For a term cooked up during a 1950s Dartmouth College summer research project, primarily as way to market computer scientists’ work to funders, ‘artificial intelligence’ has come a long way. There’s barely an electronic device or application that you can use today without being offered the chance to enhance your experience with AI.
It’s for each user to determine if these constant offers of help are any more useful than 1990s era Windows’ digital helper Clippy. This author has seen some benefits, and many distractions and frustrations. The production of this article, for example, was supported in part by AI: not in its conception, structure or choice of words, but by cutting the time spent on producing a reasonably accurate transcription of interviewees’ remarks, which were then manually checked before being included. All of us will have from time to time used AI tools to read text written in languages we do not know, or even to hold a conversation while travelling.
It’s the linguistic uses of AI that have proved transformative. Aptly, the key development here are transformers. These were developed at Google, as the company built its translation tools. While the computer science behind them is complex, the key point to understand is that they allow for faster pattern matching than previous systems. That innovation allowed companies like OpenAI to develop systems like ChatGPT, which produces text that matches what we would expect to see from a thinking speaker.
There’s a real ‘wow’ factor to engaging in apparent conversation with a machine. But it’s important to remember that these systems are not doing the same thing as we do when we talk. They are not searching their memory for facts or established assumptions, building their statements step-by-step based on reason. Rather, they are scanning vast libraries of text and producing an imitation of a human response.
Winning the imitation game does not convey actual knowledge. That can become a real problem in the construction and engineering sector, where quality and safety is assured with the sign-off by a human expert. If you end up in court over a decision made using AI, it will be no defence to say ‘the computer told us to do it’. Someone – a flesh and blood, thinking human being – must take responsibility and check that any decision can be backed up by evidence and reasoning.
Those vast troves of data pose another risk in a highly litigious sector: data outside your control may be used against you in court. That was a key source of transatlantic contention in the 2000s, as the largely German writers of the European mobile crane standard, EN13000, mandated the use of data loggers. For buyers of these cranes in the US, where the courts are a frequent aspect of business, the idea that litigants may be able to access complete data on the history of a machine was horrifying.
Today, as the world’s most AI-aligned car company regularly and selectively posts customer data online within days of an accident, it’s worth considering how any data you share is stored and used, and the commitments any third parties have made on their response to court requests for its release.
Reliance on third parties poses another challenge for buyers of AI-based services: with the costs of AI processing for end users currently rising and many services moving from fixed-cost subscriptions to token-based pricing, some tools may become much more expensive while others, offered for free as part of a software package, could be withdrawn or made a paid add-on.
And, while the data centre construction business currently appears to be booming, it’s worth considering whether there is truly a big enough market to support this investment. Without a real revolution in how we work (with AI replacing humans and delivering hundreds of billions of dollars of real revenue) many of these investments may, in two or three years, be merely unpowered sheds full of burnt-out GPUs.
But none of these causes for concern are reason enough to ignore AI. From site selection and surveying, through design and estimating, to machine control and collision prevention, AI-based tools may offer real advantages. While some of these tools do store data out of your control, many commit to tight regulation on its use and sharing, while others offer local processing on a computing system you own. None of those we looked at closely offer to replace human decision-making and responsibility.
We’ll follow these tools through the construction lifecycle, seeing what they offer from site selection through to construction and ongoing maintenance.
Every project starts with finding a site. New York-based James Stirrat-Ellis co-founded Build.inc – a developer of autonomous, or ‘agentic’, AI systems for the built environment – inspired by his twin interests in architecture and coding. The ambitious start-up aims to replace staff-intensive professional services firms with a more slimline approach, making heavy use of AI agents. So far, the company’s key wins have been in site selection, primarily for the very data centres that aim to power the AI revolution.
“We combine real estate, private equity folks with architects and designers, with technologists, with development consultants,” Stirrat-Ellis says.
“We do a lot of site selection work today, specifically within infrastructure. The timelines are compressing and finding powered land for data centres, for example, is increasingly difficult.” That data on land is fragmented, he points out, and this presents a rich environment to demonstrate the benefits of AI: the system can work through many different parameters, such as likely planning decisions, transport links and power availability, often using unstructured data in multiple languages to determine the best site for a project.
One of the key services any traditional architecture, engineering or construction professional offers is their signature, on a design or project plan, taking responsibility for its accuracy and regulatory compliance. Stirrat-Ellis says that Build.inc will still offer this. “We’re not just generating this information or generating these outputs and then giving them to the customer. We have our internal domain team, who’ve worked on gigawatts of data centre projects. Before anything leaves Build.inc, that goes through the human verification of an industry professional. When we issue an output to a customer, that is signed by a professional.”
Manchester-based tech firm ANS describes itself as a ‘digital transformation’ specialist, using cloud and digital services to improve how clients work. The company recently worked with Northumbrian Water as the utility moved from manual coastal surveys to the use of drones, collecting readings including pH levels, location and temperature.
While the drones could cut the time spent conducting surveys, they also created a mass of data: roughly 10,000 samples per year, per site. The challenge was to make this capable of being streamed, stored, and analysed, potentially using AI tools. ANS worked with the client to connect each drone sensor to Microsoft’s Fabric data platform, making it immediately usable for reporting and decision-making, and available for future AI use cases.
Making project data usable is key to the success of AI systems, and to more traditional automations. At collaboration specialists Revizto, that is core to a recent update. Marc Schuetz joined the company last August as chief product officer, from Bentley. While the company is not using AI directly in its own collaboration platform, it has implemented the Model Context Protocol (MCP), developed by Anthropic, to allow customers to connect the AI tools they find useful to the platform.
This allows customers to ask questions of the platform – What are the most critical issues in this project? What issues haven’t been resolved in a week? – using simple prompts, and to create dashboards based on them. “We’re doing a lot of work right now on how to build the best tool for specific personas and how to make sure they are adopted in the fastest way,” explains Schuetz. “The single biggest barrier of adoption of new technology is time.”
The adoption of MCP aims to address this, allowing Revizto users, their clients and peers to connect the tool to third party services and to control them with chat-like prompts. By focusing on personas, like an architect, project manager, or structural engineer, Revizto can create sample prompts relevant to their work as a starting point, easing adoption.
Dutch consulting engineer Arcadis recently formed a partnership with AI platform provider Nomic and will use Nomic’s tools for a broad range of activities, from permitting and planning to major infrastructure delivery. Nomic’s AI agents will help teams accelerate delivery work such as drawing reviews against firm standards, code compliance checks, first-pass submittal review, RFI research and BIM coordination.
US technology firm Trimble has built its business on the interface of the digital and physical, with its scanning systems being used for tasks like adding as-built data to BIM models. For the mechanical, electrical and plumbing sector, some of its most useful tools come at the estimating stage.
Brooke Stewart is Trimble’s director of product management for its estimating and take off solutions – and while her focus is now on construction technology, she started her career as an estimator in the electrical trades.
Stewart perhaps summed up best what makes AI different from other systems: “I tend to be very puristic in my view of AI. We’re very explicit when we’re leveraging AI. Whereas if we’re not actually leveraging it, I will refer to it as an automation.” AI is perhaps best understood for what it isn’t: it isn’t a series of automated steps.
A key step for any estimator is takeoff. “Takeoff is the systematic process of counting and measuring everything that you see on construction drawings that make up the M&E systems,” explains Stewart. “And that generates quantities for a contractor. So quantities of things that they see on these construction drawings. Those quantities get applied to these rule-based kits. The rule-based kits translate those quantities into real-world material parts that get purchased and installed.”
Accurately pricing the materials needed is key to a successful estimate: getting it wrong could get your pitch ruled out on a price basis, or it could see you stuck with costs that leave no space for a fair margin.
One fear sparked by AI is that tools like this will replace the work of junior staff, cutting important rungs out of the bottom of the career ladder. But, ironically, so important is the careful manual work of takeoff to successfully winning profitable contracts that it is often performed by the most experienced and skilled staff.
Trimble realised that AI could be used to work through the 2D drawings still supplied by designers and tier-one contractors, counting switches and measures conduits. But general-purpose AI tools weren’t up to the job.
“People have declared that they can use Claude or Gemini, give it construction drawings and it can go and find objects. The reality is that the accuracy has been measured at sub-30%, because these general large language models are just not trained in the domain and data needed.”
Trimble has used its decades of expertise to design an AI tool, Live Count, that will tackle two very specific tasks. “First, they remove the manual step of setting things up, the steps that precede takeoff. Very specifically, it’s the setting of the drawing scale and the setting of the drawing name. If you’re doing that thousands of times, that time adds up,” Stewart explains.
“Second, and most impactfully, we’ve been able to train the AI models to automatically count and measure objects on a drawing.”
The tools are designed around human review. “The AI goes and figures out the counts and categorises them. Then it shows its work to the estimator. It leaves the estimator fully in control of the output. The AI is just doing the heavy lifting on the first pass.”
When Trimble put the tools through an intensive beta testing programme with clients, a common source of feedback was a request for a grid view of the results. This was implemented with a confidence indicator added.
The beta programme allowed Trimble to measure the time saving made by using the tools and the real-world testers saw a 62% time saving.
The tool has been designed around specific regions, incorporating relevant domain knowledge and hosted on local data centres ensuring regulatory and standards compliance. It is now available in both Trimble’s US package Accubid, and its UK counterpart, Total Estimating.
As work progresses, whether on new build or maintenance projects, crews in the field can use Cooper, a natural language assistant from US software firm Simpro. A component of Simpro’s Lightning platform, Cooper ties together design functions like take-off with customer relationship management and invoicing, along with tools for teams in the field. A product like this works best when it has access to large amounts of company data.
Raffi Tchakmakjian, chief product and business development officer, says that Simpro ensures customers are not exposing themselves to new risks when sharing this information. “Where a request is processed by a third-party model provider, it is under terms that prohibit that provider from training on customer data or retaining it beyond the request.”
“We do not hand over customer data on request. We respond only to legally valid, binding orders, but have not had such a request arise.”
As Simpro makes use of third-party platforms, it has made sure that any changes in compute costs are not passed on to clients. “Pricing is contractually capped at a maximum CPI + 3%, for customers on long-term agreements. If AI infrastructure costs double again next year, that is our exposure, not our customers.”
Tchakmakjian explains that part of the role of an AI-based service provider is to pick the best compute provider, changing this behind the scenes if needed. “We are also looking at our own proprietary models for certain workflows where the token cost is completely is under our control. The real value to our customers is our intelligence and context, and if a certain model no longer fits the parameters of quality or cost, we make the decision to re-route to other models or to our own models if it fits the workflow.”
Tools like this demonstrate the potential of AI in the office-based development and management of projects. But the technology also has applications on site. Jonathan Guest formed Safety Shield Global after losing someone close to him in a site accident. The company uses cameras and AI-enabled human form recognition to identify people coming close to working machines. It is increasingly widely installed on site equipment, and included as a factory option by many manufacturers, including Komatsu.
The system offers real improvements over other forms of human detection. Jeremy Meredith, Safety Shield’s director of public affairs, says that the system offers 99.6% accuracy and a false positive rate of around 0.05%: a key concern on busy UK sites, where accidents must be avoided, without frequent unnecessary stoppages.
The AI detection happens on the machine itself, using ‘edge computing’ whereby data is processed close to where it is generated, in this case using a small but powerful embedded computer. It offers much faster detection than is possible with chip-based systems, such as doorbells, or anything cloud-based. As well as boosting accuracy and speed, it also eliminates privacy concerns. The camera data is stored only for a few seconds to allow for processing, with options to upload when a risk occurs. It is owned by the client and, when shared to the Safety Shield Global cloud, stored on hardware in the same country.
For clients who do elect to upload incident or risk data, Safety Shield has launched a new AI-based product called Nexus that can quickly work through risks, categorising and prioritising, both to deliver individual alerts and to allow health and safety officers working across sites to identify broader issues.
Alex Oakley, senior digital lead at Safety Shield, led the development of Nexus. “Construction has largely been driven by lagging indicators,” he says. “Accident reports, near misses, investigations after something’s already happened. What AI does is change that from being a lagging indicator to being a leading one. It provides predictive analysis, gives you the information up-front so you can see the full picture on site.”
Meredith adds, “Every machine, camera, sensor, digital system is producing information continuously. But no safety manager or project director can realistically process thousands of events each day.
“Nexus is for both the safety officer on site and the health and safety director who’s maybe in charge of hundreds of machines across several projects. This gives insightful data in real time so that site management teams can go out immediately to review a situation where a trend is identified.
“It might be that people require training about what they should be doing in and around machinery. Site management teams might establish a need for greater supervision or realise that a safe system of work needs changes.”
All AI systems are more useful the more data they have access to. Meredith says that Safety Shield, and human form recognition more broadly, has been mandated by tier-one contractors on many sites, and on critical projects like Sizewell C, Hinkley Point C and the Lower Thames Crossing.
An obstacle to more widespread mandates, and better insights, is a lack of standards. “There isn’t a standard that we would consider the correct – or any – benchmark for systems that are accurate enough to deliver assured human form recognition,” says Meredith, who is working with the Construction Plant-hire Association to come up with a standard specification for the adoption of the systems.
Vision systems, while not always considered AI, are closely related in terms of their pattern-matching powers. One day, they might be powerful enough to allow for the development of cobots, or collaborative robots, able to work safely alongside humans on site.
Other computer vision systems are, in the right conditions, suitable for use driving vehicles. US equipment manufacturer John Deere has developed an advanced vision processing unit (VPU) for its agricultural equipment in the US, implementing it first on its See & Spray technology.
In 2025, the company developed the Autonomous 460P articulated dump truck, able to travel autonomously around a quarry. The VPU has now been made available to European machine builders and OEMs for use in the development of their own autonomous equipment. In the next few years we may see computer vision systems able to handle conditions on construction sites, with far more obstacles and risks than a cornfield or quarry.
Once work is completed on infrastructure projects – and for decades afterwards – AI-enhanced computer vision systems have another role to play: in monitoring ongoing maintenance needs. This July, KaarbonTech, a software company based in Bournemouth, launched Smart Shot AI, a new monitoring capability that combines regular image capture with AI review to identify visible change and emerging risk. It gives operational teams a clearer picture of how assets and locations behave over time, without replacing professional judgement or established inspection regimes.
The capability can show whether debris repeatedly returns after clearance, whether water continues to rise, whether an obstruction worsens or whether an issue naturally clears. This creates a more complete evidence base for deciding when a site visit is needed, where maintenance should be targeted and how an asset is performing across its operational lifecycle.
The company says that solar-powered, self-running units can be deployed at locations where additional visibility would support decision-making, including culverts, trash screens, watercourses, high-risk junctions and other hard-to-reach or rapidly changing sites. Configurable image capture, night vision, secure upload and AI-led review allow teams to assess change through a central dashboard and focus attention where it is most needed.
Got a story? Email news@theconstructionindex.co.uk
