By Phil Molé, MPH

In Part One of our blog series on agentic AI, we introduced the concept of AI agents and discussed some of the things they can do.

The goal now is to discuss what they should do, and what they shouldn’t do.

In what follows, you’ll learn why AI agents are not created equal, and what you should have top of mind as you’re exploring the potential of agentic AI for EHS management.

Agents Need Guardrails. AI Agents Are No Exception.

First, let’s start with a quick recap of some of the key points from Part One of this blog series, specifically the general discussion of what agents do.

In Part One, you learned that agents act in pursuit of goals. The most important feature of an agent is not how much it acts, but whether it’s acting responsibly and with oversight. For example, not just anyone can be an FBI agent because many people can’t pass the background checks or make it through the rigorous training program. Even if they do make it through those hurdles, they can only act with the approval of the FBI director.

In other words, the definition of an agent, any agent, includes the concepts of authorization, trust, and boundaries. No credible agent operates without oversight.

The same is true for AI agents. And in the case of AI agents for EHS, even more so.

Agentic AI can plan, act, and adapt. That’s what makes it powerful. But in EHS, where decisions directly impact human health and safety, unchecked autonomy is dangerous.

For example, consider the possibility of an AI hallucination, particularly in contexts of EHS. An AI hallucination is a seemingly valid insight, recommendation, or conclusion generated by AI that isn’t supported by real safety data, best practices, established standards, or the specifics of a given situation.

A hallucination in the context of a lower stakes AI operation, like a form-fill or generation of text summarizing meeting minutes, can be a minor annoyance, easily corrected through a little work on the back end. But in a high-risk environment, a hallucination about the appropriate control to use in job task involving hazardous equipment can make a severe injury or fatality (SIF) more likely.

Possibilities like this are the reason why industry experts are already clear on this point: Agentic AI in EHS should follow a human-in-the-loop model. That is, actual humans should validate and strengthen AI outputs before action is taken.

It’s worth pointing out that a human-in-the-loop approach does not mean humans are infallible. We all know they’re not, but that doesn’t mean their expertise doesn’t count. The point of good AI is to supplement that expertise and strengthen judgement calls. That’s why it also matters in the types of AI that is used. Ensuring you leverage purpose-built AI, trained on real EHS datasets is essential. It’s the combination of reliable AI with human intelligence that produces the best results.

A human-in-the-loop approach is the only way to make sure that inaccurate AI outputs don’t go unreviewed and create risks instead of controlling them. A missed hazard, an incomplete root cause, or an unsafe recommendation from AI can all potentially lead to major accidents.

That’s not just a software problem. That’s a safety outcome. And it comes from the choices you make as well as those you don’t make.

The Kinds of AI Agents Worth Having in EHS

It follows that to avoid these kinds of outcomes and get the outcomes you want, like fewer injuries and better engagement with safety initiatives, you need to make better choices. You don’t just need AI for AI’s sake, or AI agents because they’re the new thing. You need purpose-built AI for EHS and the kinds of AI agents worth having. This means you need to be precise about the roles that agents should play.

The chart below reviews different use cases for agentic AI, summarizing the issues that can occur without AI support, the problems involved in using without oversight, and the benefits of a human in the loop approach.

Task What It Looks Like Without AI Support What It Looks with an AI Agent, But Without Oversight The Human-in-the-Loop Approach
Incident Description Analysis

 

 Incomplete descriptions, lacking important details needed to understand how the incident happened.

 

  

An agent may automatically assess your incident description and implement changes it considers improvements. It might be right; it might not be. But if it’s not, and you haven’t reviewed its changes, the agent may not only fail to improve your description but also introduce new problems. That in turn undermines root cause analysis and selection of actions since both depend on a good description.

 
  • The investigator enters an initial description, then engages AI to assess the description and offer improvement suggestions. The EHS professional can assess the AI output and choose whether to accept or reject them.

 
Root Cause Analysis (RCA)

 

 Choice of common, but unhelpful causes, like human error, or guesses on causes unsupported by the incident details.

 

  

Once you’ve finished your incident description, an agent can potentially identify root causes. Again, it may do so correctly and hopefully will do so based on training on real EHS datasets. The problems come if the RCA is inaccurate and doesn’t get reviewed by a human SME. Wrong root cause selections prevent you from having the knowledge needed to address underlying risks.

 

 

 
  • The EHS professional chooses whether to engage AI RCA. When engaged, the agent assesses the incident description in light of its dataset on incidents and associated root causes, and surfaces the most relevant root causes. The EHS professional chooses whether to accept or reject the RCA identifications.

 
PSIF Identification Progress identifying PSIFs, if made at all, is slow. The selection process is marred by subjectivity.

 

 Here, the risk runs two ways. Attributing PSIF status to less serious incidents distracts you from focusing on the right things. On the other hand, missing PSIF risks means missing the chance to address them before a major accident happens. Still, it’s better to have at least a few false positives instead of false negatives to avoid missing PSIF risks.

 

  

  • After entering the incident description, the EHS professional can choose to engage the agent to assess if the incident is a PSIF. The agent then reviews the incident details and assesses, based on its training on real EHS datasets, whether the incident is a PSIF. The EHS pro can either accept or reject the determination, keeping in mind that having some false positives is a lot better than having false negatives.

 
Corrective Actions Since many activities generate actions, management and follow-up is time-consuming and difficult. Actions may not address every root cause identified.

 

 Potentially, some agents may automatically recommend actions after the incident description and RCA are complete. But what if the actions aren’t a good fit to root causes, and what if the risks involved are severe?

 

  

  • After completing an incident description and RCA, the EHS professional can decide whether to engage the agent to identify appropriate actions. The agent then uses its training on safety incident datasets to find and recommend actions that are most closely associated with the identified root causes. The EHS pro reviews the action selections and can either accept or reject them.

 
JSA Task Description Task descriptions may lack enough details to identify job hazards or to assess the extent of associated risks.

 

 Some AI agents may automatically assess your JSA job task descriptions and then implement changes to improve them. If they do improve them, that’s great, but if not, the whole foundation of your JSA is eroded.

 

  

  • An EHS team member writes a draft job task description and can choose to engage the AI agent to assess the descriptions and recommend improvements. EHS team members can either accept or reject the suggestions.
JSA Hazard Identification  

Users who haven’t done many JSAs, or don’t have experience recognizing hazards in job tasks, may miss the mark on hazard identification. For instance, someone who hasn’t encountered a specific hazard before won’t recognize it.

 

 After your job descriptions are done, some agents may then automatically move on to identifying hazards. You don’t want those selections made without your review because incorrect hazard identifications will prevent you from selecting good controls.

 
  • After finishing the task description, an EHS professional can engage the AI agent to use its dataset training to identify hazards. The EHS pro can choose to accept or reject the identifications.
JSA control selection  

Here again, lack of background in operational risk management limits the accuracy of selections. Errors in task descriptions and hazard selections can cascade to missing identified controls.

 

  

Once you’ve finished your JSA description and identified hazards, some agents would then immediately take the next step and select controls. But controls are really where the rubber meets the road. Human review is needed to ensure that selections are appropriate and will reduce, if not eliminate, the risks.

 

  

  • Once the EHS professional has completed task descriptions and identified hazards, she/he can engage the AI agent to identify effective controls to reduce risks.

 

 

If you review the different situations laid out in the chart above, you’ll see a pattern. It goes something like this:

  1. 1. The traditional ways of doing many foundational EHS management tasks take time. That time, in turn, makes EHS teams less efficient, with reduced bandwidth to focus on proactive safety management measures, like leading indicators.
  2. 2. These administrative burdens play a big role in keeping EHS pros trapped in a reactive safety management cycle. Technology, including AI agents, are one of the tools EHS teams are exploring to help them escape. But while agents can certainly facilitate faster turnaround times for tasks, agents that operate without a human in the loop may be missing safety risks or even creating new ones.
  3. 3. A human-in-the-loop approach provides all the advantages of technology, while still involving review and approval by an SME. The greater the workplace risks associated with specific tasks, the more important the human-in-the-loop approach becomes.

This is what human-in-the-loop really means: Speed, yes, but speed with purpose, that doesn’t come at the expense of sound judgment.

These conclusions also align with the framework of the EU AI Act, which establishes four risk categories for AI:

1) Unacceptable risk: AI that represents an inherent threat to fundamental rights and human dignity, such as biometric categorization systems that infer sensitive characteristics, including sexual orientation, political opinions, or philosophical beliefs.

2) High risk: Examples include autonomous driving or biometric identification of individuals involved in criminal activities or investigations. These activities must meet strict requirements and obligations in the EU market, such as rigorous testing, transparency and human supervision.

3) Limited risk: AI systems such as chatbots that generate content and are subject to transparency obligations, such as informing users that AI generated the content so that they can make informed decisions concerning further use.

4) Minimal/negligible risk: Applications, such as AI-driven games that have minimal baseline risks which can be easily controlled, such as via spam filters.

The key takeaway as you move from high to low-risk categories under this framework is that, aside from the moral issues involved in some AI applications, AI can sometimes perform tasks that are inherently higher in risk, and it’s necessary to limit that risk. Human supervision is one of the most effective ways to do that, so AI doesn’t perform autonomous actions with potentially severe consequences.

The goal is better decisions, made faster, with the right level of human expertise applied at the right time.

Take a Holistic Look at AI for EHS

To get the most from agentic AI, you need to take a step back and take a more holistic view of AI for EHS. Circling back to an earlier point, you get the best results when high-quality AI trained in the right way augments human expertise.

For example, look at the entries in the chart above for RCA, PSIF identification, and actions selection.

In each of the best scenarios for each task, the one involving human-in-the-loop agentic AI, the AI is specifically trained on real EHS datasets to understand the real-world contexts of the information it reviews and acts upon.

It can accurately and quickly identify PSIFs because its algorithm was honed on vast amounts of data for types of incidents and associated hazards that carry PSIF risks. It can select appropriate actions because its training helps it rapidly flag correlations between specific root causes and the types of actions that would most effectively address them.

Man Clipboard Industrial Site 1200x628 EHS professional

In the best AI for EHS software, this algorithmic training occurs under the structured supervision of real, human EHS SMEs. The whole AI for EHS software platform is purpose-built for EHS, meaning that real EHS experts have trained it on real EHS data, for use by real EHS professionals. The output is therefore attuned to the finer details of context and is relevant to the specific EHS task performed in a way that the output of general AI capabilities could not be. And as an extra safeguard, an EHS expert can still review and approve the AI output.

One takeaway here is that you really need human-in-the-loop not just when it comes to evaluating the outputs of AI agents, but also on the front end, in the training of the software itself.

Another takeaway is that it’s not only what agents do that’s important, but also how they’ve been trained to do it. If generic AI is bolted onto EHS workflows, then an agent completing tasks it’s not adequately trained to do won’t save time in the long run or move the needle on reducing risks.

Key Takeaways on Evaluating Agentic AI

As agentic AI becomes more common, EHS teams will increasingly face the challenge of differentiating between helpful AI and problematic AI. Not all agents are designed with EHS in mind because some are just general-purpose tools retrofitted for safety workflows.

Others emphasize autonomy as a selling point, without clearly defining boundaries. That’s where problems start. Instead, EHS teams should be looking for AI that is:

Purpose-Built for EHS

Make sure any AI that you are considering, including agentic AI, isn’t a general tool or adapted from another domain. Look for proof points that it’s been designed and trained using real-world safety data for use in real safety workflows.

Shaped by Human Expertise

The AI worth wanting is the AI that’s not only been built, but is also continuously refined, with input from EHS subject matter experts on the front end.

Designed with Oversight

Remember, an emphasis on oversight is baked right into the concept of an agent. Clear checkpoints with human oversight are not optional and are essential in high-risk environments.

There’s a tendency in AI conversations to frame autonomy as the end goal, but in EHS, that framing doesn’t hold.

The future isn’t about AI replacing safety professionals. It’s about AI working alongside them, augmenting and extending their reach, sharpening their insight, and helping them focus on what matters most.

Organizations that succeed will be the ones that embrace a balance among these factors:

  • 1. Speed with purpose
  • 2. Automation with accountability
  • 3. Intelligence with human judgment at the center

Organizations that achieve this balance and work with partners that understand these factors will act sooner and smarter.  The right AI won’t take you out of the loop. It will make your role in that loop more impactful than ever.

Looking for More AI Insights?

View additional resources on AI and EHS and strategies for adopting a proactive management approach, including:

Also, be sure to follow our blog or more information on the latest EHS news and insights.

Ready to See What Purpose-Built AI Looks Like?

VelocityAI capabilities, powered by Vēlo, are human-centered and purpose-built for EHS, designed to integrate with core workflows, surface actionable insights, and support safety leaders in driving measurable outcomes. Schedule your demo to see for yourself.

Cta Insert Schedule A Live Demo