The RICS guidance on the use of AI came into effect in March 2026. Since then there has been a lot of discussion around the use of AI within surveying practices, particularly around privacy and the security of client information.
In reality, the issue is not AI itself.
The issue is where the language model is located, how it is accessed, what happens to the data sent to it and whether that data remains within the control of the surveying practice.
To understand where we are going, it is worth looking at what has happened in the past when new technology has been introduced.
When I started as a graduate surveyor, drawings were produced by hand and sent to local reprographics companies for copying. Photographs were taken on rolls of film and sent away to laboratories for processing.
At that point, client information was regularly leaving the office environment and entering a third-party environment.
Usually that movement of data happened at a local level and involved people physically handling the information.
Plans for local developments were seen by local people working in local businesses who, if they wished, could easily have discussed what they had seen outside of work.
The industry dealt with that risk through confidentiality agreements, employment contracts and professional standards and, perhaps most importantly of all, trust.
The information left the office, was processed elsewhere and then returned.
As surveying moved into the digital age, the process remained largely the same even if the format changed.
Paper drawings became CAD files.
Photographs became digital image files.
Documents became PDFs and spreadsheets.
The data still left the office and travelled to third parties for processing, but now it existed on the sender’s hardware, the recipient’s hardware and in transit between the two.
That introduced new risks that did not exist previously. If a drawing went missing between the office and the reprographics company, it was generally quite easy to work out what had happened to it.
Over time, many firms reduced those risks by bringing parts of that process in-house. They printed their own drawings, printed their own photographs and invested in their own hardware and software.
The data no longer had to leave the building.
Then cloud computing arrived.
Rather than storing information on a local server, files were stored in a remote location and accessed over the internet. Information was once again leaving the office environment, albeit in a different form.
Traditionally, cloud computing within surveying practices has largely been concerned with storage rather than processing.
Storage brings its own advantages.
Data can be replicated across multiple data centres, often in different geographical locations, ensuring that if one facility suffers an outage, the information remains available elsewhere.
AI processing introduces a slightly different challenge.
When the processing itself is taking place remotely, simply replicating data across multiple locations does not necessarily provide the same level of resilience. The processing capacity also needs to exist elsewhere and be capable of taking over the workload.
Recent events have shown that this is not purely a theoretical issue.
The disruption to data centre infrastructure in the Gulf during the conflict involving Iran in March 2026 demonstrated how businesses can be affected when a heavily relied upon data centre or region suffers an outage.
There is also the issue of power consumption.
The data centres undertaking AI processing require enormous amounts of electricity and cooling infrastructure to support the hardware operating within them.
When all of these factors are considered together- security, resilience, geopolitical risk, energy demand and infrastructure dependency- there is a reasonable argument for moving towards locally hosted, business-specific language models operating within the firm’s own infrastructure.
With storage, the challenge was largely where to keep the information.
With processing come additional considerations.
At this stage you may be wondering why I have spent so long talking about reprographics companies, film processing and cloud storage.
The reason is quite simple.
The use of language models today is not much different from those earlier processes, as many people think.
- Data is sent away in one format.
- Processing takes place elsewhere.
- The result is returned in another format.
That is exactly how many AI systems operate today.
You provide the input, the language model processes it and returns something different from what you originally supplied.
Understanding that process is crucial to understanding how AI will eventually be integrated into surveying firms.
At present, most language models used by businesses are hosted within large data centres somewhere in the world.
When a surveyor enters a prompt, uploads a report or submits photographs for analysis, that information is sent to the language model for processing.
Depending on the provider and the service being used, some of that information may be stored for logging, security monitoring or future model improvement.
This is where many of the concerns around confidentiality arise.
Historically, some AI providers have used user input and output data as part of future model training.
That does not mean your report suddenly appears in somebody else’s answer, but it does mean firms need to understand exactly what happens to their information once it leaves their control.
As this has become more of an issue in business, some providers allow users to opt out of this process.
Providers of larger business and enterprise services often disable future training by default.
It is crucial that surveying firms understand the difference between the various types of service being offered.
Earlier I mentioned how reprographics and printing gradually moved back in-house.
I suspect we will see exactly the same thing happen with language models.
Not long ago, running a language model locally required specialist knowledge and expensive hardware.
That is changing rapidly.
Today it is possible to run capable language models on hardware that would have been considered completely unsuitable only a few years ago.
Going forward, I suspect we will see firms using multiple language models, each trained or configured for a very specific purpose.
One model may deal with report writing.
Another may deal with building regulations.
Another may deal with defect diagnosis or planned maintenance advice.
The surveyor may never know which model is being consulted in the background.
The software will simply decide which specialist model is most appropriate for the task and return the result.
The larger surveying firms with dedicated IT departments will probably be the first to adopt locally hosted language models operating entirely within their existing infrastructure.
For them, the attraction is obvious. The data never leaves the business.
Managed service providers will then begin offering hosted private language models for firms that do not wish to maintain their own systems.
Smaller practices will almost certainly adopt simpler plug-and-play solutions that can operate on local hardware without requiring specialist knowledge.
The technology itself will become less important than where it sits and who controls it.
Ultimately, I do not believe the long-term future for surveying firms is sending confidential information to a publicly accessible language model hosted somewhere else in the world.
I suspect the future for many surveying firms is local models, trained for specific tasks, operating within the firm’s own infrastructure and under the firm’s own control.
In many ways, we may simply be returning to how many previous technological advances were integrated: the processing stays close to the business.
Only this time it is not the person undertaking the processing who is sitting in a reprographics office or a photographic laboratory.
It is a piece of hardware loaded with specialist software sitting quietly in a server rack.