Generative AI for efficient insight into society and working life
Generative AI will have a significant impact on the Norwegian business sector. For NHO (the Confederation of Norwegian Enterprise), it is important to explore both the opportunities and challenges related to the technology, as well as its use within their own organization. Together with Itera, NHO has identified several areas where generative AI can create value and developed two prototypes for information retrieval for NHO's employees and members.
Where can generative AI create real value for a societal actor like NHO? How can it provide immediate and accurate information based on thousands of data sources? How can hallucinations and imprecise answers be avoided?
In collaboration with Itera, NHO explored these questions and developed prototypes for two AI-driven digital assistants that will provide immediate and accurate answers to complex issues for NHO employees and members.
- The first solution compiles a large amount of data to provide NHO employees with answers to questions related to NHO's many viewpoints and opinions within society, technology, and working life. The solution enables the organization to quickly identify and communicate its positions effectively and uniformly to all stakeholders.
- The second solution enables NHO's many members to receive quick answers to relevant issues that an employer may face - all based on data from NHO's own employer portal, Arbinn.
Realising concrete value from Generative AI
NHO, like most large enterprises, has a large amount of data. Combined with generative AI, this formed a good starting point for creating value from the technology. NHO also had many ideas about where generative AI could be used.
The challenge was to qualify where the greatest potential for creating value lay and what data, processes, and organization were required to realize it. Finally, it was important that learning and development of the solution went hand in hand.
Two value-creating solutions
Two functional prototypes of smart assistants based on generative AI were developed. The solutions address different use cases, rely on different data sources, and target different audiences - both create direct value for NHO.
NHO has many viewpoints and inputs across society, politics, and business. It can be challenging to identify relevant viewpoints and positions from the many associations.
The first solution, therefore, compiles a large amount of internal and external data, allowing NHO employees to ask the question, "What does NHO think about..". The solution enables the organization to identify and communicate its viewpoints effectively and uniformly to all stakeholders.
The second solution targets NHO's members. The solution uses generative AI and data from NHO's employer portal to provide members with quick and accurate answers to various legal and non-legal issues an employer may face. The solution ensures members receive precise and up-to-date information. In addition, the solution can streamline the time usage of NHO's own lawyers.
The solution is developed modularly, facilitating the addition of more use cases in the future.
Avoiding hallucinations to ensure trust
Hallucinations refer to situations where the AI system produces information not based on facts or the original data the language model was trained on. In other words, it makes up content.
The ability to create entirely new content is one of the strengths of generative AI, but it can also be a serious problem when wanting to use a language model to provide precise, fact-based answers based on specific data.
Hallucinations can be a serious problem for several reasons:
- Misinformation: AI systems producing imprecise or incorrect information can lead to wrong decisions, especially in application areas requiring high precision, such as law.
- Trust: Frequent hallucinations can undermine the trust users have in the AI system. If users do not trust the answers they receive, the value of the system is limited.
- Security: In certain scenarios, hallucinations can be exploited by malicious actors to deceive the system and produce desired but incorrect results.
- Costs: Correcting and managing the consequences of incorrect AI responses can be time-consuming and costly.
With current language models, it is impossible to completely eliminate hallucinations in practice. It was therefore crucial to address and minimize hallucinations given by the solutions. This was achieved by reducing the model's creativity (temperature) and adapting the solution so that if it did not find an answer in NHO's dataset, it would rather return "I don't know" than make up an answer.
Photo: Moment Studio.
– Itera impressed us with their commitment, effort, and relevant expertise. This project and its deliverables are an important input for our continued work with AI.
Project Manager and AI-enthusiast, NHO
At the starting line for practical AI application
Through this project, NHO has both mapped and qualified the opportunities generative AI has to create value for employees and members. Two concrete prototypes can create tangible value, and challenges around both data access and hallucinations are addressed, ensuring trust and scalability in the future.
NHO's own exploration of AI is just beginning. Future plans include further development and scaling of the solutions created, as well as continued exploration and realization of new areas where AI can transform and create value.