Hamachi says its patent keeps client data out of LLMs

Hamachi.ai says it has received a patent for a method that masks sensitive client information before it reaches a general-purpose large language model, a system the advisor-facing platform describes as letting it run agentic AI while treating client records as restricted material. The co-founders behind it are a cross-section of advisory technology—Eric Clarke of Orion, Brian McLaughlin of Redtail, Mike Wilson of AdvisoryWorld and Mustapha Baassiri of Advizr—and Wilson, now the chief executive, argues that account statements, tax forms, trust documents, CRM records and portfolio data should not be casually uploaded into general-purpose LLMs at all, and that the protection has to sit upstream of the model rather than live in a policy document.

RIABiz, which reported the announcement, sets it against warnings about runaway AI from a departing Anthropic staffer and the isolation guarantees the large model providers extend to enterprise buyers, and the distinction Hamachi is selling is location: per the RIABiz account, the masking happens outside the hyperscalers' own environments, so the unmasked record never lands inside their systems. Babu Sivadasan, who spent two decades as executive vice president of engineering at Envestnet and now runs Jiffy.ai for wealth managers, said by email that he applauds Hamachi for taking privacy seriously and putting an architecture around masking data before an LLM processes it; his next thought is the one worth rereading—there are many ways to achieve that masking, and Hamachi holds a patent on its own approach. A patent covers an implementation; it does not certify an outcome, set a security standard, or move any part of the obligation off the RIA deploying the tool, and no two firms route client data through the same stack.

Where does the masking actually happen—inside the vendor's environment, at the model's API boundary, or before anything leaves the firm's own systems? What does the platform retain after a task closes, and do the terms permit client data to train a model? What happens when the AI feature rides a usage meter the practice never negotiated, or when agents begin touching client records with no human in the loop?

A masking layer that sits outside the hyperscaler is a defensible architecture, and the advisory pedigree behind it is real; the protection an RIA can claim is only as good as the answer to where the data travels and who answers for it. The client who asks where the tax return went will not be satisfied by a patent number.