Twin1 AI, based in San Mateo, California, has emerged from stealth with a 20 million USD seed round (about 3.2 billion yen). What the company sells is a digital twin that captures each employee's knowledge and the way they make decisions. Law firms and financial institutions have been using it for more than a year, and Twin1 says it now absorbs 30 to 50 percent of the communications work that knowledge workers would otherwise handle themselves.

※1 USD = 159 JPY

Keeping individual context instead of the company average

Most enterprise AI systems ingest a company's documents in bulk and return the same answer no matter who asks. Twin1 starts from the opposite premise. Co-founder and CEO Dr. Lewis Z. Liu argues that in a knowledge organization the human is the atomic unit of knowledge, and that the point of AI is not to flatten individual expertise into generic output but to amplify and connect it.

Each twin is therefore grounded in the user's working context, including email, meeting records, documents and connected workplace systems, and it keeps evolving as that context changes. It also does not live in a separate window. Twins respond inside the tools where work already happens, such as Slack, Microsoft Teams, Outlook, Gmail, Google Drive and SharePoint.

Six layers of controls decide who gets an answer

An AI that speaks on someone's behalf carries as much risk of leakage as it does convenience. Twin1's answer is six interlocking layers of controls, combining rules-based and AI-based judgments and applying them to both human-to-human and human-to-AI exchanges.

Three things drive those controls: enterprise policy, permissions inherited from existing access settings, and human approval. The user decides which data the twin may read and which colleagues may query it, so information does not surface for anyone who was not already entitled to it. Choosing finance and legal as first markets, where entitlement rules are spelled out in contracts, looks like a direct consequence of building that layer first.

The Twin Network connects the twins to each other

Above the individual twins sits a coordination layer called the Twin Network. It lets a twin find the right colleague, gather permission-aware knowledge and hand work across teams. The design deliberately stops short of acting over the head of the person it represents.

For enterprises there is also a Model Context Protocol server. It gives outside AI agents and internal applications a single interface for pulling governed context from one twin or from the wider network and acting on it. Deployment options include SaaS, single-tenant and private cloud, a structure the company frames as avoiding dependence on any single model or infrastructure provider.

Law firms and banks as customers, Eigen alumni as founders

Twin1 spent more than a year in production with partners in legal, financial services and energy before going public. Named customers include the law firms Linklaters, Orrick and Dechert, the financial institution Customers Bank, and the energy company Aegis Energy. Orrick is both a customer and a strategic investor in this round.

The company was founded in 2025 by Dr. Lewis Z. Liu, Tom Cahn, Huiting Liu and Dr. Jonathan Budd. Three of the four came out of Eigen Technologies, the London document AI company that raised more than 80 million USD (about 12.7 billion yen) before being sold to SirionLabs in 2024. Several investors from that era joined this round as well.

Bessemer Venture Partners, Tribeca Venture Partners and Aramco Ventures co-led the seed. The money goes to hiring in San Mateo and London, go-to-market work and further platform development. Liu has told industry press that the company is working through a pipeline of more than 400 prospects and plans to follow the enterprise product with a self-service version for smaller firms and individual professionals.

Summary

Twin1's 20 million USD is a bet that runs against the standard enterprise approach of pooling company knowledge into one model. Keep knowledge attached to the person who holds it, and pass it to colleagues or AI agents only where existing permissions already allow. If that design holds up, the hours professionals spend on routing questions and relaying answers should visibly shrink. More than a year of production use, with demanding law-firm customers secured first, makes for a steadier starting point than a purely conceptual pitch.

※The image is for illustrative purposes only.