Exclusive: ZeroDrift applies small language model to prevent AI-generated compliance violations

Investment firm ZeroDrift Inc. has introduced Command, an AI-powered compliance platform designed to identify and rectify regulatory violations in both human and AI-generated communications. The startup aims to address the challenges posed by the increasing reliance on AI agents for communication, which generates vast volumes of messages that are difficult to manually review.

Command operates as a control plane, automating the detection and correction of regulatory issues before messages are sent. It applies corporate policies and industry regulations to emails and AI outputs, flagging problems such as misleading claims, unauthorized disclosures, and nonpublic information references. ZeroDrift positions Command as an alternative to traditional compliance systems that review communications after the fact or require extensive human intervention.

The platform intercepts communications in real time, comparing them against a database of regulatory requirements and company policies. If potential violations are detected, Command can flag them for review or optionally rewrite the content to ensure compliance. The system consists of four components: Policy Studio, which converts regulatory documents and internal policies into enforceable rules; Anchor, a small language model that analyzes content for violations; Guard, which applies policies at the point of communication entry or exit; and Command, which records decisions and explanations.

Command can pass, warn, repair, or block messages. Financial regulators hold companies responsible for AI-generated communications, just as they would for messages written by employees. ZeroDrift's Command includes a sandbox for developers to test messages and view violations, proposed corrections, and differences between original and revised text.

The Activity Center provides an audit trail of flagged content, invoked rules, and final dispositions. The platform can be accessed via API and integrated with Microsoft and Google email systems as an SMTP relay, inspecting inbound and outbound communications. ZeroDrift claims its specialized Anchor 3.0 model detects at least twice as many violations as general-purpose frontier models while reducing latency to milliseconds.

However, the company has not released detailed detection or false-positive results. Command is a risk-reduction layer, not a guarantee of violation detection, and the company does not guarantee any outcome. ZeroDrift plans to extend Command's coverage to WhatsApp, SMS, Slack, and Microsoft Teams, with pricing based on input and output tokens processed.