AI Work Code Prediction for Legal & Accounting Efficiency

Enhancing Legal and Accounting Workflows with AI: Insights into Work Code Prediction

By:

Ryan Alshak

Legal and accounting firms face a turning point: adopt advanced AI to revolutionize timekeeping and billing or remain mired in inefficiency. This blog explores Laurel's breakthrough in work code prediction using Siamese Network Embeddings, showcasing a path towards unprecedented operational efficiency and strategic advantage. Join us as we delve into the specifics of our technology and its impact on the professional landscape.

Pioneering Precision: Siamese Network Embeddings in Timekeeping

Our innovative approach using Siamese Network Embeddings for work code predictions is a technical leap forward, driven by the need for greater accuracy and efficiency in timekeeping. By analyzing billable entries through advanced AI, we've developed a system that not only understands the nuanced requirements of different firms but also adapts to them. This method stands out for its ability to capture the complex relationships between various work activities and their appropriate codes, significantly reducing the margin for error and enhancing the fidelity of billing data.

Operational Excellence Through AI: Beyond Code Prediction

Every work code that Laurel's AI accurately predicts represents a significant step towards operational excellence. This precision eliminates the need for timekeepers to manually enter codes, reduces the oversight required by partners, minimizes finance's compliance checks, and lowers the chance of client billing disputes. This streamlined process enhances the experience for all involved and leads to better financial outcomes. With Laurel, timesheets become a review-only task, accelerating the billing process and directly impacting revenue recovery. Our technology ensures entries reach billing systems faster, enabling quicker turnaround to bills with unparalleled billing fidelity and quality, laying the groundwork for revenue increase or cost reduction through strategic analysis.

Technical Foundations: Shaping the Future of Billing Accuracy

Our technical exploration into Siamese Network models and contrastive loss serves as the foundation for a transformative user experience in timekeeping and billing. This advanced AI framework is specifically engineered to discern and categorize billable activities with a high degree of accuracy, directly translating complex data into actionable insights. This capability ensures that each work code prediction is not just a step towards automation but a leap towards operational efficiency and accuracy. The impact on the user experience is profound—timekeepers are relieved from the manual drudgery of coding entries, partners gain confidence in the accuracy of billing data, and finance teams benefit from streamlined compliance checks. Moreover, clients appreciate the transparency and accuracy in billing, fostering trust and satisfaction. By meticulously aligning our AI technology with the real-world needs of our users, we're not only enhancing the quality of billing data but also enabling firms to harness this data for deeper analysis and strategic decision-making. The result is a billing process that is not only faster but smarter, facilitating quicker turnarounds on invoices and improving revenue recovery through increased billing velocity. In essence, our technology does more than automate; it elevates the entire billing ecosystem, ensuring that firms can focus on their core work with the assurance of having a state-of-the-art billing process in place.

Embracing the Future: How Laurel's AI Solutions Transform Professional Services

Laurel's AI-driven technology is a cornerstone for firms striving in the digital age, offering more than mere efficiency gains. It transforms timekeeping and billing into precise, automated processes, directly enhancing operational performance and strategic decision-making. Explore our detailed technical paper to understand the tangible benefits and innovations at play. Don't let outdated manual processes hinder your firm. Visit Laurel's Resources to see how our solutions can specifically streamline your workflows and improve profitability. Embrace the technological shift with Laurel and redefine your approach to professional services.

Ryan Alshak

Founder and CEO

Ryan was previously a litigator at Manatt Phelps, where he was ignominiously awarded the country’s worst timekeeper award. While Laurel’s current vision is to automate timekeeping for all professional services, ultimately, Ryan is on a mission to make the world aware of its time.

Posted on

February 22, 2024

Using Contrastive Loss and Siamese Networks to Train Workcode Embeddings.pdf - Google Drive

Page 1 of 8

Work Code Predictions Using Siamese Network Embeddings

Authors:

Jim Howard

Laurel Data Team

Moulay-Zaidane Draidia

Laurel Data Team

Vincent La

Laurel Data Team

Alec Clowes

Laurel Data Team

For legal and accounting timekeeping, the Uniform Task-Based Management System (UTBMS) defines a series of work codes that are used to classify invoices. For Laurel, every billable unit of work (called an entry), requires codes. The code requirements and specifications for selection are unique to each customer.

In this article we cover how Laurel handles work codes, describe our approach to create work code embeddings, and provide some results using these embeddings to predict work codes from a narrative and initiative.

Background

To better understand work codes and how they relate to billable entries, we present a short background on Laurel’s approach to representing entries and workcodes in general.

Laurel Entries

Laurel has the ability to capture granular work information from a user. We call this granular work activities. An entry may contain any number of activities (including zero).

Activities typically contain an activity title plus some metadata about the activity such as the type of activity (email, document, meeting, etc). For our current approach, we use only activity titles, but could extend to using activity metadata in the future.

Combining all activity titles and computing the embedding from a pre-trained LLM and taking a convex combination with a pre-trained embedding on the narrative allows us to create a

Work Codes

The timekeeping industry relies heavily on the usage of work codes as part of billing processes.

These codes play an essential role in the organization of billable work by providing operational visibility to service charging firms, and by facilitating accountability for customers. Within the legal industry for example, the Uniform Task-Based Management System (UTBMS) is used to classify bills into discrete categories spanning a wide range of practice areas and tasks. These codes, by supplementing possibly lengthy natural language descriptions of the rendered services, constitute useful labels that facilitate quantitative analysis.

The UTBMS provides an overarching coding framework with a variety of code types. Each category can refer to a specific practice area (e.g. Patents, Bankruptcy, Litigation etc...) or to a particular kind of work. Each project has a defined set of possible code types, each having a list of agreed upon codes encompassing the scope of the expected work. It is common for service providing firms and clients alike to create custom codes to fit their needs, these codes can be firm, engagement, or even individual specific.

While specific coding practices can differ between firms, most firms require activity codes and task codes. Both activity and task codes have a well defined set of potential values. There exists one set of 28 activity codes (expanded in 2013 from the original set of 12). Task codes are broken down into a hierarchy from a defined phase (Litigation, Intellectual Property, Bankruptcy, etc) and specific task. Each phase has a different set of potential tasks. With most phases having between 20 to 50 valid task codes.

Sample Codes

Activity Codes

Litigation Task Codes

In practice, the breadth of possible codes combined with the tediousness of labeling each billable unit of work (entry) makes assigning work codes an administrative burden for timekeepers. This is the pain point we seek to address by automatically predicting codes for each billable entry.

The Problem

We want to build an accurate work code prediction model, but despite work codes having a well defined schema, most firms do not rigidly adhere to this schema. We’ve found that most firms either add custom codes, or only use a subset of the available codes. For activity codes, Laurel customers have on average 26 different selectable codes. Some firms use only the original set of 12 codes, while others add custom firm specific codes not present in the defined UTBMS set.

Sample Activity codes not in the UTBMS schema:

For task codes, the practice of using custom codes is even more pronounced. For a given matter there are on average about 47 potential task codes with some clients having plans with more than 400 potential codes.

Sample Bankruptcy task codes not in the UTBMS schema:

Because code schemas are not universal, Laurel continually ingests codes from clients, new codes may be added or updated at any time. It is not uncommon for thousands of new codes to be added daily. The main challenge to solve then is to create an accurate prediction of the work code from all information about the entry that is able to predict with little to no ground truth data.

Conclusion

Laurel’s innovative AI solutions are set to reshape the future of timekeeping and billing in the legal and accounting professions. With precision, speed, and the ability to adapt, our technology stands ready to empower firms, ensuring they can achieve operational excellence and deliver optimal value to their clients.