A new discovery feature identifies passports, driver licences and other high-risk documents by their visual layout. It does not read the text on them, and it keeps working when the image is blurred.
The vision model capability has been added to the Concentric AI Semantic Intelligence platform. The company describes it as an industry-first for a data security governance product.
It analyses the content of an image and identifies the document type from visually consistent characteristics. Targets include the biographic page of a passport or a driver's licence.
The underlying premise is that visual information in a data record may carry more signal than the text does. Organisations are obliged to protect all the sensitive data they hold. In practice that data is spread across environments in many formats, which makes discovery the hard part.
The pitch is a cost and reliability argument against optical character recognition. OCR consumes considerable computing resources. It must process every pixel in an image to detect characters and convert them into machine-readable text.
OCR also struggles to detect information accurately in low-quality images. Concentric AI notes that an attacker using AI could still extract enough from those same images to be useful. A scan too poor for OCR is therefore not a scan that is safe to ignore.
"Many sensitive documents carry a distinct visual identity," said Dr Madhu Shashanka, co-chief technology officer and co-founder of Concentric AI.
"Not taking advantage of these visual cues leaves valuable signals untapped and limits the capability to discover and protect sensitive data," Shashanka said. He described the approach as going beyond discovery of textual data, to find sensitive information regardless of format.
Beyond standard identity documents
The feature is not limited to passports and licences. Documents unique to a single organisation often carry equally characteristic visual signatures.
Concentric AI says the capability lets its partners build dedicated models for customers. Those models find sensitive data by visual consistency within a specific environment, rather than by generic text patterns. Standard internal forms, claim documents and structured templates all fall into that category.
The capability sits within the vendor's unified AI and data security governance platform. That platform runs on patented language models in what the company calls its Semantic Intelligence Engine.
The platform is positioned to discover, govern and protect business-critical and compliance-sensitive data under a single policy framework. It covers data at rest, data in motion and data used by AI systems.
Concentric AI positions the wider platform around safe AI adoption. The stated aim is to govern data as AI systems begin to consume it, under the same policy framework used for data at rest.