Unsiloed AI Banks $7.5M Seed to Turn Enterprise Documents Into Data
Unsiloed AI raised a $7.5M seed led by Felicis to structure enterprise documents for AI agents, claiming 99.78% accuracy on 84,000 extracted data points — all self-reported.
By Olivia Hart
4 min read

What's News
- Unsiloed AI raised $7.5 million in seed funding led by Felicis, reported November 18, 2026.
- Its parser scored 88.0 on olmOCR-Bench in a company-run May 2026 evaluation, versus 66.0 for Reducto Agentic and 39.9 for Unstructured.
- A company-reported test of 542 financial documents yielded 99.78% accuracy across roughly 84,000 extracted data points, independently unverified.
- The San Francisco startup, founded in 2025, operates with seven employees and prices from one cent per page.
- Customers include an unnamed bank, a law firm running the software on its own servers, and Bolna AI.
Unsiloed AI has raised $7.5 million in seed funding led by Felicis to build technology that turns enterprise documents into structured data for AI agents, TechCrunch reported on November 18, 2026. Y Combinator, Orange Collective and Uncorrelated Ventures participated, alongside angel investors Aarthi Ramamurthy and Beth Turner. Seven employees are chasing a market crowded with specialists and cloud giants.
CEO Aman Mishra said the proceeds will fund research and engineering hires, an expanded sales effort and proprietary datasets. That is a standard seed-round promise, and it comes with the standard caveat: the money buys runway, not customers.
Y Combinator's company profile lists Unsiloed as a San Francisco company founded in 2025 and a member of its Fall 2025 batch. Mishra serves as founder and CEO, Adnan Abbas as founder and CTO; the two developed the technology as former IIT Kharagpur classmates before joining the batch.
What does the platform actually do?
The product overview describes a platform that converts PDFs, images and spreadsheets into JSON and Markdown for large language models and AI agents. Three operations anchor the system:
- Parse, which preserves document structure in Markdown;
- Extract, which returns requested fields against a JSON schema;
- Split, which separates multi-document files and divides long documents into retrievable chunks.
Extracted fields can carry confidence scores and citations pointing to their locations in the original document, letting reviewers check uncertain values before downstream use. The platform connects to S3, SharePoint, Drive, Snowflake and document-management systems, with deployment options in a virtual private cloud, on premises or in an air-gapped environment.
The founders describe a dual-stream architecture that keeps extracted content in one stream and document layout and hierarchy in another, followed by a decoder that applies industry-specific terminology and outputs JSON or Markdown. Attention-guided heatmaps direct processing toward numerical columns, merged cells and section headers while preserving multi-page tables and relationships between clauses.
Mishra framed the underlying reliability concern in blunt terms: "The problem with LLMs is they're inherently probabilistic. There is going to be some deviation induced. There is going to be some hallucination involved, and that has a massive negative impact downstream." The pitch, in short, is determinism where language models guess.
How solid are the accuracy claims?
In Unsiloed's own May 2026 evaluation, its parser scored 88.0 on olmOCR-Bench, a document-parsing benchmark, against 66.0 for Reducto Agentic and 39.9 for Unstructured. The run used 8,413 unit tests across 1,403 PDF pages and an unmodified version 0.4.27 of the olmocr scorer, covering tables, mathematical notation, reading order, long or tiny text and degraded scans.
Every number in that paragraph came from the company itself. Unsiloed even flagged its own methodology's limits: strict string matching can flag harmless formatting or notation differences, while academic-oriented benchmarks can underrepresent charts, forms and other elements common in production documents.
A separate company-reported test on 542 financial documents produced 99.78% accuracy across roughly 84,000 extracted data points. The documents came from one enterprise customer, and the comparison covered major AI models, Microsoft's Azure Document Intelligence and the customer's in-house extraction tool. The results had not been independently verified.
Who is paying, and at what price?
Mishra said Unsiloed is working with customers in finance, legal services and healthcare. An unnamed bank uses the software to process documents. A law firm runs it on its own servers for contracts and old case files. Voice-AI startup Bolna AI uses it to ingest knowledge bases for building voice agents.
Competition comes from specialized document-processing startups Reducto and Unstructured, products from Microsoft and Amazon, and general-purpose AI models businesses already use for extraction. Mishra said pricing starts at one cent per page — for a 10,000-page migration, that is $100 before any volume discount. He attributed cost control to directing computation toward the parts of a document that need the most attention rather than processing every section uniformly.
The immediate test is conversion: Unsiloed must turn early deployments and self-reported accuracy results into broader enterprise adoption against specialist providers and far larger platforms. Seven people and $7.5 million buy a credible benchmark score; they do not yet buy an enterprise sales force.
Source: unite.ai
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,312 articles