A surveyor signs off a Level 3 building survey. At the foot of the report sits a short paragraph stating which parts of the document AI helped draft, and asserting that the surveyor takes full responsibility for the content. The interesting detail is not the sentence. It is where the sentence comes from: it is generated from the actual record of every AI suggestion the surveyor accepted, modified or rejected during the job — not written from memory, and not pasted from a template at the end.
We built that disclosure into a bespoke surveying operating system for a UK residential surveying practice. It maps to the disclosure expectation RICS now sets for AI in surveying work. But the structure underneath it is not specific to surveying at all. It is a pattern any regulated profession can copy: solicitors, accountants, actuaries, auditors, valuers, financial advisers. This post takes that surveying disclosure apart and rebuilds it as a generic template.
Key takeaways
- A disclosure is only as good as the record behind it; "produced with the assistance of AI tools" cannot answer which parts, which model, and who checked them.
- RICS published its first standard on responsible AI use in surveying in September 2025: the member must disclose where and how AI was used and remains professionally responsible.
- Five components make a disclosure computed rather than typed — an append-only decision ledger, sticky human verification, a generated disclosure, bounded use, and graceful degradation.
- The load-bearing line is "the named human takes full responsibility"; the law does not accept "the AI did it" as a defence, so the disclosure must name and prove the reviewer.
- The transferable structure (ledger, verification, generated disclosure) generalises to any regulated profession, though each has its own regulator and wording to check.
A disclosure is only as good as the record behind it
Most AI disclosure today is a sentence bolted on at the end. "This document was produced with the assistance of AI tools." It is honest in spirit and useless in practice, because it cannot answer the only questions that matter when something goes wrong: which parts, which model, and who checked them.
RICS published its first standard on the responsible use of AI in surveying practice in September 2025. The principle it sets out is plain: a member must disclose, in a way the client can understand, where and how AI was used, and the member remains professionally responsible for the output. That is the same logic the regulators reach for everywhere AI touches a regulated judgement. The UK's pro-innovation AI principles, led by DSIT, name appropriate transparency and accountability among their five cross-cutting principles — applied by each profession's own regulator rather than by a single statute.
So the test of a disclosure is not whether the words are present. It is whether the words are true and provable. That moves the work upstream, away from the closing paragraph and into how the AI was used in the first place.
The surveying build, taken apart
The disclosure in the surveying system was computed, not typed. Five components made that possible.
An append-only decision ledger. Every time the AI suggested text — a condition rating narrative, a defect description, a summary — the system recorded the model used, the input, the output, and the surveyor's decision: accept, modify or reject, with a named surveyor and a timestamp. The table is append-only by construction. No update or delete operation exists against it, so the record cannot be quietly tidied after the fact. This is the same append-only ledger pattern we use across our builds.
Explicit, sticky human verification. A section touched by AI is flagged until a person verifies it. Once verified, it stays verified; it cannot be silently cleared by a later edit. The professional's sign-off is a deliberate act, recorded as one.
A disclosure generated from the record. The closing statement is assembled from the ledger. If the ledger shows AI-assisted sections that a named surveyor accepted or modified, the disclosure says so and asserts that the surveyor takes full responsibility for the content of the report. If the ledger shows no AI involvement, it says no AI technology was used. The sentence cannot drift from what actually happened, because it is derived from what actually happened.
Bounded use, not a free-form chatbot. AI was used at a small number of structured touch-points, each returning constrained output, never as an open assistant writing whatever it liked. Narrowing where AI can act is what makes the record legible.
Graceful degradation. If the AI step failed, the system fell back to deterministic drafting rather than blocking the surveyor. The disclosure stays accurate either way.
None of this is about the model being clever. It is about the record being honest.
The generic disclosure template
Strip out the surveying specifics and the same five elements generalise to any regulated profession. The mapping is direct.
| Component | Surveying instance | Generic requirement |
|---|---|---|
| Decision ledger | Per-suggestion accept / modify / reject, named surveyor | An append-only record of every AI contribution and the named professional's decision on it |
| Sticky verification | Section flagged until a surveyor verifies it | Explicit, recorded sign-off that cannot be silently reversed |
| Generated disclosure | RICS-style statement built from the ledger | Client-facing statement derived from the record, not written from memory |
| Bounded use | Five structured touch-points | A defined, finite set of places AI may act, each with constrained output |
| Degradation | Deterministic fallback drafting | The output remains valid and the disclosure accurate if AI is unavailable |
A defensible disclosure for any profession should be able to answer four questions, in writing, from the record:
- Where was AI used in producing this work? (Which sections, tasks or judgements.)
- How was it used — to draft, to summarise, to extract, to check? (Never to decide on its own.)
- Who is the named person who reviewed and is responsible for the output, and when did they sign off?
- What happens to the assertion of responsibility if the record shows the AI was not actually checked? (It should not be possible to assert responsibility without the verification step.)
If your disclosure cannot answer those four from a record rather than from recollection, it is a marketing sentence, not a control.
Why "the named human takes full responsibility" is the load-bearing line
The phrase that matters in the surveying disclosure is not about AI at all. It is "the surveyor takes full responsibility for the content of this report." That sentence is the whole point. AI assists; a named person decides and answers for the decision.
This is the principle every UK regulator is converging on, under different names. In financial services there is no AI-specific rulebook, but the Senior Managers and Certification Regime already attaches individual accountability to the outcomes of a function, including where AI now performs work a person used to do — and the FCA and PRA reaffirmed a technology-neutral approach on 1 April 2026. Where AI processes personal data, the ICO is the lead regulator under UK GDPR; the Data (Use and Access) Act 2025 brought new Articles 22A–22D into force on 5 February 2026, preserving the right to human review of significant automated decisions. The ICO's updated guidance on automated decision-making and profiling was still in draft at the time of writing — its consultation closed on 29 May 2026, with final guidance expected in summer 2026 — so treat its detail as not yet settled.
The common thread: the law does not accept "the AI did it" as a defence. It expects a person to own the output. A disclosure that names that person, and proves they verified the work, is how you evidence that ownership when challenged.
What this looks like as an engineering decision, not a policy
The temptation is to address disclosure with a policy document and a training session. That produces a disclosure people can forget to apply. The surveying build addressed it as architecture, which produces a disclosure that cannot be forgotten because it is computed.
The practical moves, in order:
- Constrain where AI can act. A finite set of structured touch-points, each with bounded output, not an open chatbot. You cannot record what you cannot delimit.
- Log the decision, not just the output. Capture the professional's accept / modify / reject against each AI contribution, with name and time, into a record that cannot be edited after the fact.
- Make verification an explicit act. Unverified AI content is visibly flagged; sign-off is recorded; it cannot be silently undone.
- Generate the disclosure from the log. The client-facing statement is derived from the record, so it is always true.
- Fail safe. If AI is unavailable, the work still completes deterministically and the disclosure still reflects reality.
This is the difference between governance written about a system and governance written into one. We have built it both ways and only the second kind survives an audit. Our broader approach to this — controls that hold by construction rather than by intention — is set out on our trust page, and our case studies show the same pattern across surveying, property operations, invoice automation and public-sector evidence work.
A note on scope: RICS's AI standard governs surveying, and other professions have their own regulators and their own wording. The template here is the structure — ledger, verification, generated disclosure — not a claim that one profession's standard applies to another. Check your own regulator's rules for the exact disclosure your work requires.
Last reviewed: 29 May 2026.
If you are working out what an AI disclosure should say for your profession — and how to make it provable rather than aspirational — talk to us. We build the record before we write the sentence.
Sources: RICS responsible-AI standard (September 2025) · RICS Home Survey Standard · DSIT pro-innovation AI principles · FCA approach to AI · ICO on the DUAA 2025
