Our Director of Health & Safety, Jennene Lyda, M.S., CSP, CIH, is a key voice bringing industrial hygiene into the data center conversation – including at AIHA Connect 2026, where she presented on occupational and environmental hazards in data centers.  

It’s a conversation increasingly shaped by AI. As we see it reshaping data center infrastructure, we’re further considering the risks involved in keeping the people who work in it safe. 

How the risks are evolving 

As we see it, AI infrastructure isn’t introducing categories of risk that didn’t exist before – it’s increasing their scale. Higher-density AI servers require more power and more cooling, which translates directly into louder equipment and higher voltage loads. Electromagnetic field exposure is a further consideration we track closely: the human health effects aren’t yet fully understood, though the EU already has specific regulations in place to manage exposure. 

Even the fixes carry their own trade-offs. Water cooling, for example, reduces noise but introduces water-related risk in its place – a reminder that managing risk in this space is an evolving discipline, not a problem solved once and left alone. 

Automation adds another layer to that same pattern: it doesn’t remove exposure so much as relocate and disguise it. Battery rooms, generator testing, refrigerants, and noise in spaces designed around equipment and efficiency rather than people all carry hazards that are easy to miss once a process becomes automated – precisely the kind of nuance that having dedicated industrial hygiene expertise embedded within our business is designed to catch. 

The change is often audible before it’s anything else. Walking into a data hall fitted with new AI server racks is a noticeably different experience from a traditional one. And as more AI infrastructure comes online, that difference is only becoming more pronounced – one more reason we treat ongoing assessment, rather than a single fix, as the right approach. 

Designing around the real work 

That same instinct shapes how we approach exposure assessment itself. Data halls aren’t occupied the way a conventional workplace is; time on site is sparse and episodic, concentrated into maintenance windows rather than spread evenly across a shift. A standard eight-hour time-weighted average can miss the task-based peaks that happen during those windows entirely, so the more useful approach is designing sampling around the activity itself and anticipating risk based on what the work actually involves. 

Some of the most useful data doesn’t come from dedicated monitoring equipment at all. Building management and data center infrastructure management systems already track temperature, pressure and runtime hours for operational reasons – information nobody originally designed as industrial hygiene data, but which can double as a continuous exposure surrogate once you know how to read it that way. 

It’s also why we don’t see safety and uptime as competing priorities. In an industry built around availability, the more useful question isn’t how to balance safety against uptime, but how to design controls that strengthen reliability while reducing exposure – so both are achieved together, rather than traded off against each other. 

Safety by design 

Building safety considerations into decisions from the outset, rather than responding to them once they surface, is the ideal for any data center development. As Jennene explains, “part of what that looks like in practice is what we plan around in the first place: traditional safety planning often centers on incident rates, but low recordable numbers have never reliably predicted a serious event. We organise planning around high-energy hazards instead, and pull stakeholders into risk planning at the design stage rather than waiting until mobilisation”. 

Collaboration between our safety function and the teams making day-to-day delivery decisions means the ability to stay closely connected – flagging, for instance, that one piece of equipment may carry more risk than an alternative before it’s installed, rather than after. That proximity is reinforced by our leadership team, for whom safety is treated as a standard consideration, built into how we operate. 

Building dedicated expertise into the business 

Much of Jennene’s role reflects our growth phase directly. Rather than maintaining an already-established safety programme, a significant part of her work is building the operational safety infrastructure we need as we scale – including anticipating industrial hygiene risks that dedicated expertise, still rare across this sector, is only beginning to bring into focus. 

Staying embedded in facility design and day-to-day operations is less about reliance on a specific tool and more to do with trust and credibility, built by demonstrating the value of safety input consistently enough that it becomes part of how decisions get made. As Jennene notes, “that shows up in how we train, too: rather than training people to recall a fixed scenario, we increasingly focus on rehearsing judgement, including what to do when the procedure in front of them doesn’t quite fit”. This allows for building the capacity to absorb a surprise, not just proving that the rules were followed.  

That principle extends to how we think about automation. As more operational tasks – switching processes, for example – become automated or robot-assisted, we don’t expect the need for human oversight to disappear. AI and automation can reduce direct human exposure to certain hazards, but someone still needs to be accountable and oversee what’s being done. 

Why human judgement still leads 

Jennene explains that, “in practice, it’s less a policy than a set of habits – built on the assumption that error is normal and blame fixes nothing. Pre-task conversations focus on the crew naming what could go wrong, rather than reciting a checklist. When something does go wrong, the response starts from a different question: not who missed a step, but what made that make sense at the time”.  

We also walk tasks periodically with the people actually performing them, to see where the written procedure and the real work have quietly drifted apart, then feed that back into fixing the procedure or adding a control. Near misses and safety events go to facilitated learning teams involving the people who were there, rather than a traditional top-down investigation – and every one of those sessions closes on the same question: which control would have caught this, regardless of who was on shift. 

It’s the same scepticism that shapes how we think about AI tools and data more broadly: data itself is a tool, and understanding its biases matters as much as the numbers themselves. It’s also important to acknowledge that AI, whatever its other capabilities, has no morality or fidelity of its own. For us, that’s precisely why human judgement remains central – not despite the growing sophistication of the tools available, but alongside them. 

It’s the standard we hold ourselves to at Yondr: bringing safety and industrial hygiene into the conversation early while staying agile and anticipatory, rather than reacting to risk once it becomes visible. We continuously evaluate, adapt and reassess as technology, and the industry around it, keeps moving.