AI Contract Management

by Engineering | May 21, 2024

AI Contract Management

Data entry is a key task that Large Language Models (LLMs) can alleviate for users. Large companies, in particular, receive a vast number of contracts and may have dedicated teams to sift through them, search for specific sections or clauses, and assign them to colleagues. These teams also add notes, page and section numbers, follow-ups, and other annotations. Often, these teams might not be lawyers or native speakers of the contract’s language, leading them to spend significant time Googling or, more recently, prompting an LLM to understand a clause.

Our current technology uses custom models to quickly identify obligations and insights for lawyers to review. However, it doesn’t assist with data entry or search functionalities. To address this gap, we’re testing a new interface aimed at relieving this burden. This interface will allow users to chat with a section using predefined or custom prompts. Once we’re satisfied with its performance, we’ll move on to the next requested feature: using a combination of our models, large context LLMs, and Retrieval-Augmented Generation (RAG) to interact with an entire document or even a whole project full of documents.

NOTE: The images here are of a prototype. When we finish evaluating it, it will be styled to look like our existing app.

The review process starts just as it does in our production app, except that here, Mistral has been prompted to only look at indemnity-related clauses. We’ve then simplified our current extraction of insights plus sub-insights to ‘insights-only’ and an ‘Approve/Delete’ option. This is the new Step One of Review: the user gives immediate feedback about good or bad extractions, without going through the existing process of reviewing categories, actors, and frequencies.

Prototype Screenshot

When the user has a complete set of approved insights, they can move to Step Two. This is a new screen for sub-insights, and it’s here that we can relieve the load of data entry work. We’ve pre-filled it with some common columns, but they can be easily customized for a user’s contract management integration. Initially, the rows for a new insight are empty…

Step Two Screenshot

But once an insight is opened, the LLM is prompted to run a first-pass over the selected sub-insights—here that’s type, frequency, page and section numbers, and section title. The user can edit them, or ask the LLM to try again, and once they’re happy, they can be saved into the table above. This is a basic flow working with LLM; additional flows are needed for logging when the insight was saved and allowing it to be shared or signed off by a senior colleague.

Sub-Insights Screenshot

So far, we’ve looked at LLM capabilities for extraction, but we can also use its generative powers for chat. In the screenshot below, we are trialing 'open' chat about the particular insight; maybe the user has a very specific question or wants to know about other relevant sections in the contract, instead of looking at this clause in isolation.

Chat Screenshot

Red Flags

We’ve included two predetermined prompts, one being ‘red-flag issues’ that stem from this insight, but also in the context of the contract as a whole. This is looking at one clause, but you can read about our full-document “Negotiator” on our blog.

Non-lawyers and Non-native Speakers

One of the significant advantages of using LLMs in contract management is their ability to bridge language barriers. For multinational companies dealing with contracts in multiple languages, or for start-ups without in-house legal teams, LLMs can provide translations and summaries, making it easier for non-native speakers and non-lawyers to understand and manage these documents. This capability ensures that all team members, regardless of their proficiency, can contribute effectively to the contract management process.

Summary Screenshot

By leveraging LLMs, users can quickly extract relevant information from contracts, populate databases, and ensure that all necessary details are accurately captured. This not only saves time but also minimises the risk of human error, leading to more reliable and efficient contract management.

We’ll be posting updates as we roll out these new features to revolutionise the way we manage contracts.