AI-driven precision for compliance requirements

See how Enhesa’s in-house AI team has tailor-made specialist tools that improve regulatory understanding and compliance — from legal text to site inspection

Andrea Pennisi Marco Ramos

by Andrea Pennisi, Marco Ramos

Artificial intelligence (AI) has revolutionized language processing, especially in fields like legal compliance, where interpreting complex documents is critical. AI models such as Natural Language Processing (NLP) tools are employed to analyze, extract, and classify obligations, duties, and regulations from legal texts, transforming how compliance is handled. In industries like healthcare, environmental, and safety compliance, AI tools are used to analyze regulations, ensuring facilities and operations meet legal standards, thereby reducing risks and streamlining audit, monitoring, and reporting processes. These technologies enhance efficiency, accuracy, and regulatory adherence across diverse sectors.

Our AI-driven approach enables us to deliver precise insights and guidance, ensuring compliance is easier and more accessible, so our clients can focus on what matters most — running their businesses smoothly and confidently. At Enhesa, we’re leveraging AI to enhance compliance across industries through two innovative projects: NOMOS and Fusion Vision.

  • NOMOS focuses on extracting obligations from legal texts using advanced NLP techniques
  • Fusion Vision integrates AI to streamline the analysis of images, ensuring facilities meet regulation compliance

In this article, we’ll explore how we’ve developed these two AI projects to provide our customers with reliable, real-time tools to ensure their operations remain aligned with evolving legal standards, minimize risk, and improve efficiency.

Understand legal texts with NOMOS

NOMOS helps streamline the interpretation of legal documents, making it easier for businesses to ensure compliance by automatically detecting obligations, even when scattered across multiple sentences or hidden in complex “legal language”. Through NOMOS, we’ve seen up to 90% less time spent on legal discovery tasks, significantly accelerating the process. For example, what might traditionally take months to cover can now be completed within a matter of weeks, drastically improving efficiency in navigating regulatory landscapes.

 

Why is NOMOS needed?

Interpreting legal texts is challenging due to their complex language and the subtle ways obligations are articulated. Long, complex text and long sentences can affect the interpretability of the text itself, which becomes even more challenging with the presence of multiple legislations and languages that vary in structure and format.

The diversity and complexity of different legal systems around the world requires specialized approaches to process and interpret texts that are not only long and complicated but also framed differently across regions. This variation makes it difficult to create an all-in-one solution, as legal documents from different jurisdictions follow unique standards, further complicating the task of ensuring precise and efficient compliance analysis.

Moreover, the nuances in how obligations are framed contribute to making the text even more difficult to understand. An obligation in legal terms is a “legal constraint imposed by law”, often requiring precise interpretation. Traditional AI methods struggle to identify these obligations, as they are frequently spread across multiple sentences or buried within intricate legal structures.

 

How does NOMOS help?

NOMOS (Navigating Obligation Mining in Official Statutes) is an AI-powered method designed to extract legal obligations from complex legal texts. It uses a combination of Word Embeddings (WE) and Positional Embeddings (PE), together with Temporal Convolutional Networks (TCNs) to accurately identify and classify obligations.

  • Word Embeddings (WEs) focus on understanding relationships between words. For example: knowing “lawyer” and “attorney” are related.
  • Positional Embeddings (PEs) track where each word appears in a sentence, which is crucial in legal texts. For example, in the sentence “The company shall comply with regulations,” the word “shall” states an obligation.
  • Temporal Convolution Networks (TCNs) model sequential data to process these embeddings, extracting temporal relationships between words. The model classifies legal sentences by identifying patterns and obligations across lengthy and complex legal texts, providing accurate predictions using fewer parameters than state-of-the-art models, such as the BERT methodology. For example: Detecting the phrase “shall comply” across different legal contexts.

To effectively manage the variety of legal documents Enhesa works with thousands of legislations and hundreds of languages, NOMOS preprocesses the texts by removing irrelevant sections like footnotes, headers, footers, and preambles. These parts often don’t contain obligations and can increase “noise”, complicating the classification. By focusing on the essential content, we ensure a clearer and more accurate extraction of legal obligations. Additionally, to enhance its global versatility, we trained multiple AI models tailored to specific legislations and are integrating a multi-language tokenizer, allowing our system to analyze legal texts worldwide.

 

Reduced energy consumption

We prioritized environmental considerations throughout the development process of NOMOS, carefully optimizing the model to run efficiently on a CPU.  It therefore delivers real-time performance without relying on more resource-intensive hardware, reducing energy consumption and minimizing the environmental footprint of our solution, while still maintaining results that rival those of modern, generative AI systems.

You can learn more about the algorithms behind NOMOS in our academic journal publication.

Better site inspection with Fusion Vision

At Enhesa, we guide organizations through the complex, fast-changing landscapes of regulations, frameworks, and guidelines, aiming to create modern solutions to help our clients achieve — and even go beyond — compliance.

One such area is safety and organizational guidelines in workspace environments. For instance, an office might be assessed for proper signage, safe storage of equipment, or adequate ventilation, while factories are often scrutinized for adherence to safety protocols, correct usage of protective gear, or equipment maintenance standards.

Historically, these guidelines’ compliance is assessed by an auditor, which can have the following drawbacks:

  • Time-consuming: Audits can take hours, especially in large factories or sprawling office buildings
  • Subjective: Auditors may interpret guidelines differently, leading to inconsistency
  • Costly: Regular compliance checks require a significant investment of resources
  • Human error: Manual assessments are prone to oversight or inaccuracies. Humans are not typically experts in all areas of audits so may focus time on immediate dangers such as factory settings while ignoring the ever-important risks present in offices or shared workspaces

In recognizing the limitations of this scorecard-based system, the need arose for a more efficient, scalable, and consistent approach that could automatically assess an environment in real-time and eliminate human bias.

With the rise of artificial intelligence (AI), we have been creating a new project to automate this process: Fusion Vision utilizes an AI algorithm capable of analyzing images of offices and factories, identifying which guidelines need to be met, according to each jurisdiction that workspace belongs to.

This project consists of two core components: the Vision-Language Model (VLM) and a Retrieval-Augmented Generation (RAG) system.

 

Vision-Language Model (VLM)

The VLM interprets the image content — whether it’s the placement of safety signs, the arrangement of office desks, or the positioning of heavy machinery in a factory. These models are trained on vast datasets of labeled images, giving them the ability to “see” and “understand” what’s happening in a visual scene much like a human inspector would, and create a detailed content-aware description of the image.

 

Fusion Vision example: an open-plan office with desks and people working in the background

An example image used for Fusion Vision testing

 

Fusion Vision’s description of the image

This image depicts a modern office workspace. The room is well-lit, possibly by natural light from the windows and artificial ceiling lights. The space is equipped with multiple desks arranged in rows, each featuring desktop computers, monitors, office chairs, and various personal items and office supplies. There are also a few potted plants scattered throughout the room, adding a touch of greenery.

In the background, there is a glass partition, and beyond it, we can see a group of individuals gathered around a desk, seemingly engaged in a discussion or a collaborative task.

The office design appears to be open-plan, promoting an environment of collaboration and communication. The color scheme is neutral, with grays and whites dominating the scene, which gives the space a professional and clean appearance. The flooring is carpeted, likely contributing to noise reduction in the workspace.

There are a few people visible in the image, focused on their work, which suggests the photo was taken during business hours. The general atmosphere of the office is calm and orderly.”

 

We then transform that description into an embedding vector — because machines “understand” numbers and not words, per se — and feed it into our second component, a Retrieval-Augmented Generation (RAG) system, a type of Generative AI.

 

Retrieval-Augmented Generation (RAG) system

Our proprietary RAG based system takes the image description’s vector and searches a knowledge base (in our case, our proprietary regulatory database dating back 30 years across 400 jurisdictions) to find matching vectors through a similarity metric. This returns relevant documents, policies, or standards and generates a detailed explanation of whether the identified elements comply with the relevant guidelines.

For the image shown above, and for the French jurisdiction, our AI system found the following guidelines’ headings and code names:

  • 110501: General ergonomic requirements
  • 110502: Visual screen equipment
  • 110401: General workplace requirements
  • 110404: Lighting
  • 110403: Temperature
  • 110402: Ventilation
  • 110409: Workplace stress
  • 0903: Confined spaces
  • 110407: Rest areas

If the image had shown a cluttered fire escape, the RAG system would reference safety regulations, explaining the violation in terms of fire safety compliance, unique to the jurisdiction defined by the user. In a picture of an office space in the US it will find a completely different set of regulations than in China, emulating how an audit expert in each of those countries would act.

Advancing compliance management with AI

At Enhesa, we’re at the forefront of revolutionizing the role of AI in law and compliance. Through projects like NOMOS and Fusion Vision, we’re not only enhancing the accuracy and efficiency of legal obligation extraction and compliance verification, but also pioneering sustainable AI practices.

By leveraging cutting-edge technologies such as Natural Language Processing and Vision-Language Models, we’re shaping the future of compliance and helping organizations navigate the complexities of legal and safety regulations with innovative, scalable, and environmentally conscious solutions. Together, we’re driving the future of AI in law, ensuring a safer, more compliant world.

To find out more about how Enhesa uses AI to enhance the solutions and services we offer our clients, take a look at these additional resources…

Regulatory content and sustainability intelligence

Creating better compliance management with AI

AI is great for data-driven compliance solutions, when used correctly. See how Enhesa’s experts use AI processing to deliver excellent solutions to clients.

Regulatory content and sustainability intelligence

From data to decisions: Enhesa’s AI-powered compliance

See how Enhesa’s in-house AI team are pushing AI-driven enhancement to bring expert insights and knowledge to the world of regulatory compliance.

Regulatory content and sustainability intelligence

Enhancing regulatory content with AI-powered metadata

Learn about how Enhesa is using AI to unlock content discovery with metadata, auto-classification, and chemical entity recognition.

Regulatory content and sustainability intelligence

Overcoming bias in AI

Using AI in our solutions brings huge benefits to our customers, but it must be accurate and untainted by bias — find out how we do it!

Share