Introducing Business Sensitive Data Detection

Mavs AI can now use a fleet of small language models (SLMs) to identify data that is sensitive to the business, not just PII, and protect it without blocking the prompt or stripping the context AI needs to perform, allowing businesses to do deep knowledge work with AI.

Today we're announcing business-sensitive data detection, a new capability from Mavs AI that identifies business-sensitive information in AI prompts through meaning and context, rather than static pattern matching.

What is Semantic Data Detection

Unlike tools that match a fixed format or a file's label, Mavs AI reads each prompt, including the contents of attached files, to understand what it means and judge what is sensitive in the context of the business, in real time.

In a live prompt, Mavs AI can distinguish whether "Project Maverick" is an internal codeword or whether "maverick" simply refers to an animal or is being used as an adjective. Until now, the industry has struggled with a fundamental problem: tools built to prevent data loss look for structure, so they catch formatted PII but stay blind to the confidential business data that actually decides a company's fortunes. Mavs AI's new capability brings this data into scope, so enterprises can adopt AI for their most important knowledge work without fear of leaking confidential information or intellectual property.

How We Figured Out Semantic Data Detection at Runtime

Mavs AI runs a fleet of small language models that read each prompt the way a careful human reviewer would, judging sensitivity by meaning rather than structure. Because the models are small, they run inline on every prompt at sub-second latency. And because judgment comes from language, not rules, a company can describe in plain terms what is sensitive instead of training a model to match patterns. This closes the contextual detection gap that legacy tools leave open.

Developed by our data science and engineering team, the core capabilities of business-sensitive data detection include:

• Meaning over patterns: Sensitivity is judged from the meaning of the prompt, not a static keyword list or regex, so novel and never-before-seen confidential concepts are caught rather than missed.
• Coverage beyond PII: The capability identifies what is sensitive to the business, such as confidential project and deal codenames, unannounced pricing, M&A terms, unreleased roadmaps, and key accounts.
• File contents, not file labels: Mavs reads the contents of attached files the same way it reads a prompt, judging them by meaning rather than by origin, location, or access permissions. Sensitive values found in a file are desensitized before it is processed, which reduces false negatives.

More Prompts Processed, Not Blocked

Detection matters because of what it makes possible, not just what it catches.

Today a lot of high-value knowledge work never reaches AI, because there has been no safe way to put confidential data in front of a public model without sacrificing it. The only option is therefore to block the prompt.

With this release, we’re changing that. Prompts and workflows that used to be blocked can now be freely processed with business sensitive documents and conversations, without compromising on security.

Business-sensitive data detection is available in the Mavs AI platform. It is model-and platform-agnostic, working with OpenAI, Claude, Gemini, and other models, and can be deployed in the cloud or within a customer's private environment. It fits the workflows teams already run, so protection happens inline with the request.

To learn more about Mavs AI or to request a demo, please visit us at https://mavsai.ai/demo.

We started Mavs so the work that matters most doesn't have to stay off-limits to AI. This release is a big step toward that.

Ride with us,
The Mavs AI Founders