Raj Bakhru of Datasite explains that the company stopped building its own document parsing technology due to the complexity and scale of unstructured financial data, opting instead to partner with Reducto to improve efficiency, quality, and focus on higher-value AI applications. This strategic decision, combined with leveraging proprietary data and specialized tools, enables Datasite to accelerate AI innovation, maintain compliance, and deliver advanced solutions for M&A transactions.
In this discussion, Raj Bakhru, Global CTO at Datasite, shares insights into the company’s journey with AI, particularly why Datasite decided to stop building its own parsing technology and instead partner with Reducto. Raj recounts his background as a technologist and founder of Blue Flame, a vertical AI startup focused on M&A, which Datasite acquired. Initially, Blue Flame built its own document parsing and ingestion pipelines, but as Datasite integrated the technology into its platform—handling over 60% of large transactions globally—the complexity and scale of unstructured data in virtual data rooms (VDRs) became apparent. Datasite needed a fast, secure, and compliant way to process billions of pages of sensitive transaction documents, which led them to leverage Reducto’s specialized parsing tools.
Raj explains that the decision to outsource parsing was driven by the numerous edge cases encountered in real-world financial documents, such as complex Excel spreadsheets with merged cells and macro-enabled files, which made in-house parsing costly and inefficient. The internal team found themselves spending excessive time on support and troubleshooting rather than focusing on higher-value AI applications like building advanced agents for diligence analysis and document querying. By partnering with Reducto, Datasite could accelerate development, improve quality, reduce costs, and allow their engineers to concentrate on creating revenue-generating AI products rather than reinventing foundational technology.
The conversation also touches on the evolution of vertical AI and the competitive landscape. Raj shares how Blue Flame initially aimed to grow independently but chose acquisition by Datasite to access unique proprietary data sets, including Datasite’s vast VDR content and Grata’s private company database. This combination provided a differentiated data advantage that enhanced their AI capabilities. Since joining Datasite, the team has integrated these data sources closely, improving model performance, reducing token costs, and enabling more sophisticated AI-driven workflows such as generating investment committee memos and PowerPoint decks.
Raj further discusses the strategic approach Datasite takes regarding build versus buy decisions. With a large engineering team of about 250 people, the focus is on dedicating top talent to core, revenue-generating projects rather than solving problems already addressed by specialized vendors. This approach maximizes efficiency and innovation, allowing Datasite to scale AI solutions rapidly while maintaining high quality and compliance standards. The company aims to grow revenue significantly by continuously launching new AI products and expanding their solution offerings, leveraging partnerships like the one with Reducto to handle foundational tasks.
Finally, Raj highlights ongoing challenges, such as optimizing AI model performance across cost, speed, and quality dimensions, and the complexity of benchmarking models for diverse document types. He mentions Datasite’s custom evaluation framework, VDRBench, which tests AI accuracy on real transaction documents and helps ensure their solutions outperform general-purpose models. The discussion underscores the importance of robust data extraction as the foundation for effective AI applications in finance, emphasizing that getting parsing right is critical to delivering high-quality answers and insights to users.