This is a guest post for the Computer Weekly Developer Network written by Thomas Levi in his role as senior director of AI and ML at Agiloft.
Agiloft is customisable, no-code business process automation platform known for its Contract Lifecycle Management (CLM) solutions. Designed to streamline and automate complex workflows, Agiloft enables organisations to manage contracts, legal processes, service desks and other business operations with minimal technical intervention.
As explained by Andy Patrizio, CLM is a systematic approach to managing a business contract at each stage. CLM combines technology and business processes to streamline and automate tasks including the drafting, negotiating, procurement, executing, monitoring and renewal of contracts.
Levi writes in full as follows…
The criticality for companies to manage risk in the face of geopolitical, regulatory, or economic swings has never been more acute. Tapping into the limitless intelligence of contracts can help businesses understand who is responsible, who is liable, who is going to do what and at what cost. I’d argue that having that information at your fingertips is the greatest power of Contract Lifecycle Management (CLM) technology. Having a single solution to manage every obligation at the touch of a button utilising AI is the true strategic value of CLM, beyond just managing the operational efficiency of processing contracts.
For CLM, both large language models (LLMs) and small language models (SLMs) offer valuable capabilities. Choosing between them or combining them depends on the task’s complexity and desired performance. While LLMs are often employed due to their extensive parameters, some tasks in the contract review process are repetitive and simple and can be handled by an SLM.
Take the automatic identification and extraction of standard contract clauses like “termination”, “indemnification”, or “payment terms”. An LLM’s full power isn’t always necessary for these tasks. A specialised SLM, trained specifically for such clauses, can perform these tasks faster and at a lower cost.
SLMs for CLM
For example, you could employ an SLM trained to identify and extract all clauses related to “liability exclusively”… and when presented with a contract, the SLM would quickly locate all relevant instances, saving time and resources during the review process.
Similarly, an SLM could be trained to extract data on all payment schedules and deadlines. When applied to a set of contracts, this SLM would consistently and accurately pull out the relevant dates and financial information, automating a task that is often tedious and error-prone if done manually.
A hybrid approach – using SLMs for routine tasks and reserving LLMs for more complex ones – can often provide a sweet spot for optimising performance, ensuring accuracy and speed while controlling operational costs. Think about it: an SLM could tackle the initial screening of contracts, identifying standard clauses and flagging potential issues. Meanwhile, an LLM could provide an in-depth analysis of complex legal wording or drafting bespoke contractual provisions.
Intelligent routing
Intelligent routing can support the implementation of a hybrid strategy, but your AI system needs to accurately discern the complexity of a given task and direct it to the best-fit model. Routing can come down to various factors, including the type of contract
(e.g., procurement, sales, legal) or the type of clause being analysed. For instance, a contract related to intellectual property might be routed to a specialised LLM trained in patent law and licensing agreements, while a standard non-disclosure agreement might be processed entirely by a suite of SLMs.
The decision between LLMs and SLMs should not always come down to choosing one over the other but rather as a strategic deployment of resources based on the specific needs of the task. [As has been discussed at length already, we know that] LLMs excel at complex reasoning and tasks that require broad knowledge, like interpreting legal precedents or drafting nuanced clauses. SLMs, on the other hand, are perfect for repetitive actions like data extraction and identifying standard clauses. Thorough experimentation and comparative analysis will go a long way in determining which model, or combination of models, is right for your CLM needs.
One of the significant advantages of SLMs is their size. They are significantly faster to train, require fewer computational resources and are generally more cost-effective. That makes them ideal for on-premise or private cloud environments where data security and control are top priorities. That being said, there is no reason SLMs have to be limited to these environments.
The environmental impact of language models is another critical consideration. While SLMs typically exhibit a smaller carbon footprint, the overall sustainability of their deployment depends on factors such as the energy consumption associated with training and deployment. If many SLMs are required to replicate the functionality of a single LLM, the environmental benefits diminish.
SLMs are not without inherent limitations either. Their performance may degrade when confronted with tasks that deviate significantly from their training data and can be susceptible to biases. Therefore, it is essential to rigorously evaluate and test their performance to ensure their reliability and accuracy.
Specialised datasets
Thomas Levi, senior director of AI & ML at Agiloft.
The emergence of domain-specific LLMs introduces another layer of complexity. These models, trained on specialised datasets, can sometimes outperform SLMs in certain areas but often come with higher operational costs. Choosing a domain-specific LLM or an SLM depends on your CLM application and needs. For example, if your legal department needs to analyse complex contracts in corporate law, a domain-specific LLM might be the better choice. An SLM would be more efficient for more routine tasks like clause extraction.
SLMs are an excellent fit for more focused tasks with limited scope and complexity. If you need a chatbot to answer questions about contract terms, such as payment schedules or delivery dates, SLMs are perfect for the job. However, an LLM would be necessary if you want that chatbot to handle more complex queries, like interpreting legal language or assessing risk.
Ultimately, every industry should consider the strategic use of SLMs. The key is to align the model’s size with the task’s complexity. In CLM, this means leveraging SLMs for routine tasks such as extracting straightforward clauses and saving LLMs for deeper legal analysis and drafting. Businesses can optimise costs, enhance efficiency and improve their operational processes by focusing on the smallest effective model.