· via TechCrunch
Unsiloed AI raises $7.5M seed to turn complex enterprise documents into agent-ready data
Unsiloed AI raised a $7.5 million seed round led by Felicis to convert tables, charts and complex document layouts into structured data for AI agents, with each extraction traceable back to its source.

The funding
Unsiloed AI, a San Francisco startup building software that prepares enterprise documents for AI agents, has raised $7.5 million in seed funding. According to TechCrunch, the round was led by Silicon Valley venture firm Felicis, with Y Combinator, Orange Collective and Uncorrelated Ventures participating. Angels Aarthi Ramamurthy and Beth Turner also invested.
The problem it targets
Companies rushing to deploy AI agents keep hitting the same wall: much of their institutional knowledge sits in documents full of tables, charts and intricate layouts that language models struggle to parse reliably. TechCrunch notes that errors are costly everywhere, but the stakes are highest in banking, healthcare and legal work, where such documents drive daily operations.
Unsiloed's software dissects complex documents, pulls information out of tables, charts and other elements, and converts it into structured data that agents can consume. Each extracted item is also linked back to its location in the original file, so customers can verify outputs rather than trust the model blindly.
Co-founder and CEO Aman Mishra told TechCrunch that large language models are probabilistic by nature, and the deviation and occasional hallucination they introduce can have serious consequences downstream for businesses relying on the results.
Mishra founded the company with Adnan Abbas, a former classmate from the Indian Institute of Technology Kharagpur, and the pair went through Y Combinator's Fall 2025 batch after developing the technology.
Customers and technical approach
The startup already serves customers in finance, legal services and healthcare. According to TechCrunch, an unnamed bank is using the technology for document processing, and a law firm has deployed it on its own servers to handle contracts and archived case files. Voice AI startup Bolna AI is a named customer, using Unsiloed to ingest knowledge bases for building voice agents.
On the technical side, Unsiloed uses what it calls dual-stream vision models, which process text tokens and image tokens separately, paired with a technique it calls heat map-based chunking to identify the most important parts of a document. The system then concentrates computing power on figures, numbers and charts rather than treating every section of a page identically. A specialized decoder takes industry-specific terminology into account, helping the system distinguish terms that mean different things in different business contexts. Mishra said this selective allocation of compute also keeps costs down.
Benchmark claims, with caveats
Unsiloed shared results from a test on 542 financial documents supplied by an enterprise customer, reporting 99.78 percent accuracy across roughly 84,000 extracted data points. The comparison included major AI models, Microsoft's Azure Document Intelligence and the customer's in-house extraction tool, though TechCrunch notes the results have not been independently verified.
The startup also claims it outperformed competing document-processing tools on olmOCR-Bench, a public benchmark from Allen AI covering 1,403 PDF pages and testing how systems handle tables, mathematical expressions, scanned documents and complex layouts. Unsiloed reported a score of 88.0, versus 66.0 for Reducto's agentic system and 39.9 for Unstructured. Those tests were conducted by the startup itself.
A crowded field
Unsiloed is far from alone in the market. TechCrunch points to specialized document-processing rivals such as Reducto and Unstructured, platform offerings from Microsoft and Amazon, and the common practice of businesses prompting general-purpose AI models to extract information directly.
Pricing starts at one cent per page. The company currently has seven employees and plans to use the new funding to hire researchers and engineers, expand its sales efforts and develop proprietary datasets to improve the technology.
Why it matters
The round is a signal about where enterprise AI tooling is heading. The first wave of business AI adoption focused on generating text and answering questions; the next wave depends on agents that can take action, which requires trustworthy, structured input drawn from the documents where companies actually keep their records. Document parsing looks mundane, but it is becoming the bottleneck between AI ambition and AI execution, particularly in regulated industries where a misread number carries real liability. The fact that Unsiloed pairs extraction with provenance — showing exactly where every data point came from — reflects a broader shift toward verification as a core requirement for enterprise AI, not an afterthought. Felicis putting seed capital behind a seven-person team in a space contested by Microsoft and Amazon suggests investors see document-to-data infrastructure as a durable layer of the AI stack rather than a feature that general-purpose models will simply absorb. Whether Unsiloed's benchmark edge holds against larger incumbents will decide if that bet pays off.
- #ai-agents
- #startups
- #fundraising
- #document-processing
- #enterprise-ai