Berggren | Building an AI-powered virtual IP expert

Recordly and Berggren virtual ai ip expert

Berggren is a full-service European intellectual property (IP) law firm with
extensive expertise and a vast database of IP-related materials and documentation. Their experts handle complex and highly specialized work that requires deep domain knowledge and precise language.

As large language models (LLMs) matured, Berggren wanted to explore how AI could support expert work without compromising quality, domain specificity or security.

Berggren partnered with Recordly to investigate how AI could utilize their proprietary data to assist their experts and improve productivity by analyzing and interpreting masses of IP documents in a secure way. This project was supported by Business Finland's Generative AI funding that enabled Berggren to explore (Gen)AI technologies within their domain.

In the initial phase of the project, we conducted various workshops with different stakeholders to identify the most impactful use cases for the client's internal processes.

For each selected case, an initial solution was drafted to establish a direction, refined through developer-IP expert collaboration, and then built out into a complete, production implementation. These solutions have been widely tested by domain experts and are in active rollout phases.

While LLMs are powerful tools for generating text, applying them to expert-level IP work presents unique challenges. This field requires precise terminology, domain-specific reasoning, and consistency across complex documentation. For AI-generated outputs to be useful in this context, they must align closely with the organization's knowledge base, professional standards and highly specific workflows.

Another challenge lies in utilizing Berggren's large unstructured datasets effectively. The data, spanning millions of documents and thousands of patent families, contains valuable information but exists across multiple systems in formats designed primarily for human use. Preparing this data so that it could be accessed and interpreted by an AI system required identifying relevant sources, understanding their structure, and building pipelines that preserved the meaning and relationships within the data while translating it into a format usable by language models.

Berggren’s field of work lives and dies by context. We invested significant effort in understanding the underlying processes and ensuring the AI had access to the right information at the right time so that its analysis was always based on relevant, well-scoped data. This demanded close collaboration with the client, whose deep process knowledge was essential in shaping not only how information was structured and delivered to the AI, but also what outcomes were expected from it. In addition, experts must be able to interact with the system easily and trust that the generated outputs support rather than disrupt their workflows.

The solution

Recordly designed and implemented a solution built on Microsoft Azure, hosted entirely within EU regions to meet data residency requirements. This ensures that all data, both Berggren's and its customers', always remains within Berggren's own environment and is never used to train AI models. The solution combines data ingestion, indexing, generative AI, and a user interface into a complete solution for AI-supported expert work.

The architecture combines structured processing steps with agentic capabilities to tackle multiple different business processes.

A large amount of effort across the solution went into defining precise processing rules and guidelines, as well as refining the knowledge retrieval from the indexed knowledge base before generating process outputs. This ensures that outputs are grounded in Berggren's own data rather than general training knowledge. All AI models are run on EU-based Azure OpenAI endpoints, and authentication is handled through Microsoft Entra ID, tightly integrated with Berggren's existing identity management infrastructure.

Recordly and Berggren virtual ai ip expert - photo owned by Berggren 2026 — Image owned by Berggren, 2026

Under the hood

The system was built with a FastAPI backend and a Next.js (React) frontend, each running in Azure App Services. Azure OpenAI provided access to LLMs. A pipeline was developed to ingest data from on-premise M-Files into Azure Blob Storage and an Azure PostgreSQL server.

Testing and evaluation played a central role throughout the project. Working directly with Berggren's domain experts in short iteration cycles allowed us to improve fast and keep development aligned with the most important business needs. A combination of evaluation metrics on a predefined test set and expert feedback, based on using the application in real work tasks, was key to improving the system and guiding further development.

Recordly and Berggren virtual ai ip expert 2 - photo owned by Berggren 2026 — Image owned by Berggren, 2026

The results and business impact

The impact on daily work has been tangible. Experts in Berggren's largest business line report that the solution frees up a significant portion of time spent on routine tasks such as initial data collection and synthesis, allowing them to focus on work that requires professional judgement. The applications are now in active daily use by Berggren’s experts.

Behind the applications, the project established data processing pipelines connecting Berggren's internal knowledge systems and databases to the AI layer. These pipelines are designed to be extended to future use cases, providing a scalable foundation as Berggren identifies new opportunities for AI-supported work.

Berggren's organizational capabilities were also strengthened. Domain experts gained firsthand experience of what AI can and cannot do, and what developing internal AI systems actually requires.

At the organizational level, the practical experience sharpened Berggren's ability to evaluate build-versus-buy decisions for future AI technology investments, grounded in real project experience rather than assumptions.

“Our experts are now able to put more focus towards their highest-value tasks and move away from repetitive work. The key was building something our experts actually want to use. When people can see exactly where the AI's answers come from and verify them against the source documents, trust follows naturally.”

— Jorma Vanhala, European Patent Attorney, UPC Representative

“Recordly helped us integrate AI into expert workflows in a practical and responsible way. With the solution, we can give our professionals better tools to work with the knowledge we already have. When our experts spend less time on routine work, they have more room to focus on what our clients value the most; deep, considered analysis.”

— Heikki Kinnunen, Head of Digitalisation & Operations Development, VP of Greip

Jorma Vanhala Berggren - image owned by Berggren 2026 (1)