Everyone wants AI in their EHS program. Boards are asking for it, vendors are racing to ship it, and safety leaders are under pressure to show they have a plan. But there's a catch: AI is only as good as the data underneath it, and that data starts with the frontline worker filling out a form at the end of a long shift.
That's the frontline paradox, and it was the focus of our recent fireside chat featuring Eric Stevens, CTO and co-founder of EHS Insight, and Nathan Goldstein, Senior Manager of EHS&Q research at Verdantix. Their conversation drew on research from Verdantix, an independent research and advisory firm. Per Verdantix, "The 2026 Verdantix global corporate EHS survey gathered insights from 302 senior decision-makers across 25 countries and 24 industries, primarily from those employed in large enterprises and director-level roles."
Here's what stood out.
1. Data quality is the foundation, and the top AI use case
Improving EHS performance is still the leading digital strategy priority. But according to Verdantix, "Data quality, standardization and governance are also gaining importance, with 28% of firms ranking them first and 25% second."
The same theme shows up in AI plans. Of the 12 AI use cases Verdantix tracked, "Cleaning, organizing and improving EHS data quality leads adoption, with 60% of firms having either widely rolled it out or looking to increase deployment."
Nathan called this a foundational weakness: "You can't build on a bum foundation." It's also a catch-22. The most popular AI use case in EHS is fixing the very data that AI depends on. As Eric put it, it's hard to see the benefit of layering on AI analyzers and agents if you haven't secured good data to begin with.
Garbage in, garbage out isn't a new idea. What's new is how much more is riding on it.
2. Better data starts where it's captured: with the frontline
Why is EHS data so hard to get right? Eric pointed to the friction frontline workers face every day. They step away from operational duties to fill out forms they're sometimes unsure about. Submissions get rejected and sent back. The result is people entering what they think is right, skipping steps, or doing the bare minimum.
Historically, the fix was more training, more supervisors and more workflow steps. EHS Insight has taken a different approach: putting AI coaching directly into the data entry experience. That can mean:
- Suggesting clearer, more complete incident descriptions
- Prompting for a photo when a description mentions something that wasn't captured
- Prompting for more detail based on what appears in a photo
- Better answers. The connector gives the AI the metadata and business context behind the data, so it guesses less about how records relate. Instead of a small exported slice, the AI sees the full picture the user is allowed to see.
- An audit trail. No more spreadsheets floating around and being uploaded into unsanctioned tools. Access runs through a provisioned, reviewed connector to an approved AI vendor.
- Less friction. Users stay in the AI tool where they already work.
The goal is to replicate an experienced EHS manager at the moment data is captured. Workers get real-time coaching that helps them finish faster and with more confidence. Management gets better data that flows into reporting, trend detection and serious injury and fatality (SIF) precursor analysis. "Everyone wins," Eric said.
Nathan reinforced the point with a story from Verdantix's RFP work. A haul truck driver at a northern Canadian mine had to juggle three separate apps on his phone just to log miles, hours, and once a threshold was triggered, mental health resources. The most successful software selections, he noted, are the ones that gather input from the frontline early.
3. MCP: bringing EHS data to the AI tools people already use
MCP (Model Context Protocol) is an industry standard that lets data sources connect securely to AI assistants like Claude, ChatGPT and Microsoft Copilot.
Today, most people analyze EHS data with AI by exporting a spreadsheet or PDF, uploading it to a chatbot and asking questions. EHS Insight's MCP connector removes that export-import step, and Eric explained why that matters beyond convenience:
4. Governance: from shadow AI to approved AI
"Is my data safe?" is the question on everyone's mind, and the survey backs that up. Verdantix found that "70% of respondents identify AI safety, governance and regulation considerations as the 'most important' or a 'very important' barrier." Eric framed it as a choice among three paths:
- Ban AI entirely (some companies do)
- Let employees use free or personal AI accounts, uploading company data into an unmanaged ecosystem
- Sign an enterprise agreement with an approved AI provider and connect to it in a governed way
The MCP connector is built for the third path. It brings AI use out of the shadows and into a space IT already controls, often with a vendor the organization has already approved.
Just as important is who can see what. Some vendors ship AI features designed only for power users with access to everything. EHS Insight took the opposite approach, building on standard protocols like MCP and OAuth so the AI connects with the user's own credentials. It sees only the records and fields that user can already access. A site manager's AI session stays scoped to their site. Data outside their permissions is never made available to the AI at all. Auditability, response quality and guardrails all build on top of that foundation.
Nathan noted that governance is a growing differentiator in Verdantix's upcoming EHS Green Quadrant benchmark, where vendors are stratifying between scattershot AI, targeted use cases and wait-and-see approaches, with a growing subset pursuing MCP.
5. Designing for trust when AI isn't deterministic
"In the age of AI, the real currency is truth," Nathan said. So how do you design AI people can trust?
Trust is a real hurdle. According to Verdantix, "Concerns around accuracy and hallucinations are key barriers for 66%, while 59% cite a lack of transparency and auditability of AI outputs."
Eric's answer: stop treating AI like a calculator. People expect software to give them the same correct number every time. AI behaves more like a coworker who usually understands the question, but sometimes only partly does.
Early AI features were often single-shot: data in, answer out. That encouraged people to over-trust the output. Today's better designs position AI as a coach or collaborator, something you interact with iteratively, ask where it got its data and question how it reached a conclusion. Eric argued that software vendors have a responsibility to build that understanding into the product design, rather than assuming users already know AI's limits.
Nathan added that explainability will only grow in importance as regulators catch up and start asking how AI systems arrive at their recommendations.
What’s Next
Asked for a prediction, Eric came back to the fundamentals: everything downstream breaks without good data in.
Beyond that, he expects that as people rely more on tools like ChatGPT, Claude and Copilot, they'll expect their EHS data to be there too. Nobody wants to context-switch across 10 or 15 apps to piece together an answer. He compared it to the shift from desktops to mobile: work follows where people already spend their time.
As Nathan summed it up, the goal hasn't changed. It's making sure everyone goes home with 10 fingers and 10 toes. AI that serves the frontline, not just the boardroom, is how we get there.
Watch the Full Conversation
Watch: Ask the Expert to hear the full discussion between Eric Stevens and Nathan Goldstein.
Source: Verdantix, Global Corporate Survey 2026: EHS Budgets, Priorities And Tech Preferences, August 2026.
