Why Datasite Stopped Building Its Own Parsing with Adit Abraham, Reducto & Raj Bakhru, Datasite

Vertical AI companies operating on highly sensitive, complex unstructured financial data (like M&A data rooms) should buy specialized document extraction/par

Why Datasite Stopped Building Its Own Parsing: A Case Study in Build-Versus-Buy for Vertical AI Infrastructure

Abstract

This synthesis examines the build-versus-buy calculus confronting vertical AI companies operating on highly sensitive, structurally irregular financial documents. Drawing on the operational experience of Datasite, a virtual data room (VDR) platform through which approximately 60% of M&A transactions exceeding $100 million are executed, and its acquisition of the vertical AI startup Blue Flame, the analysis argues that document extraction and parsing constitute a solved infrastructure problem best procured externally. This frees internal engineering capacity for retrieval-augmented generation (RAG) and agent systems that generate revenue. The methodology combines founder-level retrospective account with internal benchmarking data. Findings indicate that extraction quality, not answer-generation sophistication, dominates end-to-end system performance: a custom evaluation harness attributed a 22% improvement in recall rate to outsourced parsing. Practical implications include resource-allocation guidance for engineering leadership at domain-specialized AI firms handling high-stakes, low-structure document corpora.

1. Introduction

Vertical AI applications - systems narrowly specialized to a single industry's workflows and document conventions - have emerged as a dominant commercial pattern for large language model (LLM) deployment. Their value proposition rests on domain-specific data and workflow integration rather than on model innovation itself. This positioning creates an acute resource-allocation question: which components of the stack warrant internal engineering investment, and which should be procured externally from specialized vendors?

The M&A domain offers an unusually demanding test case for this question. A virtual data room (VDR) is the controlled repository through which transaction counterparties exchange diligence materials; its contents are heterogeneous, frequently spreadsheet-heavy, and classified as material non-public information (MNPI), triggering stringent regulatory and certification obligations. Any AI system operating over such content must simultaneously satisfy extraction fidelity, latency, cost, and compliance constraints, with failure in any dimension producing visible, high-stakes errors for finance professionals.

The central thesis advanced in this synthesis is that extraction quality is the binding constraint on downstream answer quality, and that specialized parsing infrastructure should therefore be bought rather than built - not because extraction is unimportant, but precisely because it is both critical and increasingly commoditizable. This argument runs counter to an intuitive assumption that answer-generation logic constitutes the primary engineering challenge in RAG systems. The analysis proceeds by establishing organizational background, examining extraction difficulty and proprietary data advantages, detailing security constraints unique to MNPI handling, and consolidating resource-allocation and evaluation methodology into actionable technical findings.

2. Background and Related Work

The operational perspective synthesized here derives from a career trajectory spanning quantitative research and trading systems at Goldman, Highbridge, and Kepos Capital, followed by the founding and sale of a cybersecurity firm to a private-equity-backed acquirer, and a subsequent role leading M&A and corporate strategy. Blue Flame, a vertical AI solution for the M&A sector, was founded approximately eighteen months prior to the account discussed. Its acquisition by Datasite occurred mid-fundraise: during negotiations over a venture capital term sheet, Datasite proposed acquiring the company outright rather than participating as an investor. The approximately 50-person Blue Flame team was subsequently integrated across all Datasite virtual data rooms.

Several frameworks organize the analysis that follows: a build-versus-buy decision framework applied to engineering resource allocation rather than procurement economics alone; a cost-speed-quality optimization triangle used to prioritize system trade-offs; and an observed barbell effect describing how AI-coding-tool leverage is distributed unevenly across engineering seniority levels. These frameworks collectively situate the Datasite case within a broader pattern of vertical AI firms reassessing which infrastructure layers merit in-house ownership as third-party tooling matures.

3. Core Analysis

3.1 Extraction as "All Edge Cases"

Prior to acquisition, the Blue Flame team built its own parsing, document, and ingestion pipeline under bootstrapped, cost-sensitive conditions. This experience surfaced the structural difficulty of financial document extraction: the domain is characterized, in the account's own terms, as one where "it literally is everything is an edge case." A representative failure mode involved a pie chart whose legend's shade of blue caused a leading vision model to misread the underlying values, even though the identical data rendered as a bar chart was read correctly. Finance spreadsheets compounded this difficulty further, featuring merged cells, macro-enabled XLSB files, and inconsistent template structures that generated near-unbounded combinations of parsing failure. The resulting support burden - private equity clients reporting incorrect RAG answers traceable to upstream parsing errors - motivated the decision to switch extraction infrastructure to a specialized vendor, Reducto, so engineering effort could redirect toward agent development rather than rebuilding already-solved extraction problems.

3.2 Platform Scale and Proprietary Data Advantage

Datasite's scale provides context for why extraction quality carries outsized leverage. The platform processes approximately 55,000 data rooms annually, serves over 500,000 monthly active users, and handles billions of pages of content, with volume having grown approximately 40% in the eighteen months following the Blue Flame acquisition. Datasite additionally owns Grata, a private-company database containing roughly 20 million private companies plus transaction and contact data used for sourcing and origination. Combining VDR content with Grata's proprietary dataset produced a data asset unavailable to any competitor, and consolidating this data closer to the models reduced both token costs and run times while increasing answer quality - illustrating that extraction and data-architecture decisions interact directly with the cost-speed-quality triangle described in Section 2.

3.3 Security Constraints Governing MNPI

Because data rooms contain the world's most sensitive transactional data, including MNPI, Datasite had never previously permitted content to leave its Azure-hosted platform, maintaining deliberately high walls around data room content. Integrating an external parsing vendor therefore required substantial diligence: significant time was invested with the vendor's team to ensure content protection and continued compliance with applicable regulatory regimes and client-mandated ISO certifications. This underscores that the buy decision was not undertaken casually but was conditioned on verifiable security guarantees commensurate with the sensitivity of the underlying data.

3.4 Resource Allocation and the Barbell Effect

Datasite employs approximately 250 engineers organized around five major revenue-generating initiatives spanning new products, consolidations, and new monetization models. The deliberate strategy of outsourcing solved problems such as parsing is explicitly tied to a growth target: achieving 20% annual growth on approximately $1 billion in revenue requires roughly $200 million in incremental revenue each year, which the account attributes to new AI products and expanded solutions rather than to reinvestment in already-solved infrastructure. A secondary organizational observation concerns a barbell effect in AI-augmented engineering productivity: AI coding tools confer outsized leverage - described as 10x or more - on senior engineers possessing strong mental models of the codebase, while junior engineers also gain meaningful leverage because LLMs can generate documentation and scaffolding on their behalf. Mid-level engineers were observed to benefit comparatively less from the same tooling.

4. Technical Insights

Several implementation-level findings merit direct documentation. Datasite's data science team measured a 22% improvement in recall rate attributable specifically to adopting Reducto's parsing output, providing quantitative support for the thesis that extraction quality dominates downstream performance. The production architecture feeds Reducto's extraction output into Turbo Buffer, used as the vector database, with graph RAG queries executed via Python code layered on top of the vector store and parsed content. The ingestion pipeline must accommodate large, unpredictable bursts of content, including up to a terabyte of data submitted on short notice by investment bankers, implying throughput and reliability requirements beyond simple batch-processing assumptions.

To evaluate system changes rigorously, Datasite built a custom benchmark, VDR Bench, comprising a representative virtual data room of several hundred documents paired with several hundred questions of known answers; this benchmark is rerun on every system change and is reported to outperform frontier labs' direct performance when the same files are dropped into general-purpose tools such as Claude Projects or OpenAI Projects. Datasite is considering open-sourcing VDR Bench and potentially co-authoring a paper on it with OpenAI. A further operational trade-off noted is inconsistent agent-level tool-call behavior across model variants - one cheaper model exhibited "tool call spamming" relative to Opus despite the cost differential, illustrating that cost savings at the model layer can be offset by inefficiency elsewhere in the agent loop.

5. Discussion

The Datasite case suggests a broader pattern relevant to vertical AI companies generally: as document parsing and extraction infrastructure matures into a specialized, vendor-serviceable layer, firms competing on domain expertise gain more from redirecting engineering talent toward agent orchestration, retrieval architecture, and proprietary data integration than from maintaining in-house extraction pipelines. This is particularly salient in domains such as finance, where document heterogeneity (merged cells, macro-enabled spreadsheets, chart variants) produces a long tail of edge cases disproportionate to the engineering effort available at any single vertical AI company, however well-funded.

A notable tension in the account concerns security: outsourcing extraction for MNPI-classified content requires extensive vendor diligence, suggesting that the buy decision is not universally available to all data-sensitive verticals without comparable compliance investment. This raises an open question for future investigation - namely, how smaller vertical AI firms without Datasite's scale or vendor-negotiating leverage can achieve similar security assurances when adopting third-party parsing infrastructure.

The barbell effect in engineering productivity and the emphasis on custom benchmarking (VDR Bench) both point toward an industry trend of treating evaluation infrastructure as a first-class asset, potentially as valuable as the production system itself, since benchmark fidelity determines whether build-versus-buy decisions and model-switching decisions are made on sound evidence.

6. Conclusion

This analysis demonstrates that in document-intensive vertical AI applications, extraction quality functions as the primary determinant of downstream answer quality, reversing a common assumption that most engineering value resides in answer-generation logic. Datasite's experience - quantified through a measured 22% recall improvement following a switch to specialized parsing infrastructure - provides concrete evidence for prioritizing procurement over in-house development of solved extraction problems, conditional on rigorous security diligence for sensitive data. Practically, engineering leadership at vertical AI firms handling complex, high-stakes unstructured data should weight build-versus-buy decisions using outcome-based benchmarks such as VDR Bench rather than cost or build-speed alone, reserving internal engineering capacity for the domain-specific agent and retrieval systems that most directly drive revenue.

Sources

• Why Datasite Stopped Building Its Own Parsing - Adit Abraham, Reducto & Raj Bakhru, Datasite - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub