By Phil Molé, MPH

If you use a generic AI tool to write a safety checklist, you’ll probably get something that looks convincing. It may be well organized, clearly written, and full of familiar terminology.

But does it accurately address the hazards at your facility? Does it account for your specific equipment, processes, and existing controls? Can it distinguish an incident with minor consequences from one that could easily have caused a fatality?

Those questions matter because EHS professionals need more than plausible, surface-level answers. They need reliable information that helps them make decisions about management of real risks.

AI has enormous potential to support that work. But getting meaningful results requires more than adding a chatbot to an EHS application. It requires purpose-built capabilities grounded in EHS data, shaped by subject matter experts, and connected to the workflows and business context that make information useful.

1) EHS Decisions Require EHS Context

General-purpose AI tools can help summarize information, brainstorm ideas, and draft communications. But their broad capabilities don’t automatically equip them to evaluate specialized EHS problems.

Consider an incident report that says, “An employee narrowly avoided being struck by a falling object.”

A generic tool might improve the sentence. An EHS-focused capability can go much further and assess what details are missing from the description of the incident and then recommend improvements. For instance, it can give you specific prompts to address questions like:

  • What fell?
  • From what height?
  • How heavy was it?
  • Where was the employee standing?
  • What controls were supposed to prevent the event?

Those details establish the foundation for understanding what happened and how serious it could have been. For the next step of root cause analysis (RCA), the details determine the underlying, systemic reasons why the incident happened.

Two Ehs Professionals

Purpose-built AI is trained and validated on relevant, representative EHS datasets for the tasks it performs. For capabilities that generate answers from documents or business records, it should also draw on reliable sources and the appropriate operational context.

Training, access to information, and validation serve different purposes. A system might have access to your incident records without having demonstrated that it can reliably identify serious injury potential. An AI feature might recognize safety terminology without understanding how to apply it within an investigation.

That’s why “AI-powered” tells you very little by itself. You need to know what the capability was designed to do, how it was trained to do it, and how its performance was evaluated.

2) Human Expertise Belongs Throughout the Process

AI for EHS should reflect human expertise from development through everyday use by EHS teams.

This point builds on the first one about purpose-built AI. Effective, purpose-built AI is effective precisely because of the input of real human subject matter experts during the development and training of the models. During development, EHS subject matter experts help define meaningful problems, evaluate data, assess outputs, and identify potentially dangerous errors. Their involvement helps ensure that a technically impressive model also supports sound safety practices in real EHS workflows.

That design also considers the end use of the software, and leaves room for real EHS professionals to work with and assess the output of the AI.

For example, an AI system might recommend employee retraining after incidents. An experienced safety professional would ask whether those recommendations overlook equipment design, maintenance, work planning, or other organizational factors.

EHS professionals should be able to review recommendations, examine supporting information, challenge conclusions, and decide what action is appropriate. That review needs to be meaningful. A person clicking “approve” without enough information or authority to question the result provides little protection.

Regulators recognize these concerns. The EU AI Act includes requirements for human oversight of systems classified as high-risk. That risk classification depends on the system’s intended use and the Act’s criteria; it does not automatically apply to every EHS AI application.

In the UK, the Health and Safety Executive’s (HSE) regulatory approach addresses AI where it affects workplace health and safety, including its use in machinery and safety-related applications. HSE is also examining how reliance on AI can affect human performance in major-hazard settings.

The practical implications become especially important with agentic AI, which can take steps or initiate actions toward a goal. As a system gains the ability to act, organizations need clear boundaries around permissions, review, escalation, and human intervention. For example, HSE guidance makes clear that EHS professionals can’t simply outsource all of their risk management to AI.

For EHS professionals, the goal should be AI that strengthens their judgment, instead of replacing it, and gives them more time to apply it.

3) Purpose-Built AI Builds Strength Upon Strength

The value of purpose-built AI becomes especially clear when you look at the lifecycle of a typical incident investigation.

Each step depends on the quality of the information and decisions that came before it. AI can support that progression when its capabilities are designed to work together.

Start with a stronger incident description. AI can assess whether a report contains enough detail and prompt the reporter to fill gaps. Those prompts should help people document facts they know and identify questions that require follow-up.

Use that information to support better root cause analysis. A more complete description gives investigators a stronger starting point. AI can then suggest contributing factors and questions to explore, while investigators verify those possibilities against interviews, records, and other evidence.

Evaluate potential severity. The actual outcome doesn’t tell the whole story. An event that caused no injury may still have had the potential for a serious injury or fatality. AI trained and validated for potential serious injury and fatality (PSIF) identification can help flag events that warrant closer review. You can even set up automatic notifications when a PSIF is detected so you can start reducing those risks as quickly as possible.

Develop more effective corrective actions. Purpose-built AI can build on improved incident descriptions and RCA by recommending effective corrective actions to address the specific root causes. Recommendations become more useful when they reflect the event, its contributing factors, and the controls that failed or were missing. Investigators can then evaluate proposed actions against the hierarchy of controls and the realities of the work.

Two Ehs Pros (back View)

Each improvement supports the next. Better reporting supports better analysis. Better analysis supports more informed prioritization and corrective action.

A generic tool working from a short, isolated prompt may miss those relationships. Connected, purpose-built capabilities can carry relevant information through the investigation while keeping people involved at each decision point.

4) Connected Workflows Make AI More Useful

AI can help organizations get more value from their EHS data. But that’s only possible if the underlying information needs to be accurate, accessible, and connected.

That’s a significant challenge for many organizations.

The 2026 EHS 360 Benchmark Report found that 59% of respondents agreed that data quality and availability limited their ability to make informed EHS decisions. Disconnected software contributes to this problem.

Download 360 Report Canva Cta

An organization might manage incidents in one system, ergonomics assessments in another, contractor qualifications somewhere else, and corrective actions in spreadsheets.

Each system may work reasonably well on its own. But the connections between them are often where important insights are lost.

Consider an incident involving repetitive motion.

An incident management system might document the injury, while a separate ergonomics tool contains assessments of the task. If those systems don’t communicate, EHS professionals may need to manually piece together the information necessary to identify contributing factors and evaluate controls.

The same problem can occur when an incident reveals weaknesses that should be addressed by a JSA.

A true connected EHS platform helps address these gaps by bringing related data, workflows, and management activities together. For instance, a corrective action for an incident can automatically prompt you to create a JSA for the job task involved, or to conduct an ergonomics assessment if a musculoskeletal disorder (MSD) injury happened. That can support more consistent processes, enterprise-wide analytics, shared administration, and better visibility and control of risk.

It also creates opportunities for more useful AI.

When AI capabilities are embedded within connected workflows and have access to appropriate organizational context, they can support decisions using relevant information rather than treating each activity as an isolated task.

Chat functions that understand your business context can make EHS information easier to find and use. Their value comes from connecting a question to the information needed to answer it.

Proprietary AI Models Give EHS Expertise a Safe Place to Grow

When you’re evaluating AI for EHS management, it helps to ask what the software provider controls. Does it develop and maintain models for specific EHS tasks? Can it improve those models when testing or expert feedback identifies a weakness?

That’s where a proprietary AI model can add value. A provider that controls its own model can shape its development around the problems EHS professionals need to solve, using relevant data, input from subject matter experts, and testing that reflects real workplace scenarios.

Consider potential serious injury and fatality (PSIF) identification. An incident that caused a minor injury may have had the potential for a much more serious outcome. A model developed for that task can be trained and evaluated to recognize relevant circumstances, helping investigators identify risks that the actual injury alone might obscure. Human reviewers can then assess those findings and determine the appropriate response.

Control over the model also creates opportunities for targeted improvement. The provider can investigate recurring errors, refine how the model handles specific scenarios, and evaluate changes against established EHS performance criteria. That gives product development teams a direct way to turn expert feedback into better tools.

Another major advantage of proprietary models is that since they’re not publicly accessible like the models of commonly available commercial AI applications, they don’t risk exposing potentially sensitive EHS data. For example, if you use ChatGPT, you need to send data to OpenAI, which is a layer of security risk. But if you use a proprietary model, you don’t need to have that security exposure. This is especially important if you consider that some EHS data, like ergonomics assessments, may involve sensitive personal information.

Of course, proprietary alone isn’t proof of quality. The value comes from what supports the model: relevant training data, expert involvement, rigorous validation, and ongoing oversight.

One important question to ask is: How does the provider’s control over its AI translate into more dependable support for your EHS decisions?

MCP Connects AI to Tools and Data That Make It Useful

Owning and developing an AI model is one part of the equation. Connecting it to the information and tools people use every day is another. That’s where Model Context Protocol (MCP) comes in.

MCP is an open standard for connecting AI applications to data sources and tools. In an EHS platform, those connections could help an assistant to access specific information, such as details about industrial hygiene management at the site.

MCP can work alongside a proprietary AI model, and in fact, it ideally should. Remember, MCP describes how applications connect to tools and data. It doesn’t determine who owns the models involved.

When you have MCP capabilities, you’re better equipped to efficiently and accurately retrieve information you need, whether it’s from external or internal sources.

Let’s consider external sources first. You may sometimes need reference information, such as occupational exposure limits (OELs) from agency websites, such as OSHA or National Institute of Occupational Safety and Health (NIOSH). You could ask a standard commercially available AI chat tool to that for you, but to do it, the tool needs access to the external website(s). That access is not typically via MCP, which limits the success of the information retrieval. But depending on the website, they may have built an MCP layer. If you’re equipped to communicate with that layer, your chat tool can connect to and access the database in a manner that is effective for your chat function to search through, which increases the success level.

How does MCP help you answer questions about your own data within the EHS software platform you use? Suppose you have a need a quick answer about your most recent IH sampling results. Without a chat feature, you’d need to answer the question by manually hunting through data, whether in hard copy or electronic form, which is very time consuming. With a chat tool unsupported by MCP, you’d potentially retrieve information faster, but might need to do a succession of searches or queries of increasing refinement to get the answers you’re looking for. But when your software platform has an internal MCP, you have the right mechanics to facilitate access to your information and quickly get an accurate and understandable answer to your query.

Contractor With Tablet A

Think about how these qualities of MCP link to earlier points about purpose-built AI with EHS context expertise and the value of chat functions that know your business context. When you have proprietary AI trained on real EHS datasets, a chat function that understands your business and MCP, you’re set up to have connected and accurate EHS data insights.

Generic AI Can’t Deliver Data Privacy and Security

EHS data can contain sensitive information about employees, injuries, chemical inventories, and business operations. Before putting that information into an AI tool, you need to understand where it goes, who can access it, and how well the platform protects it.

Start with a straightforward question: Will your information be used to train models available to other customers? Don’t assume the answer based on whether a tool is proprietary or general-purpose. Ask the provider to explain its policies on model training, data retention, and third-party access, including what happens to information submitted through chat functions or uploaded documents.

You need software that has security built into the platform and its AI capabilities. That includes encryption, role-based access controls, audit logs, and defined processes for managing security incidents. All the considerations above intersect with applicable privacy requirements, including the EU General Data Protection Regulation (GDPR) and the California Consumer Privacy Act, as well as broader cybersecurity practices.

ISO 27001 has become the world’s leading international standard for information security management systems (ISMS), jointly developed by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC). It provides organizations with a structured, risk-based framework for protecting sensitive information and continuously improving security practices across people, processes, and technology. NSF International Strategic Registrations (NSF-ISR) certifies organizations to ISO 27001.

When evaluating EHS technology providers, organizations should look for documented security and privacy controls, appropriate independent assurance, and clear AI governance practices. An ISO 27001 by NSF-IR certification is a good indicator that a software provider has done its due diligence on data privacy and security. Be sure to check the certification scope and how they apply to the services you’ll use.

Purpose-built EHS AI should make useful information easier to access while protecting sensitive data. The more context an AI tool can use, the more important it becomes to govern that access carefully.

Choose AI That Understands the Work

When evaluating AI for EHS, seeing the software in action is certainly important, but you need to go beyond what you see. Ask a potential partner how the capabilities perform within your team’s workflows.

What EHS data was used to develop and validate them? How were subject matter experts involved? Can users review and challenge recommendations? Do the capabilities connect across workflows? Can the assistant answer questions using your organization’s information while respecting access controls?

These questions help distinguish a useful EHS capability from an impressive interface.

At VelocityEHS, our approach to purpose-built AI centers on bringing EHS expertise, relevant data, and connected workflows together to support the people managing workplace risk. It’s why we designed and trained VelocityAI to deliver the reliable insights you need for specific tasks like incident management, JSAs, chemical management, contractor management and ergonomics. And it’s why Ask Vēlo chat delivers insights that are context-specific to your business.

Because the real measure of AI in EHS is what happens after the answer appears: a more complete investigation, a better-informed decision, or a more effective action to protect someone at work.

Reach out today to set up a meeting, so you can see what purpose-built AI on a true connected platform looks like.

Ai Built For The Reality Of Ehs Canva Cta