03 Dec 2019 How Natural Language Understanding improves speed and accuracy of Contract Intelligence

Contract Intelligence - the Lifeblood of Contract Management

When it comes to contracts, every business learns to deal with several important contractual facts of life. One is that contracts need to accurately protect the company’s business interests while adhering to acceptable legal practices and regulatory requirements in relevant jurisdictions. Another is that contracts must be faithfully administered in order to serve their basic business functions. Finally, contract management costs a lot. And it’s not just the cost if creating and administering contracts. It is the cost of creating and administering contracts inefficiently. Various industry estimates put the cost of inefficient contracting anywhere from 5% to 40% of the total contract value.

Where does all that money go? Some of it goes into researching terms and conditions to make sure new contracts are consistent with standard practices and dependent business activities. Some goes into simply understanding what contracts say so they can be administered properly. Some goes into losses due to missed opportunities or poor contract management. In fact, if you take a close look at contract lifecycle management, almost every aspect of it is rooted in a necessary but often under-recognized function. That function is contract analysis – examining and understanding the key terms and provisions of contracts. Contract intelligence brings AI capabilities to contract analysis.

The traditional approach to contract analysis, and the approach still used by most businesses today, is a manual one. At large enterprises, teams of lawyers spend tens or hundreds of thousands of hours annually interpreting existing contracts. It is often boring, repetitive work that is time-consuming, prone to error, and costly both in people-hours and business losses due to mistakes.

With all that digital transformation has to offer, why do businesses put up this archaic, labor-intensive approach to contract analysis? One reason is that in spite of tools like contract lifecycle management (CLM) software and artificial intelligence (AI) driven contract analytics, many contract intelligence solutions have limitations. But that is changing, largely because of an entirely new approach to natural language understanding.

Shortcomings of Most AI-Driven Contract Intelligence

The promise of AI-based contract intelligence is huge. Here are a couple of examples:

  • One limited experiment pitted lawyers against a commercial AI contract intelligence application to analyze five contracts. The lawyers took an average 92 minutes to perform the task with 85% accuracy. The software analyzed those five contracts with an accuracy of 94%, and it completed the task in 26 seconds.
  • JPMorgan Chase & Company implemented an AI contract intelligence system to analyze commercial loan agreements. They report saving 360,000 hours of contract review time in the first year.

As compelling as these cases are, only a small number of companies have adopted advanced contract intelligence solutions, and most of these are very large enterprises. However, there’s a lot of interest. Like many professionals, Lukas Müller, Legal Counsel Group IT & Enterprise Activities at Zurich Insurance Company Ltd, is very interested. “I have seen demos and they look promising, but we are still in the early stages.”

There are impediments. Simon-Pierre Pype, Interim Legal Counsel, sees the value of AI tools for contract intelligence, but cites these reasons for holding back: “Companies don’t really know the solutions that are on the market and to what extent they will help them; it will take them some time to find the right solution and implement the AI-tools properly, and of course there are budgetary restrictions.”

Surveys show there are two main reasons why companies are not rushing to adopt AI contract intelligence just yet:

  1. The solutions are too difficult to implement. Most AI contract intelligence systems are based on algorithms that perform statistical analysis of contract text, which includes words and word combinations. Training these algorithms to properly identify and interpret contract language requires feeding the system large numbers of documents and guiding it through a supervised learning process until it is able to perform at an acceptable level of accuracy. Companies with very large contract repositories have the most raw material to work with, but the process takes time. Many companies do not have the resources they need to go through a long implementation and training process.
  2. The solutions deliver unsatisfactory results. Some companies find the AI system delivers unsatisfactory results even after training. There can be several reasons for this. For instance, it could be the contract repository is not large enough to provide adequate algorithm training. Additionally, if a business handles many different kinds of contracts, it may be difficult to provide necessary training on all the different contract types for the system to achieve the desired performance accuracy in all possible contract situations. Contract changes present another challenge to some AI systems. Contract changes can degrade system performance until the algorithm is re-trained on the new contract language.

A fundamental characteristic of most AI language processing systems explains why some users get poor results. The basic problem is that most text analytics algorithms need more and more text to improve the accuracy of their results. However, there is a point of diminishing returns. As the contract repository grows larger, more processing resources are required to analyze all that data. You can reach a point where the cost of processing enormous amounts of text outweighs the value of incremental gains in accuracy.

Not all AI contract intelligence systems suffer from these shortcomings. A new approach to natural language processing – Natural language Understanding (NLU) – makes it possible to train the system quickly with only a small set of sample documents and still achieve high levels of accuracy.

Natural Language Understanding – A New Approach based on Semantics

The text analytics approach to contract intelligence works best in situations where the system is processing large numbers of very similar contracts. That’s because one of the challenges of Natural Language Understanding is that words have different meanings and get used in different ways. Lots of similar contracts have fewer variations in word meanings.

But in reality, large enterprises that manage many internal and external contracts covering activities in different jurisdictions and even different languages must handle a much larger range of linguistic ambiguity. This calls for a different kind of approach based on semantics.

This new approach based on a concept called semantic fingerprinting, works like the human brain, which is the most efficient language processing system of all. In the human brain, all sensory inputs, including language, are reduced to patterns. Meaning is derived through pattern matching. Semantic fingerprinting works in the same way. Text is ingested, and instead of being subjected to statistical analysis, it is reduced to a binary matrix pattern. The technical term for this pattern is a “sparse distributed representation”, but it amounts to a semantic fingerprint of that text. The system begins by creating semantic fingerprints of a large body of reference information, such as domain knowledge, that might come from textbooks. Once domain knowledge is established, the contextual meaning of additional text, such as contract text, is determined through comparisons of semantic fingerprints in a process called semantic folding.

In the case of a contract intelligence system based on semantic fingerprints, the training process looks something like this:

  • Create reference information. This is an unsupervised training process in which the system ingests relevant domain knowledge, typically the same text book content humans would study to become proficient in a subject area. It is not the same content the system will ultimately be processing. In a commercial contract intelligence product this step may already have been done.
  • Train the system to extract information from actual documents. Subject matter experts teach the system to perform specific processing tasks by providing it with the documents and examples of the expected outputs. This can be done using a comparatively small number of documents and doesn’t require any data scientists or AI experts.
  • Continuous learning. The system continues learning based on new ingested data and corrections made by actual users of the system.

This semantic approach to contract intelligence has several big advantages. First of all, it does not require a huge repository of contract data do deliver accurate results. Secondly, it is fast, and it requires fewer computing resources than traditional text analytics. Finally, the training process is much simpler and can be managed by the subject matter experts who will be using the system, even if they have no AI experience.

The Need for Better Contract Intelligence

Implementation and performance issues associated with AI contract intelligence have caused many people to stick with manual analysis. Ginny Penzell, Director of Risk Management for EOG Resources, sums it up in this way: “I have found through my 20 plus years of experience that no AI tools replace a well-trained in-house administrator.”

In many ways she is right. If you put a contract in front of a well-trained contract administrator, that person will do a good job of analyzing that contract. But if you put 1000 contracts in front of that person, performance will decline. On the other hand, a well-trained contract intelligence system based on semantic computing will analyze 1000 contracts quickly with better quality results because it never gets bored or tired.


back to opinions overview