ZeroDrift Inc., a startup that automates compliance for artificial intelligence communications, today introduced Guard for Agents, a service that lets developers turn written company policies into enforceable rules to check AI agents’ communications before they reach customers.
The launch extends the company’s compliance technology into agent workflows through an application programming interface and connectors that import policy documents from Notion Labs Inc., Linear Orbit Inc. and Google LLC’s Drive. ZeroDrift charges one cent per validation, which can cover messages sent via email, text message or Slack post.
The product builds on Command, the company’s control center for viewing, testing and managing compliance enforcement. Guard gives developers a way to embed those checks into applications, addressing the gap between policies stored in documents and the messages agents generate.
“We sit at the pre-send layer,” ZeroDrift founder and Chief Executive Kumesh Aroomoogan said in an interview. “At the generation of the message, we’re intercepting.”
Developers send an agent’s output to a validation endpoint. The service checks the text against rules and returns information about any violation, including its severity. Depending on the policy, it can supply a replacement, recommend blocking the message or escalate the issue for human review.
The controls operate on outgoing responses rather than incoming prompts. That makes the service a check on what agents communicate, with customers determining how violations should be handled.
For example, a message containing confidential financial information or promising a guaranteed investment return could be blocked. Customers can also configure the service to handle violations without human intervention.
The policy connectors use Model Context Protocol to import documents into ZeroDrift’s enforcement system. Its proprietary small language model identifies rules in the text, such as a prohibition on transmitting Social Security numbers. Customers can inspect the extracted rules in a visual interface and select which to enforce and how.
“Our small language model is able to figure out the rules from all those texts and which rules should we should enforce,” Aroomoogan said.
The company says the approach avoids requiring developers to rewrite policies in a proprietary language. Rules can be enabled or disabled while applications are running and configured for agents. ZeroDrift logs enforcement decisions, including the original message, the rule triggered, the action taken and the resulting output.
Each check uses Anchor 3.0-Preview, a model the company designed to detect policy violations and produce compliant rewrites. Aroomoogan said ZeroDrift post-trained an open-source model using attorney-labeled data. The company also uses customer metadata and decisions to improve its model.
Guard includes an OpenAPI specification that developers can use to generate clients in languages such as Python, TypeScript, Go and Java. Integration requires a request to the validation endpoint.
ZeroDrift sees interest expanding beyond written communications. Aroomoogan said companies developing voice agents are seeking checks on what systems say to customers, particularly in financial services and healthcare.
AI startups selling into regulated industries are another source of demand. “People need to move fast when it comes to startups, and they just want to use us to plug in as a compliance layer,” he said.
The technology can also check communications from conventional marketing and customer relationship management systems, without a language model. Automatic rewrites apply to machine-generated content, not human-authored messages.
Aroomoogan said checks run in just milliseconds, but did not provide latency measurements or independent performance results. ZeroDrift plans to improve its model and conduct third-party benchmarks against other language models. Those tests could provide evidence for its claims of faster, less expensive validation than using a general-purpose model to review another model’s output.
