Expectations surrounding Generative Artificial Intelligence (GenAI) were enormous from the outset, but reality has now caught up with many. Although the potential of this technology remains undisputed, numerous projects are failing to progress from proof of concept (PoC) to market readiness. There are often hurdles standing between the vision and a market-ready GenAI product, such as a lack of reliability, insufficient quality, poor performance or excessive costs.
A look at the Gartner Hype Cycle shows that GenAI has passed the ‘Peak of Inflated Expectations’ and is now moving through the ‘Trough of Disillusionment’, where many projects get stuck. Using our web application qbs.ai, we demonstrate how typical challenges in GenAI projects can be overcome – and what success factors are needed to accelerate the journey to the ‘Plateau of Productivity’.
qbs.ai is an application we have developed in collaboration with the management consultancy kobaltblau. It takes the early stages of the strategic consultancy process to a whole new level: time-consuming research and the preparation of in-depth strategic analyses can be reduced from weeks and months to just a few hours. With qbs.ai, strategic potential across various industries, company sizes and strategic directions can be automatically identified and developed in just a few clicks.
The generated content is presented clearly and user-friendly within a web application, eliminating the need for tedious, manual research. Strategic insights are available in no time at all – this not only saves valuable resources but also significantly speeds up decision-making.
Whilst qbs.ai focuses on strategy consulting, the ideas and techniques are more general in nature. Whenever tedious research work is involved, it is worth considering the use of GenAI. With optimised prompts, the extensive knowledge of LLMs (Large Language Models) can be tapped into and utilised for one’s own domain. If required, this knowledge can also be combined with the company’s own data. Based on such a database, a web interface can, for example, be developed to present the content intuitively.
Possible use cases could include, for example:
As already mentioned, various challenges frequently arise when using GenAI. Typically, questions arise regarding the quality and reliability of responses, as well as performance and costs. In addition to GenAI-specific challenges, there are also classic issues, such as those relating to software architecture, UX or CI/CD.
Drawing on our expertise in the fields of software and AI engineering, we have identified and implemented the following solutions for qbs.ai:
In addition to traditional unit tests for software quality assurance, the ‘LLM and natural language’ aspect requires new approaches. Responses are non-deterministic, meaning that the same query typically yields different responses. Consequently, automated testing using traditional methods is no longer feasible. We are pursuing several approaches to address this issue.
Subject-matter experts evaluate the responses using random samples. This, in turn, provides input for optimising the prompts. At the same time, the semantic quality of responses can be tested automatically. The idea here is that another AI evaluates the quality of the generated content. This allows thresholds – i.e. target values – to be defined, such as ‘the generated content must match a reference example by 90 per cent’. This enables automated quality monitoring over time and thus provides the option to take swift action where necessary.
We parallelise queries and pre-calculate the results so that the end user sees answers almost in real time. By decoupling the generation of the knowledge base from its display via a web app, we enable an improved user experience. We keep a close eye on costs through dedicated monitoring.
The architecture of qbs.ai is based on a structured knowledge base, which is generated by GPT-powered processes on AWS (Amazon Web Services, cloud provider). This database is not populated ad hoc from a manual LLM query, but is systematically built up using specialised scripts. These scripts can be re-run manually if required. This ensures that the results remain up to date over time. An extension, such as generation via a scheduled task, can in principle be implemented as a simple cron job within the chosen setup.
The front-end and back-end are designed in accordance with best practices for software engineering and architecture. The web interface also presents the content in an appealing and user-friendly manner.
As with any professional software project, qbs.ai is operated across various stages. This enables the separation of development and production. Furthermore, automated build and deployment pipelines ensure that the latest changes are continuously integrated, tested and deployed to production.
As previously mentioned, the aim is to enable even more in-depth and personalised strategic analyses in future. This will involve combining the world knowledge of LLMs with company-specific information. Furthermore, analyses can be carried out over a defined period of time. This covers both the content-related level – such as the question ‘How do trends in an industry change over time?’ – and the technical level, for example, whether updates to LLMs or changes to prompts have a positive impact on the results. A professional setup such as this enables such enhancements to be integrated into the application quickly and easily.
qbs.ai impressively demonstrates how GenAI can efficiently generate added value when the technology’s potential is utilised correctly. Developing GenAI applications that are intended to reach market maturity requires not only AI expertise but, above all, experience in the field of software development. As GenAI becomes more professionalised, this can provide companies with leverage that creates long-term benefits.
Do you have any questions or require assistance?
You can find the Champion Story on qbs.ai and learn more about the possibilities of generative artificial intelligence for your business on our website. Please feel free to get in touch.