RICS says AI changes what comes before the signature. What should your reports show?
They should show that a named professional tested the AI-assisted work before relying on it, and how. On 19 October 2026 RICS published a practical framework for exactly that, which it calls the judgement layer. For any firm producing condition surveys, dilapidations schedules, inspection reports or technical due diligence with AI help, the test is simple: can you show what the AI was asked to do, what evidence it used, where a professional challenged it, and who signed? This article explains the paper in plain terms, the 20/60/20 split we build reporting around, and how to start.
Part 1 | The paper
What did RICS publish, and is it mandatory?
The paper is guidance; the standard behind it is mandatory. RICS published How to build a judgement layer for AI use in surveying on 19 October 2026, written by Anil Sawhney of RICS, Chris de Gruben of Artefact, James Garner of Gleeds and David Wong. It builds on the RICS professional standard on responsible use of AI, which has applied to members and regulated firms since 9 March 2026.
The standard sets the rules; the paper shows how to work to them
| RICS AI standard | Judgement layer paper | |
|---|---|---|
| Status | Mandatory since 9 March 2026 | Insight paper, guidance |
| Firm controls | Written policy, AI systems record, risk register, supplier due diligence | Connect, codify and govern |
| Clients | Told in writing, in advance, when and for what purpose AI is used | How to make judgement visible to clients and insurers |
| Professionals | Reliance on AI outputs decided by a professional and recorded in writing | Judgement at set-up, at gates and at reliance |
Introducing the paper, Anil Sawhney said AI "changes what precedes the signature, but not what the signature means." RICS also notes that contracts, insurance requirements and client expectations are already setting limits on AI use while formal regulation catches up. The paper is written for surveyors, but the same question faces any engineer, inspector or testing lab that signs a report: if this is challenged, can I show how I reached it?
What is the judgement layer?
It is how a firm organises its evidence, its know-how and its controls so AI-assisted work can be reviewed and signed off properly. RICS is clear that it is not a product you buy and switch on, and that a sole practitioner can start with one folder the AI can reach, one written rule and a note of what was checked.
Connect
The AI can reach the evidence a professional would use, and the professional can see what it relied on. Clean historic data is not a precondition.
Codify
The firm's experience is written down: templates, house wording, rates, and rules such as when a defect goes to a senior.
Govern
Limits, checks and records: what the AI may do, where it stops, who can halt the work and who signs.
Part 2 | How it works
Why is checking the final draft no longer enough?
Because an AI agent works through many steps on its own, and errors can enter long before anyone reads the draft. With a chatbot, judgement happens twice: framing the question and judging the answer. An agent can pick up the wrong evidence, carry a weak assumption forward or use a rate from the wrong type of project, and still produce something that reads well.
RICS gives a cost plan example: an agent pulls rates from a different type of project, the firm's own rule spots the mismatch, and the workflow pauses for the surveyor before anything is finalised. The paper adds a distinction worth holding on to: an automated check that has run is not professional scepticism. A system checks against criteria it was given. It does not decide, on its own, to look for what is wrong with a conclusion that looks coherent.
What is the 20/60/20 rule?
It is the baseline we start every reporting build from, and it follows the RICS paper. Your team inspects, gathers the evidence and sets the job up. The platform does the admin in the middle. Your expert checks, decides and signs. Wherever the platform reaches an agreed gate, it stops and hands the job back to a person.
The numbers are a rule of thumb. On some reports the site work is most of the job; on others the assembly is. What does not move is the shape: a person at the start, at every gate and at the end.
Which parts of a report can AI take on, and which stay human?
The clerical work in the middle, and none of the judgement. In report-heavy firms most of the hours go on assembly: matching photos to elevations, retyping handwritten readings, numbering items and fighting the template. That is what to automate.
The platform takes the admin; the professional keeps every judgement
| The platform does the admin | The professional keeps the judgement |
|---|---|
| Sorts and numbers photos and matches them to the job, room or asset | Decides what is a defect and what is not |
| Turns handwritten sheets, dictation and notes into text, flagging anything unclear | Judges severity, urgency and likely cause |
| Drafts observations in your house wording, from your own past reports | Corrects, rewrites or rejects every draft observation |
| Fills your template: numbering, cross-references, photo schedule, appendices | Forms the opinion and the recommendations |
| Prices items from your own rate library, citing where each rate came from | Decides the quantities, rates and remedies that apply |
| Flags gaps: an item with no photo, a missing reading, an unmatched rate | Decides whether the evidence is enough, or whether to go back |
| Keeps the record of sources, edits and approvals | Signs and issues the report |
Nothing is issued automatically. Every report goes out under a named person's signature, exactly as it does today. Review can happen on screen or in your own Word or Excel file, which is how our builds so far work. For where the time actually goes across a whole report, see what automated reporting actually saves.
What should the record of each AI-assisted report show?
Two things, kept apart. The system trail records what the AI did: sources used, steps taken, assumptions carried forward and checks run. The professional decision records what the named person relied on, where they challenged the output, and whether they accepted, changed or rejected it. The trail supports the decision; it cannot replace it.
RICS warns against turning this into paperwork. A useful record answers three questions: what was the system asked to do, what did the professional rely on, and what made them accept, change or reject the result? A box that only says "approved" is a form, not a record. Our AI register article covers keeping that record at firm level.
What does this look like in a real build?
Two of our builds show the pattern, with client names withheld. Both were tested on real past work before anything was relied on.
Testing lab worksheets
AI reads the handwritten values. Unclear digits are settled from the sheet's own arithmetic, never guessed, and shown to the engineer as resolved that way. Only the input cells of the lab's own spreadsheet are filled, the spreadsheet calculates, and an independent calculation checks it before an engineer approves.
Schedule of dilapidations
Drafted from the surveyor's walk-round recording into the required workbook, keeping its macros. Items the AI could not resolve are highlighted to listen back to or check against a photo. Every rate cites its library entry or past schedule, and a separate check confirms quantity times rate on every line.
On a typical dilapidations job the split runs like this. The surveyor walks the unit, dictating and photographing as normal, and uploads the files (the first 20). The platform transcribes, sorts and numbers photos by room, drafts each item in house wording and prices it from the rate library, flagging anything it cannot match (the 60). The surveyor deals with the flagged items first, decides breach and remedy, adjusts and signs (the last 20).
Can machine learning take over more of the work?
Sometimes it can widen the middle, and it never removes the ends. As the platform matches new items against your own reviewed past reports, drafts need fewer corrections, and once enough reviewed jobs exist, confidence thresholds can be set from measured accuracy on your work. Unclear readings are always flagged, however much history there is.
The set-up, every gate and the sign-off stay with people. RICS allows high-volume work to be reviewed at the level of an approved method backed by random dip sampling, but we do not build for that: every report still goes to a named person. The platform matches against your reviewed library, and no model is trained on your material, ours or anyone else's.
Part 2 | Buying
What should you ask an AI reporting supplier?
Whether its defaults hold up, because RICS treats a supplier's built-in settings as choices to test, not safeguards to assume. The supplier has not seen your client's instruction and does not sign your report. RICS also mentions ISO/IEC 42001, the AI management system standard, as something to ask about. These are the questions we would put to anyone, including us.
Seven questions for any supplier, with our answers
| Question | How AI Metric answers |
|---|---|
| Can each drafted item be traced to its source? | Each observation, reading and rate links to its photo, recording, sheet or library entry. Standard clauses come from your template |
| Is AI-drafted text kept apart from approved text? | The platform records who drafted, changed and approved each report |
| Can we set where the work stops for a person? | Yes. Gates are agreed with you at set-up, and anything unclear is flagged, not guessed |
| Can we stop the work ourselves, straight away? | Yes. You grant and withdraw our access through your own administration |
| Can we export the full record if a report is challenged? | The record is yours, with your data, templates and reports |
| Are totals calculated by a spreadsheet or code? | Yes. The AI drafts; your spreadsheet or code calculates, checked independently |
| Is our data kept out of model training, and do we know where it is processed? | No model is trained on it. We name the AI provider, where it runs and the transfer safeguard. We hold no ISO/IEC 42001 or other certification |
Our trust centre sets out ownership, access and data handling in full.
Part 3 | Where to start
Where should a firm start?
With one recurring report type, which is also what RICS recommends. Map the evidence a professional normally uses and check what the AI can reach. Write down one or two recurring decisions, such as when a defect is escalated, and set where the system must stop. Then test on real past jobs and compare with what was actually issued.
That is how our first step works: a £1,500 discovery and AI implementation workshop on one workflow, measured, with a fixed-price quote, taken off the build if you go ahead. Base builds start at £2,950, a platform in your name for one report type, live on real work in about six weeks. Hosting and AI usage run in accounts in your name, paid at cost, and upkeep is a fixed monthly fee quoted with the build. The report types we draft are on our engineering firms page, and the build itself is AI document automation.
How do firms protect how junior staff learn?
By making the reasoning visible. Gathering evidence and drafting a first answer is how junior surveyors and engineers learn, and RICS warns that if AI removes it and juniors only see finished drafts, firms gain speed now and lose capability later. Juniors should help frame the task, inspect the sources and question the assumptions before a senior signs, and the record of what the senior changed can double as teaching material.
Part 3 | The limits
What does this article not cover?
Legal or regulatory advice. The judgement layer paper is an insight paper, not a new mandatory standard; the mandatory requirements remain those in the RICS standard. This article summarises both as published on 19 October 2026. Read the standard itself and take your own advice on compliance.
If you want to test AI on one of your reports properly, book a 30 minute call and bring one you have already issued. We will mark which parts are assembly and which are judgement.
Sources
Which sources is this article built on?
Every statement about RICS above resolves to one of these, each read at source before publication.
- RICS, Applying professional judgement in AI-assisted surveying, 19 October 2026
- RICS, How to build a judgement layer for AI use in surveying, insight paper, October 2026
- RICS, Responsible use of artificial intelligence in surveying practice, professional standard, effective 9 March 2026
- CMS, RICS introduces mandatory AI standard for surveyors, March 2026
Photographs by Sven Mieke and Tsuyoshi Kozu on Unsplash, used under the Unsplash Licence.