The stakes remain high. In the EU, construction recorded 2,899 non-fatal workplace accidents per 100,000 workers in 2023 (Eurostat), the highest incidence rate among economic activities. It also accounted for 792 fatal accidents, or 24% of all fatal workplace accidents. Globally, the International Labour Organization estimates that more than 60,000 fatal accidents occur on construction sites every year.
Technology can intervene at several points in the prevention chain: anticipating risks, improving safety conversations, changing the way work is designed, warning people when danger approaches, and removing people from hazardous environments. But technology does not operate independently of the organisation. Workers need to understand it, trust it and participate in its implementation; supervisors need to remain accountable; and tools need to fit the realities of work on site.
The International Labour Organization’s 2025 global report on AI and digitalisation makes the same point: automation and smart technologies can reduce hazardous exposure, but new risks can emerge through excessive reliance on automated systems, human-machine interaction or intensified work. Worker participation and training are therefore essential to safe adoption.
Predicting risk: giving supervisors an augmented view of the site
Construction sites follow a familiar ritual each morning, with a supervisor briefing on the day’s activities. But the information available to that supervisor is often fragmented: today’s tasks and conditions sit alongside lessons contained in years of incidents and observations.
Balfour Beatty, Britain’s largest construction company, is working with Microsoft on a health and safety briefing agent designed to bring these sources together. Jon Ozanne, Balfour Beatty’s Chief Information Officer, describes the principle:
“We’re stitching together the activities that are happening onsite today, the external conditions, the work permits that might be in place, and we’re preparing a briefing for the supervisor.”
The AI agent can draw on what Ozanne describes as “millions of incidents and observations from the past”, helping identify situations whose risk may change with current conditions.
The system is designed to produce around 80% of the briefing, which the supervisor can then edit and complete.
“We’re not looking to automate to get to cost efficiency. We’re looking to augment humans to do their job better,” he says.
The supervisor remains accountable — the tool is intended to augment rather than replace judgement.
This principle is also visible in Oracle’s Newmetrix (formerly Smartvid.io’s Safety Suite, supported by Leonard’s CATALYST programme). Its AI engine, Vinnie, analyses construction images and video to identify potential safety risks. The project explored how predictive modules could be integrated into VINCI’s systems and used by operations and risk teams.
Capturing the conversation: making safety communication part of the data
If AI can help determine what should be discussed, another question is what happens during the safety conversation itself.
Australian platform RiskTalk replaces written risk assessments and paper-based pre-start or toolbox talks with a recorded group conversation. Stuart Farquharson, its founder, describes the purpose simply:
“This is the work we are doing today. This is how we want it done, and this is how it needs to be done safely.”
The team discusses the work, the hazards and the controls from a single device. RiskTalk then converts the recording into a structured record and an AI “coach” checks it against the company’s procedures and critical risks.
The approach treats the toolbox talk as more than paperwork: it remains a moment for a team to share its understanding of the work and identify what has changed. RiskTalk also explores voice analysis for signs of engagement, fatigue, stress and psychosocial hazards.
Designing risk out: moving prevention upstream
The most effective safety intervention may be to avoid creating a hazardous task in the first place.
That is the logic behind WSP’s work on modern methods of construction. Dale Sinclair, WSP’s UK Head of Digital Innovation and Global Lead for Modern Methods of Construction, argues for moving more work away from the construction site and into controlled environments.
“Our strategic goal is to move 70% of the work away from site,” he says.
On projects including Manchester Airport’s Pier Two, WSP has developed a “kit of parts” approach that reduced the number of components from around 3,500 to 67.
The company is also using 4D digital rehearsal and observing factory assembly operations to identify where design and procedures can be changed to make work safer.
Sinclair sees this as a question of where safety knowledge sits in the project lifecycle:
“Tools that automate things downstream [on operational sites] are absolutely fantastic, but why don’t we then have that knowledge upstream?”
If a risk can be identified during design, it may be possible to eliminate it before a worker ever encounters it. That requires designers, engineers, safety professionals and site teams to exchange knowledge early enough for it to influence decisions, with WSP using 4D digital rehearsals at concept design.
Addressing risks that are harder to see: heat, movement and exposure
Not every safety risk is an immediate collision or a clearly defined hazardous task. Some develop through exposure: heat stress, awkward movements, repetitive effort or cumulative physical strain.
Evalan’s Armor, for example, is designed to monitor heat stress in physically demanding work. It combines heart-rate data with an algorithm estimating core body temperature and a Physiological Strain Index, providing information to a supervisor when heat strain becomes too high. The company lists construction among its intended applications.
A similar approach is being applied to musculoskeletal risk. SoterCoach uses a wearable movement sensor to detect high-risk movements and provide immediate feedback through vibration. Its platform is intended to identify ergonomic risk in activities such as bending, carrying and overhead work, while also giving supervisors data on recurring exposure.
These systems extend digital prevention beyond the next accident to risks that can progressively damage health. But an alert only becomes prevention if someone can act on it: pause, adapt the task, provide recovery time or change the organisation of the work.
Removing people from the hazard: designing for robotics
Sometimes the safest task is the one a person does not have to perform.
At Sonomatic, engineering decisions are being made at the front end of industrial projects to allow robotic inspection of pressure vessels and other assets. Sean Peters, Project Manager for Europe and Africa, starts with the safety case for removing people from confined spaces.
“Assets are typically designed for people to go in rather than for complex robotics to go inside,” he says.
Designing access for robots can reduce the need for scaffolding and confined-space entry during maintenance.
The same principle is being explored within VINCI. Robots for Site, a collaboration between VINCI Energies, VINCI Construction and Eurovia has developed robotic solutions for tasks including drilling, sanding and material handling. The objective is not only to reduce accidents but also exposure to repetitive, demanding tasks and associated illnesses and injuries.
Another Leonard-supported project illustrates the potential of robotics for inspection. Flyability‘s Elios 3 drone has been deployed by Dodin Campenon Bernard, a VINCI Construction company, at the Avrieux shafts site of the Euralpin Lyon Turin Tunnel (TELT). The drone is used to inspect areas at depths of around 350–400 metres where geological instability made it impossible to send staff safely.
But robotics also changes the organisation of work. Workers may no longer perform the hazardous task itself, but they may have to supervise, maintain or work alongside machines. The ILO’s 2025 report highlights risks around human-robot interaction, system failures, cybersecurity and over-reliance on automated systems.
There is also another way of reducing exposure: rather than removing people from the site, technology can warn them when they are entering a dangerous zone.
Sharing Worker’s Flash Protec’ (Leonard’s 2026 SEED program) is designed around this principle. A beacon mounted on an item of equipment creates a configurable safety zone; when a worker wearing the company’s wristband enters that zone, the bracelet vibrates while the beacon flashes, warning both the person on foot and the equipment operator. The system is designed to operate without Wi-Fi or 4G and without modifying the machine, including in environments such as tunnels.
From technology to adoption: the organisational challenge
These approaches suggest a common trajectory: AI can help predict; digital tools can help understand; wearables can make invisible exposure more visible; design can eliminate; robotics can remove exposure; and real-time alerts can help people react before a situation becomes an accident.
But none of these technologies is inherently safe simply because it is digital.
There is also a risk in this trajectory: technological safety can become too easily synonymous with removing people from the equation. Workers may become recipients of safety decisions rather than participants in making work safe, while supervisors and managers may rely increasingly on systems that identify risks instead of engaging directly with how work is actually done.
This matters because safety is not only a question of knowledge, knowing where the risks are or having the technical means to control them. It is also a question of will: the collective willingness to stop, challenge an unsafe situation, adapt a procedure, speak up and take responsibility for one another.
Recent international research shows that adoption itself is a major challenge. A 2025 ASSP survey of construction professionals found that 36% cited resistance to change among staff as a barrier to adopting new safety technologies, while 31% referenced difficulties integrating new technologies with existing systems. More than half of respondents were only somewhat or not at all confident that the technologies they currently use are supported by strong evidence that they improve compliance and reduce risk.
RICS’s 2025 global research similarly found that AI adoption in construction remained limited: 45% of respondents reported no AI implementation in their organisations, while another 34% were still at the pilot stage.
The issue is therefore not simply whether a tool works technically. It is whether it functions within a system of work.
Who decides when an AI recommendation should be followed? How is accountability maintained? How are workers trained? How are alerts prevented from becoming background noise? And how can workers have a say in technologies that monitor, assess or change the way they work?
IOSH’s 2026 position is particularly explicit on this point. Ruth Wilkinson, its Head of Policy and Public Affairs, argues that AI needs to be people-centred, with worker participation, responsible governance and appropriate training built into implementation from the outset.
This brings the innovation pathways together. The future of construction safety may not depend on a single breakthrough technology, but on a continuous prevention loop: lessons from the site informing the next briefing; safety conversations feeding organisational learning; that knowledge influencing design; technology making emerging exposures visible; and design or robotics making it possible to remove people from hazardous work.
Ultimately, safety remains more than a problem of knowledge or technical capability. It is also a collective choice to act on what we know — to stop, adapt, challenge and care for one another. Technology can make that choice easier, more informed and more effective; it cannot make it on our behalf.
Ludovic Demierre’s perspective (VINCI’s Vice President of Human Ressources)
Our Health & Safety ambition is quite clear: strive for zero accidents, together, every day, and wherever we operate.
To achieve this, we rely on three pillars: exemplarity, transparency, and dialogue.
AI must be used to serve this ambition. It can be a powerfull prevention tool, helping us to better anticipate risks, highlight previously overlooked situations, and sometimes prevent men and women from being exposed to dangerous tasks.
However, the introduction of AI can also pose risks to the health of our employees, physically, cognitively, organizationally, and in terms of mental health. These risks must be assessed with the same rigor as other occupational risks, and appropriate measures must be taken.
This is one of the main reasons AI must remain at the service of people. It must never be allowed to replace professional judgement, team experience, professional expertise, or managerial responsibility.
We therefore make sure that human responsibility and professional judgement remain at the heart of every AI-assisted decision, so that every choice is understandable, explainable, and fully owned.”
AI can help us spot and identify issues sooner, understand them better, be more reactive, proactive even; but it will never replace vigilance, dialogue, and the collective responsibility of the men and women on the ground.
