At the Innovációra Magyar! conference organized by the Joint Venture Szövetség (JVSZ), KPMG Hungary reported on an experiment in which they developed a machine learning model to pre-screen corporate documents for compliance checks. The presentation was given by Ignácz Péter, a manager at KPMG.
Purpose and technical approach
The project aimed to build a system capable of preliminarily checking incoming corporate materials and comparing them against relevant business policies to determine whether a deeper human review was needed. The team considered the capabilities of large language models — the presentation mentioned OpenAI GPT4 and Meta Llama3 as examples — but the approach centered on a machine learning pipeline tailored to document processing.
The development took only a few weeks.
Results and performance
According to evaluations, the system achieved roughly 80–85% effectiveness, with variation mainly driven by the quality of the input documents. While initial drafts produced by the model were of lower quality than human-created references, its responses were generated in minutes compared to the days previously required for manual review and assessment. This acceleration allowed human experts to focus their time on cases needing more detailed, careful analysis.
Ignácz emphasized that the machine learning solution supplements — rather than replaces — expert work, providing a stronger starting point for subsequent investigations.
Challenges encountered
One of the hardest tasks was generating the appropriate regulatory questions: defining exactly which regulatory or business queries to check within documents. The team also found that clearly defined business problems are easier to handle with machine learning methods.
Risks and lessons for ML projects
The KPMG presenter highlighted three common problem areas in machine learning projects:
- issues stemming from data quality;
- unclear or insufficiently specified business requirements;
- weaknesses in collaboration between the business stakeholders and developers.
These factors affect model performance and project success, so KPMG’s experience indicates they require special attention during implementation and scaling.
Conference context
The Innovációra Magyar! event served not only as another stop on Hungary’s innovation map but also as a forum to showcase recent trends and bring together key players. Nearly 120 participants attended, including corporate leaders, government policy experts, and startup founders. Ignácz’s talk illustrated that AI already speeds up certain information-processing tasks and outlined practical applications of LLMs and other AI tools in everyday corporate work.
Conclusion
KPMG’s experiment demonstrates that machine learning can provide a fast pre-screening baseline for compliance documents with around 80–85% effectiveness and substantially reduce processing time. At the same time, the solution augments rather than replaces expert review, and success requires careful data management, precise business problem definition, and tight collaboration among project stakeholders.
