Exclusive: ZeroDrift applies small language model to prevent AI-generated compliance violations
SiliconANGLE Paul Gillin
ZeroDrift launched a tool that checks AI and human messages for compliance before they go out. It uses a small model to cut delays from seconds to milliseconds, which is the whole point.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
ZeroDrift is trying to catch regulatory mistakes before they become somebody else’s problem. The startup just launched Command, a control plane that scans human and AI-generated messages for issues like promissory claims, unauthorized disclosures and references to material nonpublic information before those messages are delivered.
That matters most in finance, where a careless line can turn into a compliance headache fast. ZeroDrift’s example is the sort of thing regulators hate: “This investment is guaranteed to return 12% annually.” The company says its system can either flag a message for review or rewrite it so it fits the rules.
Command is built from four pieces. Policy Studio turns regulations and internal policies into rules. Anchor, a small language model, checks content for possible violations. Guard sits at the entry and exit points of an organization, and Command keeps the record of what was found and why.
The platform can pass, warn, repair or block a message. It rewrites AI-generated text automatically, but not employee-written text. A human-authored message can instead be blocked, approved by a compliance officer or sent back with suggested edits. There’s also a sandbox for testing and an Activity Center that shows the flagged content, the rules involved and the final decision.
ZeroDrift says the performance case is a big part of the pitch. Aroomoogan said early approaches using Anthropic’s Claude could take more than 10 seconds to detect and rewrite each violation. With the specialized model, the company says it got that down to milliseconds and is now under one second for detection and rewriting. The preview version of Anchor 3.0 is claimed to find at least twice as many violations as general-purpose frontier models, though ZeroDrift has not released detailed detection or false-positive numbers yet.
The company is evaluating the model against 100,000 human-annotated emails. It also says the system can be used through an API, tied to specific regulations and policies, and integrated with Microsoft and Google email systems as an SMTP relay. WhatsApp, SMS, Slack and Microsoft Teams are next on the list. Pricing is token-based, but ZeroDrift isn’t sharing the details.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring AI: not a chatbot with a swagger problem, but a filter that keeps firms from emailing themselves into a fine. The real story here is speed, because compliance tools that arrive after the message is gone are just expensive apology machines. If AI is going to write more of the business world, someone has to stand at the door with a red pen.
Read more about this at: SiliconANGLE