Construction Safety AI Moves From Observation Logs To Field-Level Intervention

Construction sites don't lack for tools. They do lack tools that turn frontline observations into action quickly enough to change what happens on site. The starting point is the most stubborn friction point in safety reporting: getting a worker in gloves, in the cold, on patchy connectivity, to log what they just saw with enough detail to matter. A guided voice agent turns that moment from a form-filling task into a phone call that captures structured safety data in two or three minutes, giving teams a faster path from observation to intervention while keeping workers focused on the site instead of the screen.

Lem Prestage is Group IT Manager at Southbase Group, a New Zealand construction company that helped rebuild Christchurch’s key infrastructure after the 2011 earthquake and now operates across construction, labor, and prefabrication. He owns IT strategy across cybersecurity, integration, and AI, but his view of the role is deliberately practical: technology should reduce friction for the people doing the work. That's why he frames the safety rollout less as an AI project than as a responsibility to the workers on site, built around a simple goal: let them report what matters and get back to building.

"It can take seven to 10 minutes to do a health and safety observation on a mobile phone. With the voice agent, it fills the form in for you, sends everything through, and the average time is about two to three minutes," said Prestage. The app-based reporting process was supposed to make safety observations easier, but on site it often became another barrier. Workers lost connectivity, struggled with gloves or weather, or snapped a photo to write up later, which meant reports arrived slower and with less useful detail. Prestage tested the old process himself and found that even a straightforward observation could take seven to 10 minutes to complete. Replacing that workflow with a guided phone call cut the average report to two or three minutes, while also changing the worker’s posture on site: they could describe what they saw while still looking at the hazard, not down at a form.

• Friction removed without structure lost: "If you're on site with an iPad or a phone and you've got gloves on and it's cold or it's wet, then it's not going to be working as well," Prestage said. "Making a phone call just makes life a lot easier." The agent prompts for what the form needs, confirms it back, and flags whether the behavior sounds safe or unsafe before filing. That keeps the data structured while the worker stays heads-up. As Prestage put it, it would be ironic to have a safety incident while filling out a safety observation, and a worker on a call is watching the site rather than thumbing through checkboxes.

Faster reporting matters because the data gets better, and better data only works when the systems behind it are connected. At Southbase, that connective layer started before the safety rollout, with a facial recognition access system that needed to sync to a cloud attendance register. From there, Prestage began looking across the company’s 70-plus SaaS platforms for workflows that could be stitched together instead of managed in isolation.

• Why orchestration became necessary: "Once we had that first use case working, the question became: where else is the business losing time because systems don’t talk to each other?" he said. Onboarding came next and is now close to zero touch. The same integration discipline is what lets a spoken observation land cleanly in the safety system, turning a faster report into structured data the business can actually use.

• Good data or nothing: The input problem is where the safety rollout becomes more than a faster reporting workflow. "AI can only work with the data you give it. If the observation is thin, the output will be thin. If the data is wrong, the system will either skew the result or fill in gaps that should never have been left open," Prestage said. Frontline staff carry a KPI for observations per week, but the old entries were often minimal, closer to a tick-box exercise than a usable signal. Richer observations, captured consistently, are what make pattern detection possible rather than decorative.

• From observation to intervention: That data doesn't sit in a reporting archive. With as many as 20 projects running at once, it flows into dashboards that show when the same risk is appearing on one site or across several. "If we see five observations from one site, or 10 across a few sites, we can see there’s a systemic issue. That gives us real data to say: this behavior needs to change." From there, the team can issue a health and safety notice, raise the pattern in pre-start meetings, and bring subcontractors into the fix. Generative AI now helps draft those notices, and the next iterations extend the voice agent into workers’ first languages for a crew that includes many non-native English speakers.

The roadmap points toward agentic workflows that can read the safety data, identify patterns, draft the response, and warn teams on new projects that resemble past risk profiles. For Prestage, part of the value is removing the bias that can shape how people interpret dashboards, but the boundary is clear. The system can surface the pattern and recommend the response. It doesn't own the decision.

• The end of the loop: "AI is never going to implement the change. You’re always going to need a human in the loop," Prestage said. Trust depends on being able to see the system’s reasoning before anyone acts on it. A recommendation can look logical in a dashboard and still fail on the ground: an agent might spot steel left in the same position across several sites and suggest banning steel on the ground entirely. That's not realistic. The practical fix is keeping the site tidy and taping off the hazard, the kind of common-sense judgment Prestage is not willing to automate away.

For a firm building in an earthquake zone, under codes that keep changing because the risks keep changing, accountability is not an abstract governance principle. It's the work itself. Prestage framed the rollout as decision support, not decision-making, and expects that boundary to hold for years. AI can help teams build faster, safer, and with better information, but the destination is still a finished building people can trust. "There’s always going to be a human at the end to make the final decision," he said. "We can’t have AI make all the decisions, because there’s the safety of the workers, and there’s making sure the buildings themselves are safe once they’ve been completed."