Although the banking, financial services, and insurance (BFSI) industry is keen to leverage generative AI, many organisations across this sector are still grappling with how to establish trust in the outputs from AI models. A central focus remains on developing generative AI applications that can provide accurate and reliable answers without human error or fabrication. To achieve this, we must reflect on other mature technologies and what we’ve learned from the past, alongside what valuable insights we can implement for the future.
One such technology is search engines.
Search engines demonstrate how vast data sets can expose the risks of fallible generative AI applications. The true value of generative AI — increased efficiency, productivity, and customer service — can only be fully realised when there is trust in the logical correctness of the application and AI results.
An AI system that is reasonably accurate is better suited for applications where pinpoint precision is less important, such as selecting ads for web displays. However, for critical functions involving sensitive data, such as processing a customer inquiry related to billing or managing employee benefits information, precision is not just a nice to have. It’s an imperative.
Reimagining search and other experiences with generative AI
As widespread as concerns over the potential dangers of generative AI may be, the technology is already transforming every area of our lives and the business world. It offers a set of techniques and practices for transforming data management systems that will substantially improve business value.
An example of this would be the enhancement of search and information retrieval capabilities through advanced language models that yield more precise responses compared to traditional models. This modernisation not only speeds up operations but also eliminates the need to manually sift through results.
Integration is key here, as only by embedding it into their data frameworks can businesses realise AI’s strategic value. That, however, starts with modernising legacy systems and utilising cutting-edge data processing technologies.
Sorting the wheat from the chaff: Enhancing data reliability
Generative AI can emulate the powerful capabilities of search engines for sorting through troves of data to find trustworthy sources. Using advanced language models, generative AI can effectively summarise and synthesise information from multiple sources, enhancing its ability to provide precise and consolidated answers without the need for users to manually sift through multiple search results.
This trusted ranking guarantees the provision of high-quality results, which not only helps with accuracy but also contributes to making responses from AI applications more authentic.
Large language models as sophisticated interlocutors
Large language models (LLMs) built to be expansive with the entire internet as their training data can typically handle all kinds of topics but with varying levels of accuracy due to the differing quality of their training data. As a result, developers should consider LLMs as powerful interfaces for understanding and instructing in natural language, rather than as a gold standard for knowledge.
Grounding LLMs with proprietary and thoroughly vetted data sets allows companies to significantly increase the accuracy of their LLMs, ensuring they are highly factual tools for engineering or product modelling when accuracy is mission-critical.
Businesses can now build applications that employ last-layer masking, which is highly customised for the specific distribution of data within a business. Using retrieval-augmented generation technology, these applications enable the rapid deployment of high-performing, custom last-layer masking solutions. This process facilitates the development of customisable AI functionalities without the burden of managing complex infrastructures.
Advocating for transparency and managing expectations
While generative AI holds significant potential, its adoption comes with considerable hurdles that require careful evaluation and tempered expectations. Building trust is essential, which involves clearly articulating how AI systems function and produce results. Users need robust evaluation systems to consistently assess the reliability and accuracy of AI outputs. These systematic evaluations are crucial not only for verifying the quality of these solutions but also for enhancing user confidence and acceptance of AI technologies.
As generative AI advances, businesses need to keep up with the rapid changes in algorithms and integrate ongoing learnings into their strategies. This iterative learning process ensures that AI applications remain efficient and up to date within an ever-evolving tech ecosystem.
A Gartner prediction highlights the importance of adapting to generative AI advancements: by 2026, three-quarters of businesses will use generative AI to create synthetic data for customer-facing policies, up from less than 5% in 2023. Realising strategic value from synthetic data in a domain that bars the use of real data is challenging. However, search engines have laid some of the groundwork for sourcing accurate responses from large volumes of data. Adapting the best aspects of search engines and making them scalable and elastic is key to unlocking trustworthy AI with enterprise data.