the ground floor. Rear gable, timber near the chimney breast, damp meter reading 28%. A neighbouring lean-to roof abuts the party wall.

The redacted version loses nothing the AI needs to describe the damp issue, and loses all the names, a health detail, the lender reference and the street.

A damp finding, from redacted notes to a draft section

Take a surveyor working alone who carries out about eight inspections a month. After a visit, they have 600 words of rough notes and 14 photos. Using the routine above, they redact the notes, crop three photos to the damp patch, the chimney breast and the roof junction, and paste them into a business-plan chat with this prompt.

I am a building surveyor. Below are redacted site notes and three photos from a pre-purchase inspection of a 1930s semi-detached house. Draft the "damp and moisture" section in plain language for a non-expert client, 200 to 250 words. Use only what is in the notes and photos. Give likely causes in order of likelihood, say what further investigation is advised, and avoid costs. Mark anything you are unsure of with [CHECK].

An illustrative reply: it describes a reading of 28% on timber near the chimney breast, lists defective flashing, a failed lead tray or rain penetration through the stack as possible causes, recommends opening up and inspecting the roof junction, and flags [CHECK] on whether the reading was taken with a pin or a non-invasive meter, because the notes do not say.

The surveyor then does the part only they can. They reinsert the client's name and address in their own document, correct the wording to match what they actually saw, and decide on the recommendation. The [CHECK] was a good catch: the surveyor knows the meter type and adds it. Time saved on the section: perhaps 20 minutes. Personal data sent to the provider: none.

The parent workflow, turning site notes into reports, is covered in how surveyors use AI to turn site notes into reports. This tutorial only deals with the data side.

Plan settings that matter for client files

Settings differ by product and change often, so check the vendor's own pages before you rely on any of this. The questions are stable, even when the labels move.

• Is your content used to train the models? OpenAI says it does not train on data from ChatGPT Business, Enterprise, Edu or its API by default. Anthropic says the same for its commercial products such as Claude for Work and the API. On consumer plans of both, you can switch off use of your chats for training in privacy settings. Do this, though it does not make a consumer plan suitable for client files.
• How long is content kept? Consumer retention differs from business retention. For Claude's consumer plans, for example, Anthropic's privacy centre says chats used for model improvement may be kept in de-identified form for up to five years; with that setting off it states no fixed period beyond "as long as reasonably necessary", and deleted chats leave its back-end systems within 30 days. Business plans often let an admin set retention. Look it up for your plan.
• Who can see the chats? On a business plan, an admin may be able to see or export member activity. On ChatGPT Business, removed members' chats are kept and owners cannot see them, so leave-and-joiner processes matter.
• Where is data stored? Some business plans now let you choose a storage region. If you act for clients who need data kept in a certain place, that is worth checking.
• Which features are on? Memory, connected apps, browsing and shared links can all spread what you type. Turn off what you do not use, and never share a chat link containing client material.
• Who pays for seats? ChatGPT Business and Claude Team both need at least two seats, so a sole practitioner pays for two or uses an individual plan with extra care.

Data processing terms: what to look for

When an AI provider handles personal data for you, the provider is usually acting as a processor and you are the controller, meaning you decide why and how the data is used. Data-protection law expects a written contract between the two covering, among other things, that the provider acts only on your instructions, keeps data confidential and secure, tells you about breaches, helps with requests from individuals and deletes or returns data at the end.

In practice this lives in a data processing addendum, often called a DPA. OpenAI publishes a Data Processing Addendum for its business services, and Anthropic states that its DPA is incorporated into its Commercial Terms, so accepting the terms accepts the DPA. Read them once. Check three things: what the provider may do with your data (only deliver the service), whether it uses sub-processors and how it tells you about changes, and what happens to data transferred outside the country, where a recognised transfer mechanism should be named.

Consumer plans normally run under consumer terms without that contract. That is the main reason this tutorial's table says no to client data on them, even where the training switch is off.

Also look at your own paperwork. Your privacy notice should say that client data may be processed by service suppliers, including AI tools, and why. Your terms of engagement may need a line if clients have signed up to specific confidentiality terms. And your professional indemnity insurer may have a view on AI-assisted reports, so ask what they expect.

Three more worked cases

Case 1: a client sends their own documents

A client emails a lender's valuation, their solicitor's enquiries and a floor plan, and asks you to "run it through AI to summarise the risks". The documents carry names, a mortgage reference and financial information, and the lender's confidentiality terms may restrict onward sharing. Do not paste the PDFs. Read them, write your own notes on the points that matter, redact those notes and ask the AI to help structure the advice. If the client wants a summary produced by AI from their documents, discuss it first and get their agreement in writing.

Case 2: a photo with an occupant in frame

A defect photo of a bathroom ceiling stain also shows the vendor's grandchild's drawings on the fridge door through an open doorway, and a prescription box on the cistern. Crop tightly to the stain, or retake the photo. Check the original has no location metadata. Keep the original in your secure file, and upload only the cropped copy.

Case 3: a multi-unit block with named leaseholders

A building safety inspection lists flat numbers with leaseholders' names, some in arrears and one with an accessibility need. Use "Flat 1" to "Flat 8" labels in the AI prompt, drop the arrears and accessibility details altogether, and keep the key in your own system. The AI only needs to know that flats 3 and 6 have the cracked stair string, not who lives there.

Where client data leaks out of a survey workflow

• The file name leaks it. You upload "[surname]_14_[street]_photo3.jpg". Rename files to "photo-03" before upload.
• Metadata carries the location. Phone photos hold GPS coordinates. Strip them, or take photos with location off on a work device.
• Autofill and memory. A tool's memory feature stores a detail you typed last week and reuses it. Turn memory off for work chats, or clear it regularly.
• Browser extensions and add-ins. A summarising extension reads whatever is on screen. Disable extensions when a client file is open.
• Shared accounts. Two people use one login, so you cannot tell who saw what. Use named seats on a business plan.
• The output is trusted blindly. The AI misdescribes a defect, and you sign the report. The data routine protects privacy, not accuracy. You still check every statement against your own notes.

When a client asks what you did with their data

Sooner or later a client will ask whether AI saw their survey. Data-protection law gives individuals the right to ask what personal data you hold and how you use it, and you generally have to answer within a month. A routine you can describe in two sentences makes that painless. With the policy above, the answer is: "We use AI tools only on anonymised notes, with no names, addresses or reference numbers, on a business account with a data processing agreement." If you cannot say that truthfully, fix the routine, not the sentence.

Two practical points help. Keep your redacted notes and the AI draft in the job file for as long as you keep the rest of the job, then delete them together, so you are not holding stray copies in a chat history. And delete or archive old chats on the AI account on a schedule, for instance when you close the job, since a chat history is another place client-related material can sit. If a client asks you to delete their data, check your own retention duties first, such as insurer or professional body requirements, and answer on the basis of both.

A short sample reply for a client who asks: "Thank you for asking. We use an AI writing tool to help structure the wording of reports. Before anything is entered we remove your name, address and reference numbers, so the provider does not receive information that identifies you. The findings and recommendations are mine, checked against my own notes and photographs. If you would prefer that no AI tool is used on your instruction, tell me and I will write it by hand." Offering the opt-out costs little and tends to reassure more than it worries.

A one-page policy you can adopt today

Write this down, even in a sole practice. It shows clients and insurers that you thought about it, and it makes the habit stick. Adapt the wording.

AI use policy (draft for [firm name])
1. We use AI tools only on a business plan with an accepted data processing addendum.
2. Client names, addresses, reference numbers and occupant circumstances are never entered.
3. Site notes are redacted before use; photos are cropped and stripped of metadata.
4. Memory, shared links and third-party connections are off for work accounts.
5. Every AI-drafted passage is checked against the surveyor's own notes before it is issued.
6. Our privacy notice and terms of engagement mention AI tools among our suppliers.
7. Reviewed every six months, or when a provider changes its terms.

Review it when a provider changes its terms, which happens more often than you might expect. A quick way is to keep a list of the vendor pages you relied on and a date in your diary to recheck them.

If you also use AI in fee work, a related tutorial walks through a worked fee quote and enquiry reply, using the same redaction habit. For another view on cautious anonymising, see how to anonymise client data before you paste it into AI, and for the vendor's side of the question, the privacy policy checklist.

The closing point: your duties do not stop at the AI tool. Keep a record of what you used, when, and with what safeguards. If a client or regulator asks, a simple log and the policy above answer most of the question. For anything that goes beyond these basics, such as automated decisions about individuals, large volumes of client data or a client who requires specific safeguards, ask a qualified data-protection adviser.

Data questions surveyors ask about AI tools

Are site photos personal data?

A photo of a building alone usually is not, but it can become personal data once it shows people, number plates, family photographs, post addressed to someone, or is linked to a client name and address. Treat photos as personal data when in doubt, and crop or blur before uploading.

Do I need a data processing agreement with an AI provider?

If a provider processes personal data on your behalf, data-protection law expects a written contract setting out what it may do with it. Business plans usually include a data processing addendum you accept or sign. Consumer plans usually do not, which is one reason to avoid them for client data.

Is anonymising data enough?

Removing names and addresses helps a great deal, but data stays personal if the person can still be identified from what remains, including by combining details such as a rare property feature and a street. Redact identifiers, then reread the text as a stranger would.

Can I use AI to write a survey report if I redact first?

Yes. Paste redacted site notes, get a draft, and reinsert names and addresses in your own document. You remain responsible for the findings, so check every defect description and recommendation against what you saw. Keep the redacted notes with the job file, so you can show exactly what the tool received if a client asks.

Should I tell clients I use AI?

Your privacy notice should cover how client data is processed, including by your tool suppliers. Many firms also add a line to the terms of engagement. Telling clients plainly is good practice and avoids surprises, even where the law does not demand a specific wording.

Further reads

• AI for Land Surveyors: What Helps and What Never Replaces Field Checks — What AI helps with in survey work and what it never replaces.
• Does ChatGPT Train on Client Data? Business vs Free Plans — How business and free plans differ on training.
• Questions to Ask Before Buying AI That Touches Client Data — Vendor questions to ask before any client data goes in.
• Reusing Past Client Work With AI Without Leaking Client Data — Reusing old reports without carrying client details over.
• GDPR for Photographers: Client Galleries, Contracts and AI Tools — Another photo-heavy profession working through the same rules.
• Thermal Imaging Survey Report: Turn Scans Into Plain-English Findings — Turn scan findings into plain English with the same care.
• Responsible Use of AI in Surveying Practice: A Small Firm's RICS Checklist — The RICS responsible-AI standard turned into a checklist for a six-person practice, with named owners, filled-in examples and a 90-day plan.
• Damp Survey Report: Write Findings and Quote Lines With AI — Turn readings and site notes into a clear damp report and quote lines in ChatGPT or Claude, while the surveyor keeps the diagnosis and checks every line.
• Dilapidations Schedule: Draft Items From Lease Clauses With AI — A worked example of mapping lease clauses to dilapidations schedule items with AI, with the RICS format explained and cost and legal checks left to people.
• Schedule of Condition Template: Draft From Photos With AI, Then Verify — A fill-in schedule of condition template, plus the steps to draft entries from site photos with AI and verify every line on site.
• RICS AI Standard for Small Firms: Register, Assessment, Client Notice — Turn the effective RICS standard into a written pre-use assessment, two practical registers and client wording that can be checked.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business-data and Data Processing Addendum pages, Anthropic privacy centre articles on commercial model training and its Data Processing Addendum, and the published text of UK GDPR principles. This tutorial is general information, not legal advice.