Felicis backs Unsiloed AI in $7.5M bet on enterprise document intelligence

A fragmented stack of layered documents suspended in mid-air, with beams of light selectively illuminating key elements like numbers and charts, while the rest fades into shadow — symbolising intellig

Unsiloed AI has secured $7.5 million in seed funding led by Felicis to solve a growing bottleneck in enterprise AI adoption: the difficulty of extracting accurate data from complex documents. With AI agents increasingly deployed across finance, legal, and healthcare sectors, the inability to reliably interpret tables, charts, and multi-column layouts threatens to undermine automation efforts. Unsiloed claims its technology can achieve near-perfect accuracy by combining dual-stream vision models, heat map-based chunking, and industry-aware decoding.

Bridging the AI-document gap

Most large language models (LLMs) struggle with structured documents because they process text and visuals as a single stream, increasing the risk of hallucinations and misinterpretations. Unsiloed’s approach separates text and image tokens into distinct processing streams, allowing more precise handling of figures, mathematical expressions, and tables. The system then applies heat map-based chunking to focus computational resources on high-importance regions, such as financial figures or legal clauses. According to CEO Aman Mishra, this architecture reduces errors that could have serious downstream consequences in regulated industries.

The startup has already been adopted by customers in banking, legal services, and healthcare. One unnamed bank uses it for document processing, while a law firm runs it on-premise for contract and case file analysis. Bolna AI, a voice agent startup, uses Unsiloed to ingest knowledge bases. These early deployments suggest demand for reliable, auditable data extraction — particularly where compliance and traceability matter.

Performance claims and competitive landscape

Unsiloed reports achieving 99.78% accuracy across 84,000 data points extracted from 542 financial documents, outperforming Microsoft’s Azure Document Intelligence and an in-house tool used by the client. It also claims superior results on the olmOCR-Bench benchmark, scoring 88.0 versus 66.0 for Reducto and 39.9 for Unstructured. However, these results have not been independently verified.

The market for document intelligence is crowded, with established players like Microsoft and Amazon offering document AI tools, alongside specialised startups such as Reducto and Unstructured. Unsiloed differentiates itself through its dual-stream model and selective processing, which Mishra says also lowers operational costs — pricing starts at one cent per page. By focusing compute on critical sections, the system avoids the inefficiency of uniform processing across entire documents.

Strategic implications for enterprise AI

For businesses, the ability to trust AI-derived data is becoming a prerequisite for automation at scale. Unsiloed’s emphasis on traceability — showing where extracted data originates in the source document — addresses a key governance concern. This could make it particularly appealing in the UK, UAE, and Kazakhstan, where financial and legal institutions are accelerating AI adoption but remain cautious about black-box systems.

The $7.5 million round, which includes participation from Y Combinator, Orange Collective, and Uncorrelated Ventures, will be used to expand the team, enhance proprietary datasets, and scale sales. With only seven employees, the company remains lean, but its focus on high-accuracy, domain-aware extraction aligns with a broader shift: AI value is moving from raw model capability to reliable, context-aware application.

What it means

This funding round reflects growing investor confidence in AI infrastructure that improves data usability — not just model performance. For enterprise leaders, the lesson is clear: AI agents are only as good as the data they consume. Tools like Unsiloed’s could become essential in sectors where accuracy and auditability are non-negotiable. The risk, however, lies in over-reliance on unverified performance claims and the potential for vendor lock-in as proprietary datasets grow.

Unsiloed’s success will depend not just on technical superiority, but on proving long-term reliability and interoperability. In markets like Kazakhstan, where digital transformation is state-driven, and in the UAE and UK, where financial services dominate, demand for transparent, efficient document AI is likely to rise — but so will scrutiny.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *