Generative Artificial Intelligence (AI) is transforming the approach to creating a wide variety of content, including text, images, melodies and designs. It is a disruptive technology with the potential to help businesses increase efficiency and develop innovative approaches. In today’s business world, text-generating AI in particular is highly relevant, which is why the focus of this article is on this aspect. In another blog post, we took a detailed look at the introduction of generative AI.
Current state of the technology
During the first half of 2023, the launch of new AI models followed at a rapid pace con week after week. Various vendors launched a wide range of products for businesses and consumers in a short period of time. Some of the models on offer, which were characterised by specialised functions and state-of-the-art technologies, were particularly in demand and expensive. In parallel, many open source models were released that were more widely accessible and customisable, allowing companies to tailor them to their individual needs. The tech industry is currently still in the process of identifying the models that are most efficient for the different use cases.
Furthermore, there is an expectation that models will continue to evolve based on the current rapid pace of development and applications will need to be designed to be adaptable to new models accordingly.
Use cases of generative AI in companies
Although this blog post is primarily intended to focus on the technical challenges of implementation rather than possible use cases, we would still like to provide a brief overview of some possible applications:
- Product Development / R&D: Generative AI can be used to teach engineers important information and best practices from projects they have already done, so that existing knowledge can be used instead of newly acquired.
- Marketing & Content Creation: Generative AI helps to create marketing materials such as blog posts, social media posts or email campaigns specifically tailored to the target group and efficiently.
- Sales: In sales, generative AI can be used to process objections from potential customers quickly and efficiently or to generate follow-up emails from meeting notes.
- Service: In times of staff shortages, generative chatbots can help reduce the number of staff needed by using internal expertise to answer customers’ questions.
- Human resources: Generative AI has the ability to analyse large amounts of CVs according to certain criteria in order to make a pre-selection for the recruiter. Generative AI has the ability to analyse large amounts of CVs according to certain criteria in order to make a pre-selection for the recruiter.
Combining generative systems with in-house systems
The key challenge in implementing a generative system is how to combine internal data with AI models while complying with the GDPR. There are three main challenges here:
- Consideration of current access rights
- Administration challenge: Do documents need to be actively uploaded somewhere or do the systems communicate via a persistent interface?
- Hosting: Where should the system run and does the company have sufficient resources and expertise to manage and maintain such a system?
Most first assume the need to re-train AI models for internal use – but this is not the case. The following questions would arise if this were the case:
- How often does the AI model need to be re-trained to stay up to date?
- What resources are needed for training (spoiler: there are many) and who pays for it?
- What data is used for training? Depending on the position/department, the data basis could vary.
Given the non-triviality of answering these questions or the lack of a good return on investment, the following solution offers a better alternative:
Enterprise search systems are designed to combine information from various internal company systems and display the most relevant information to employees. A standard requirement for such systems is of course the consideration of access rights. This ensures that each employee only finds the information that he or she is actually allowed to view.
By the way: When introducing an enterprise search, one is initially faced with a make-or-buy decision. Due to its complexity, an enterprise search is usually more of a buy decision.
How can an enterprise search be combined with generative AI?
When the user submits a question or prompt to the system, the enterprise search engine searches the most relevant information. This is then sent to a general (!) generative model to find the most relevant information. This process is called Retrieval Augmented Generation and the linked blog article provides further information on this process.
To get a feel for how generative AI works and its technical limitations when used with internal company data, you should also have a basic understanding of the technology.
Which generative model to choose?
When choosing a generative model, the first question is what happens to the data in the model. If you use one of the big providers like OpenAI, then the data is processed on their systems. If you use an open source model, you can also host it on your own instances – provided you have enough resources. If not, you can often fall back on providers that offer DSGVO-compliant hosting of such models.
Which model is chosen in an individual case, however, depends heavily on the use case. It is important, however, that the architecture is developed in such a way that models can be quickly exchanged when more suitable models come onto the market. By the way: at amberSearch, we always rely on self-developed or open source AI models as standard so that we can maintain the greatest possible independence. The turbulence at OpenAI in November 2023 shows that this appears to be a sustainable strategy.
Conclusion
The biggest challenge when introducing a generative system is the question of how to connect the internal data to the generative system with a low administrative effort but in a DSGVO-compliant way. Enterprise search solutions offer a very good approach here to apply generative AI directly across multiple systems.