Sky High Thinkers

 

close
Welcome to Libryo’s Sky high thinkers blog, where thoughts on legaltech, sustainability, law, compliance and technology converge. Filter by topic or popular reads and share your thoughts with us and others. If you have a topic that you’d like to see covered, drop us an email: info@libryo.com

Stay in the know by receiving monthly email updates directly into your inbox.

What Stays Human: The Four Things AI Can't Do in EHS

Written by Malcolm Gray
on November 18, 2026

Walk into any EHS or compliance team meeting this year and you'll hear a version of the same conversation: what's AI going to change about our work? It's a fair question, and a loud one. AI is now reading regulations, extracting obligations, summarising environmental and health & safety changes, even drafting first-pass audit responses. The automation story is real, and it's accelerating.

But the more useful question - the one that gets a lot less airtime - is its mirror image: what's AI never going to change about this work?

We think there are four parts of EHS compliance that don't belong to the algorithm, no matter how good it gets:

Taste, judgment, relationships, and accountability.

Here's why each of them matters more, not less, in a world where AI is doing the rest.

1. Taste (or Context)

Taste is knowing what's worth paying attention to.

A good EHS manager reads a regulatory update and immediately knows whether it matters. A new heat stress rule in a state where most of the workforce is outdoors? That changes the game. A tweak to the reporting threshold on an obscure chemical your site hasn't used in five years? Worth logging, not worth losing sleep over. That instinct doesn't come from running a probability model, it comes from having seen enough updates to know which ones quietly change everything, and which ones only look big.

AI is already excellent at summarising what a regulation says. It's much further away from knowing whether a regulation matters to this particular site, in this particular moment, under this particular operation. That's taste or context - and taste is earned by paying attention over time to the kinds of things no training data captures well.

2. Judgment

Judgment is knowing what to do when the rules don't cover the situation.

Real EHS work is full of grey areas. A Stage 3 drought declaration says "non-essential industrial water use must stop", but is your cooling tower essential or non-essential? A new regulation technically applies to your warehouse but was clearly written for a petrochemical plant. A near-miss must be reported "within a reasonable period" - and nobody agrees on what that means. Two permits contradict each other in a way that forces you to pick.

In those moments, you don't need a summary of the rule. You need to decide what a reasonable person would do with the information available, given the context of this site, this operation, this history with the regulator, and this risk appetite. That kind of reasoning has to sit with a human, not because AI can't reason, but because the responsibility for the outcome has to sit somewhere specific.

Which brings us to the next two.

3. Relationships

EHS is often described as a technical discipline. In practice, it's a deeply relational one.

The ISO 45001 lead auditor you've worked with for three cycles trusts your account of what happened at last year's incident, because you've been straight with them before. The environmental agency inspector takes your call about a permit question because your predecessor built that relationship. The operations director listens to your warning about a new regulation because you've earned a reputation for only flagging what truly matters. The site safety committee shows up to meetings because they know you'll actually act on what they raise.

None of those transfers through software. The people on the other end of an EHS conversation - auditors, regulators, workers, operations leads - are still people, and the quality of the relationship often decides how the conversation goes. AI can draft an email. It can't build trust.

4. Accountability

When something goes wrong: a workplace injury that could have been prevented, an environmental breach that wasn't reported in time, a heat break protocol that was skipped on a red flag day, a permit condition that was missed, someone has to answer. Not an algorithm. A person, with a name, at a company, in a role.

That's why the serious conversation about AI in EHS isn't really about what the AI can do. It's about who's accountable when it does it. The best AI governance frameworks we've seen all come back to the same principle: for every service the AI touches, there's one named human responsible for the outcome. Not a team. Not a committee. A person.

Without that, you don't have AI-enabled EHS compliance. You have an accountability gap wearing a chatbot as a disguise.

The quiet truth

The more AI we build into EHS compliance work, the clearer these four things become. Taste, judgment, relationships, accountability, none of them are going anywhere. If anything, they become more valuable, because everything else is getting faster and cheaper around them.

So the real question for EHS teams this year isn't "what will AI change about our work?" It's "what are we going to do with the time, attention, and energy it gives back to us?"

The EHS professionals who will thrive through this transition aren't the ones trying to keep up with AI by doing the AI's job slightly faster. They're the ones leaning into the work only humans can do – trusting the AI to do what the Ai does, reading the room, making the judgment call, owning the outcome, being trusted by the regulators, auditors, and colleagues who actually decide how a compliance conversation goes. That's where the career grows. That's where the value lives.

Want to see how Libryo is balancing the latest AI with the human side of EHS compliance? Book a demo with the Libryo team →

 

 

Related reading

The Rise of Agentic Compliance: What AI Agents in EHS Actually Do - the companion piece on what the AI side of this conversation looks like


How to Track Regulatory Changes Across Multiple Jurisdictions (Without Losing Your Mind) - practical guidance on the automation layer that frees up the human work