Creating a Pharma-Centric AI Agent for Commercial Q&A
In this case study, we highlight a recent project exploring generative AI capabilities over complex tabular data in pharma, where Clarkston’s Data + Analytics team created a pharma-centric AI agent. Read a synopsis of the project below or download the full case study.
Download the Pharma-Centric AI Agent Case Study Here
Clarkston recently partnered with a long-standing biopharmaceutical client to develop a completely new application that better equips the client’s sales representatives with key Healthcare Professional (HCP) information and activity across their drug portfolio. Proper querying formerly required expensive, ad-hoc joins and complex filtering criteria. Even for those with technical expertise, generating such queries was cumbersome and time-consuming, with many more questions unanswerable due to the way the data was structured or limitations of the query engine itself.
The client engaged Clarkston to help streamline this process. Clarkston’s pharma-centric AI agent solution now allows for the democratization of newly gleaned insights and critical, up-to-date information across the sales team, with quick responses to complex questions and at-a-glance targeting of HCPs.
Note: An AI agent can be defined simply as a Large Language Model (LLM) with access to a tool, in this case the ability to perform tabular queries in a programmatic way, like how humans utilize SQL but with considerably more speed and precision. The rep provides the AI agent with a question (or objective) in natural language, and it is then able to reason through the underlying data using the tool, ultimately providing a helpful and accurate response in natural language, along with requisite visuals such as a table or graph.
To begin the project, Clarkston’s business analysts, in consultation with executive leadership, generated a completely new schema to properly structure the data in accordance with business processes and objectives. Following this initial phase, Clarkston’s data scientists performed extensive data cleaning, data wrangling, and feature engineering to ensure the data conformed to the schema. Clarkston’s AI engineers were then brought in to create and properly ground the AI agent on the data along with the client’s internal processes and nomenclature. Finally, a thorough evaluation of responses was performed following iterative prompt engineering.
After extensive testing in a controlled environment, Clarkston’s data engineers and architects were able to connect the requisite tech stack in a more user-friendly environment for sales reps to use, integrating it seamlessly with the client’s existing solutions and dashboards. The model was deployed to production as a fully-fledged application, complete with a streamlined user interface, Single Sign-On (SSO) access, user access control, guidance, and documentation. Additional enhancements included quick actions such as retrieving a previous question for modification, the ability to generate downloadable and searchable/filterable tables and graphs, the option for the user to submit feedback, a toggle to view the AI agent’s thinking steps for validation, and a dynamic prompt library with unique responses based on the user’s geographical location and team. With this new application, sales reps get relevant responses specific to their area, and the option to toggle the feature off for a more holistic, generalized view. The success of the pharma-centric AI agent has been huge for the client, and performance logging, metrics, and user feedback will continue to shape and improve the application.
