Artificial Intelligence Stakes its Claim in EHS Management

February 8, 2022

ehsAI’s cloud-based platform applies NLP algorithms to interpret and extract EHS requirements from regulatory documents.
ehsAI’s cloud-based platform applies NLP algorithms to interpret and extract EHS requirements from regulatory documents. This output can be exported into EHS software systems such as Intelex EHS Management Software to manage requirements and tasks and ensure timely execution.

Artificial intelligence (AI) technology, in the context of EHS management, seems a perfect fit in an essential area: the need to wade through a morass of complex and ever-changing regulations, compliance, permits, consent decrees and other critical EHS documents. It’s a human-intensive and often error-prone task that, if done improperly, risks fines and other penalties for failure to comply.

Document deconstruction – to identify the actions that EHS professionals are required to take – is difficult, time-consuming, complicated and expensive because of the manual grunt-work need from many people to pore through and analyze EHS documents. 

Vancouver, Canada-based ehsAI has developed an AI technology based on machine learning and natural language programming (NLP) to automate EHS document deconstruction and analysis. Their cloud-based platform applies NLP algorithms to interpret and extract EHS requirements from various types of regulatory documents then converts unstructured data and extracts actionable compliance obligations written within these documents. This output can be exported into EHS software systems such as Intelex EHS Management Software to manage requirements and tasks and ensure timely execution. 

According to ehsAI Chief Technology Officer Mahdi Ramezani, the time was right for AI and EHS because the past 10 years have seen significant improvements made in machine learning and deep learning that now makes it possible for AI to process and analyze massive volumes of data.

AI Supports EHS Management Decision Making

Ramezani explains that AI doesn’t make decisions in document deconstruction but supports EHS decision making by flagging compliance action items. ehsAI’s technology uploads files, scans through compliance, regulatory and other documents, identifies important items and generates an indexed list of compliance action items. AI can quickly process and analyze documents much faster than humanly possible. In this case, the intersection of people and technology allows for the utilization of information created by AI to support human decision making.  

Many day-to-day human tasks that are not complicated can be easily replaced by training an algorithm or set of problem-solving computer instructions, Ramezani says. But the people part of AI in document deconstruction is essential, not just at the back end of decision making and action, but also at the front end of laying the groundwork for EHS algorithms. 

Ramezani says that necessary instruction for building an EHS algorithm requires basic details that people must input, including things such as defining and identifying numeric chapter numbers and citations in documents, and adding indications of new paragraphs, chapters or other items. The algorithm created has an instruction sequence of things that should happen when encountering a citation and if formats are consistent, and words spelled exactly the same, then things move along.

“The problem with (certain basic algorithms) is this can be too strict a rule and can’t be generalized for new scenarios, in the case of a typo or an abbreviation, for example,” he says. In this case, an AI module must be programmed to make certain assumptions based on things like where text might be placed or located in a document, what other text might surround certain words or terms or if a spelling error or abbreviation is encountered.

The AI module as it is working may encounter discrepancies or deviations and would then make a prediction or assumption for a correct answer. Every time a prediction is wrong, however, the module incorporates that lesson and learns to make a better prediction next time. 

“We use lots of samples—thousands of documents, hundreds of thousands of pages, perhaps a million samples of sentences—to help make the module learn and improve,” Ramezani says.

Teaching ML and NPL EHS Management Workflows

Prior to applying technology to the task, EHS professionals at ehsAI spent nearly two years mapping out how the machine learning (ML) algorithms and NLP of AI needed to understand the workflows of EHS managers responsible for and managing compliance as well as the requirements in near countless regulatory and compliance legislation. It was also essential to recognize how these might relate to other requirements in other documents.

“We literally had to write the playbook…the training manual for novice AI programmers to help them think like an EHS pro and appropriately create the algorithms,” said Margery Moore, the company’s chief executive officer.  She describes this lengthy process as a massive brainstorming effort that involved intricately mapping every regulatory process on white boards and tracing through the minutia of regulations and compliance documents, as well as fully understanding the workflow of EHS professionals. 

Literally thousands of decision trees were constructed during 24 months of manual EHS analysis work that couldn’t be performed by simply scanning documents because, as Moore explains, “so many are written differently…or poorly.” Data within these documents needed to be cleaned up, normalized and structured for consistency so that it could be read.

“Many documents are not written in a consistent format and even grammar can often be challenging since regulators who author regulations aren’t always writers,” she said. 

Tens of thousands of samples and past experiences are behind accurate predictions and assumptions. Certain basic things are relatively simple to program while others are much more complex and require the higher expertise of experienced EHS professionals working with AI developers.

At the end of the day, AI used in document deconstruction saves time and human effort. How much time is a difficult question to answer, according to Moore. 

“What we can tell you is it takes about 40 to 60 hours to deconstruct a Title 5 (Title V Operating) permit and our application can do it in about 3 hours,” she says. “It might take 10 hours to do a 60-page standard. But (our application) can probably do it in about 20 minutes.”

It’s the sort of work that consultants often do for large enterprises and they would bill by their time, Moore says. “But they don’t have people to read all of the documents,” she adds. “And it’s the consultants who realize that their lunch will be eaten shortly if they don’t use tools (like AI), which should be able to boost their margins, which are already really thin.”

Learn more about how AI can be applied in many other areas of worker safety and wellness by downloading the Intelex Insight Report: Keeping the Human in Artificial Intelligence.