Gen AI RAG use cases that can deliver quick impact
A Gen AI Retrieval Augmented Generation (RAG) implementation combines the strengths of information retrieval and existing Large Language Models (LLM). RAG ensures that the output from these models is grounded in the knowledge base that is provided based on the use cases it’s expected to serve.
This ensures that models don’t hallucinate and is factually correct. In a typical workflow, the retrieval component searches the vector database for the most relevant information extracted from the prompt and adds the relevant information along with the prompt to the model to produce accurate and contextual relevant outputs.
A few examples of simple RAG use cases that can provide impact in 4–6 weeks
Generate insights from a corpus of documents
Derive insights from annual reports, existing brands reports, marketing documents, process, and research documents. This is the simplest and easiest RAG use case that can deliver immediate productivity boost to the teams that use this.
Voice of customer
Understand customer pain points, sentiment and triangulate issues and concerns that need to be addressed across a range of data sources like social media comments, reviews on marketplaces, company owned research, focus groups/customer panels, customer feedback on orders, customer satisfaction surveys etc.
In the past this used to be a rather large and resource intensive analytics exercise to be done periodically due to the effort involved. Today a good Gen AI agent implementation can provide on demand insights at a granularity (segment or cohort level) in an interactive manner with ease.
Customer Success management via Sales comments
In organizations with large sales forces as sales reps talk to accounts they uncover a range of issues like product interest, at risk behavior, concerns about specific product features, renewals etc. Recently we worked with a Fortune 50 enterprise (with a large sales force generating upwards of 200k+ comments across products and service lines) to implement a gen AI agent. This was able to provide HQ with immediate insights analyzing customer feedback data, to develop personalized action plan for each account at scale.
Medical Insights Miner (PubMed data)
PubMed data (36M+ citations, Biomedical literature) is one of the best publicly available databases with tremendous value to generate insights for pharma companies. A range of questions specific to therapeutic areas are covered in these publications. By leveraging RAG for answering researcher questions where gen AI agents provide the answers along with citations/evidence greatly enhance medicals affairs, Medical Service Liaisons productivity as they design new protocols.
Advanced RAG use cases (~4–8 weeks of deployment time)
L1/L2 Customer support assistant
Gen AI powered agent enabled better resolution mechanism for customer support team enhancing their productivity and reducing resolution time and 20% improvement in productivity. This use case includes querying from multiple structured data sources, integration with a system like Zendesk or equivalent (extraction of historical service tickets data, knowledge base, etc.), ability to provide citations of source and past similar resolutions to improve rep adoption, Q&A support in L1/L2 in customer service operations.
Compliance in customer contact centers
Enhanced Rep productivity, minimize adherence challenges for specific issues like PII data, abusive conduct etc. Large enterprises specifically in regulated industries are able to use gen AI agents to analyze rep behavior on an ongoing basis to correct compliance issues. This needs ability to handle audio data, in a compliant environment, PII handlers, and a continually changing context (e.g.; new brand guidelines) captured in RAG implementation to be effective. These agents typically also need to combine structured and unstructured data in some cases.
Employee knowledge training assessment for talent management
Analyze programs and training material to improve effectiveness for employees across various roles (sales operations, HR, etc.). These gen AI RAG implementations typically require combining training videos with tutorials with HR documentation and quantitative information about employee performance by training cohorts.
Global SOP Standardization and Legal documents
Improve contract performance significantly by addressing gaps in clauses and identify opportunities to identify deviation against policy which in turn can provide cost savings. This typically drives large impact in organizations with thousands of contract documents (ex: travel policy agreements, supplier terms, global vs local SOP protocols). Gen AI agents are very effective in quickly identifying variations in existing contracts and also help during renewal or new contracts being created at local levels. However these require advanced RAG which can absorb a variety of contract documents (often scanned) and also handle local regulatory or contextual provisions to come out with relevant insights.
Operations support assistant (manufacturing and maintenance)
Assistants to significantly improve speed of issue resolution and increase technical productivity: In organizations where complex machinery needs to be maintained on an ongoing basis (ex: Aircraft, manufacturing equipment, Hi-tech), we have seen fairly interesting gen AI implementations, where on the ground technicians need to access gen AI agents on a rugged device (like a custom tablet), provide voice input on the question, which in turn is checked against a repository of technical manuals which involve extensive technical drawings and information to find the answer. Further to this, most organizations also have prior resolution videos on how senior experts have solved the issue. In this case advanced RAG needs to index video data, retrieve the specific information in addition to the answer with citation.
What does it take to set these up?
These advanced use cases typically require an advanced RAG pipelines to improve accuracy and performance. These pipelines include:
- Advanced text extractors to handle more data formats (PDF, PPT, etc.), Dynamic Chunking, cross sectional context hierarchies to handle large corpus of data as well as relationships.
- A more precise contextual prompting layer and/or templates to identify a specific context window, using prompt enhancement techniques, using ReRankers to obtain more accurate context, using multiple indexes to identify data linked across different domains/ontologies, knowledge graph based relational search augmentation and evaluation through logging and LLM monitoring.
- These techniques will enhance the precision, relevance, and efficiency of the output of the model, adding more value than the typical RAG implementation. This layer of sophistication added into the RAG system gives more accurate results.
About the Author:
Gopi Vikranth is a principal at ZS and leads AI SaaS Products and Platforms, Customer Experience Analytics and Personalization.
Gopi, along with Arun Shastri, hosts a podcast called “Reinventing Customer Experience”. Gopi and Arun chat with digital, analytics, technology and customer experience leaders from various industries and organizations to understand how they are engaging their customers. The podcast is available on Apple and Spotify.
