AI is now part of the ordinary surveying workflow, transcribing site notes, drafting report sections, running valuation models. The RICS AI Professional Standard is clear about what doesn’t change when that happens: the surveyor stays accountable for every conclusion that reaches a client. Not the software. Not the model. You.
That principle has a name: human in the loop. It isn’t a compliance slogan, it’s a discipline with three distinct stages, and getting any one of them wrong is what turns AI-assisted efficiency into a liability.
What Does “Human in the Loop” Actually Mean?
Human in the loop means a qualified surveyor stays in control of every judgement, conclusion, and output that an AI tool touches before it reaches a client. AI can gather, draft, and organise, interpretation and sign-off stay with the person carrying the professional liability.
James Garner FRICS, Head of AI and Data at Gleeds, frames it as bookends: you take responsibility for the brief you give the AI, the AI does the work in the middle, and you take responsibility again for the output before it goes out. The AI never holds either end.
Why Doesn’t Professional Responsibility Transfer to the AI?
Because no AI tool can be held accountable, not by a client, an insurer, or RICS. Only the surveyor who signs the report can be. The RICS standard builds its five principles, accountability, transparency, professional judgement, data reliability, and record keeping, around that single fact.
This is exactly where things go wrong when a claim is made. When that reasoning was never captured clearly in the first place, revisiting it becomes guesswork.
What Are the Three Stages of Human Oversight?
Applying human in the loop means being deliberate at three points: before the AI works, while it works, and after it works. Miss any one of the three and the principle stops holding.
Before: briefing and context. The output is only as good as the brief. The best brief you can give an AI tool is the same one experienced surveyors already give themselves on site. That’s the input the whole loop depends on, if it isn’t captured at the point of inspection, no amount of review afterwards recovers it. At this stage, decide whether AI belongs on the task at all, give it accurate and complete input, and never enter client-identifying data into an unsecured tool.
During: knowing the limits. You don’t need to understand the model, you need to understand where it fails. Outputs are probabilistic, so the same prompt won’t produce the same answer twice. Generative tools can hallucinate confidently, including invented regulation references. And the tool doesn’t know your client, your building, or your local market. You do.
After: review, verify, take ownership. This is the stage the RICS standard cares about most, and it’s the one that gets skimmed under time pressure. Tim Kenny, a residential surveyor, describes writing up as inseparable from the reasoning itself: “Writing up for me is very much part of the thinking process.” That’s the point, review isn’t proofreading, it’s re-forming the judgement as your own. Check every factual claim, verify technical descriptions against what you actually observed, and be ready to explain any part of the output if a client asks. If you can’t explain it, don’t sign it.
How Do You Keep Client Data Secure When Using AI?
You keep it secure by knowing exactly where it goes before you use the tool, not after. A free-tier generative AI platform may train on what you enter, meaning client information could contribute to outputs other users see. A closed, secured environment doesn’t carry that risk, but you have to check, not assume.
Before entering any client data into an AI tool, you should be able to answer: where is it stored and for how long, is the environment closed, does the privacy policy actually permit your use, and are you acting as data controller or processor under UK GDPR in that context. If you can’t answer those, you’re not in enough control to satisfy the standard.
What Should an AI Audit Trail Actually Record?
A proper audit trail records which tool was used, what went into it, what it produced, how a surveyor reviewed it, what changed before use, and who ultimately signed off, every time, not just when something goes wrong. Record keeping is what turns “we followed the standard” from a claim into something you can actually demonstrate.
The more conversations I have, the clearer it becomes that this is the step everyone intends to do and almost nobody keeps up consistently, not because it’s complicated, but because a spreadsheet nobody updates isn’t a system. It’s one of the reasons Sitarva builds that log at the point of capture, on site, rather than asking you to reconstruct it at a desk weeks later. The surveyor still does the reviewing and the deciding, Sitarva just makes sure there’s something solid to review.
What Should a Small Firm Do First?
You don’t need new infrastructure to satisfy this, you need consistency. In order:
- Audit what you already use. List every AI tool in the practice, including ones embedded in other software, and flag anything touching client data in an unsecured way.
- Build a briefing habit. Treat every AI instruction like a brief to a junior colleague, clear, specific, and never containing client-identifying data if the tool isn’t secured.
- Make review a step, not a glance. No AI output goes into a client document without someone reading and verifying it line by line.
- Log every use. Tool, task, date, reviewer, and what changed, kept in the matter file, retained as long as your other professional records.
- Disclose it. A brief line in your terms of engagement stating AI may assist, subject to professional review, is enough. Never let a client believe AI-assisted work was entirely hand-written if asked directly.
- Revisit every six months. Tools change fast; your tool list and data arrangements should keep pace.
What Mistakes Most Often Undermine Human Oversight?
The same handful of failures show up repeatedly:
- Skimming instead of reviewing. Scanning for obvious errors isn’t the same as confirming every sentence is accurate.
- Letting AI form conclusions. It can help you articulate a judgement, it can’t make one for you.
- Entering client data into unsecured tools, regardless of whether anything goes wrong as a result.
- Not documenting AI use at all. The gap between what was found and what the report actually showed.
- Assuming the AI is current. Models are trained on historical data and won’t always reflect the latest regulation or case law.
The Standard Doesn’t Ask You to Avoid AI – It Asks You to Stay in Control
The RICS standard isn’t anti-AI. It’s asking surveyors to brief it properly, understand its limits, review its output rigorously, and stand behind everything that reaches a client. AI handles volume and drafting; the surveyor brings the inspection, the defensible reasoning, and the accountability. That division doesn’t change no matter how good the tooling gets.
If part of the reason this feels hard right now is that the record-keeping side has no home, no consistent way to capture what was seen, what was asked of AI, and what a surveyor checked before it went out, that’s precisely the gap Sitarva was built to close. Try Sitarva and see how the audit trail builds itself as you work, instead of at the end of it.