AI-Powered Safety Tools Transform Construction Sites

Last winter, a tragic incident occurred during the construction of an affordable housing project on Martha’s Vineyard, Massachusetts. A 32-year-old worker, Jose Luis Collaguazo Crespo, slipped off a ladder on the second floor and fell to his death in the basement. This incident is a stark reminder of the dangers faced by construction workers, who account for over 1,000 fatalities each year in the US, making the industry the most dangerous for fatal slips, trips, and falls.

Philip Lorenzo, an entrepreneur and executive, addressed this issue during a presentation at Construction Innovation Day 2025 at the University of California, Berkeley. He highlighted the ongoing struggle between prioritizing safety and maintaining productivity on job sites. “Everyone talks about safety being the number-one priority, but internally, it’s not always that high on the list,” Lorenzo said. “People take shortcuts, and that’s where the risks come in.”

To address these risks, Lorenzo is developing a tool called Safety AI for DroneDeploy, a San Francisco-based company. Safety AI uses generative AI to analyze daily reality capture imagery from construction sites and flag conditions that violate Occupational Safety and Health Administration (OSHA) rules with 95% accuracy. Launched in October 2024, the tool is now being deployed on hundreds of construction sites across the US, with versions tailored to building regulations in Canada, the UK, South Korea, and Australia.

Safety AI is part of a growing trend of AI construction safety tools emerging from various tech hubs, including Silicon Valley, Hong Kong, and Jerusalem. Many of these tools rely on human “clickers” to label images of key objects, but Safety AI is the first to use generative AI to reason about what is happening in an image. This advanced form of analysis goes beyond simple object detection, allowing the software to draw conclusions about potential OSHA violations.

Robots and AI typically perform best in controlled, static environments like factory floors or shipping terminals. However, construction sites are dynamic and change daily. Lorenzo’s team has developed a visual language model (VLM) to monitor these sites effectively. VLMs combine the capabilities of large language models (LLMs) with a vision encoder, enabling them to “see” and analyze images of construction scenes.

Using a “golden data set” of tens of thousands of OSHA violation images, Lorenzo’s team has trained the VLM to recognize and reason about safety risks. The model is guided through the analysis process by a team of construction safety experts who input strategic questions and tweak the prompts to improve accuracy. For example, the VLM can determine whether a person is using a ladder safely by analyzing factors like three points of contact and the position of the user.

Despite its 95% success rate, Safety AI is not infallible. Chen Feng, who leads New York University’s AI4CE lab, notes that VLMs still struggle with spatial reasoning and interpreting 3D scene structures from 2D images. Feng emphasizes the need to address the remaining 5% of errors, which can include edge cases and hallucinations.

To mitigate these issues, Safety AI incorporates older machine-learning methods, such as image segmentation and photogrammetry, to create spatial models of construction sites. These techniques help identify the most common violations, such as unsafe ladder usage, and provide an additional layer of accuracy.

Aaron Tan, a concrete project manager in the San Francisco Bay Area, believes that tools like Safety AI can be beneficial for overworked safety managers. “Having an extra set of digital eyes can save a lot of time and improve safety,” Tan said. However, he also notes that workers may resist these tools due to concerns about surveillance and privacy.

Izhak Paz, CEO of Safeguard AI, a company based in Jerusalem, has opted to use older machine-learning methods for their reliability. His team trains algorithms on large volumes of labeled footage to identify potential hazards in real-time. Safeguard AI is currently used at over 3,500 sites in Israel, the US, and Brazil.

Buildots, a Tel Aviv-based company, uses machine learning to create visual progress reports of construction sites. CEO Roy Danon emphasizes the importance of accuracy, aiming for a 99% success rate. Buildots is used by about 50 builders, primarily those with revenue over $250 million, in Europe, the Middle East, Africa, and other regions.

The integration of AI in construction safety is a promising development, but it requires careful implementation to ensure reliability and worker acceptance. As these tools continue to evolve, they have the potential to significantly reduce the number of accidents and fatalities on construction sites.