Milestone's pitch for AI in XProtect starts with where the training video came from
The Copenhagen VMS maker says its summarization and search tools were tuned on a 300,000-item library of customer video, gathered under fair-exchange agreements and anonymized before use. Summaries ship now; search arrives at the end of 2026.
Milestone Systems announced a set of AI features for its XProtect video management software on March 19, 2026, and spent as much of the announcement on the training data as on the features. The press release, datelined Copenhagen and issued by the company, describes three products: AI Search, XProtect Video Summarization and Video Anonymization. Only one of them is available today.
What ships now
The one that ships is Video Summarization. An operator sends a clip to the tool with a custom prompt and gets back a structured written description of what happened in the scene. Milestone has wired it to XProtect's rules engine, so a summary can be produced automatically whenever an alert fires, on a schedule, or on demand. The stated purpose is standardized incident documentation, the paperwork that follows every event and that operators tend to write inconsistently at the end of a shift.
AI Search, the natural-language query tool for finding events across many cameras, is promised for the end of 2026, with an on-premises version to follow through BriefCam. Video Anonymization, which redacts faces and license plates for GDPR purposes, is listed as coming soon in both XProtect and the Arcules cloud service.
The data argument
Milestone's differentiating claim is the library the models were fine-tuned on. The company says it holds more than 300,000 categorizations of real security video, collected from customers under what it calls fair-exchange agreements and anonymized with brighter AI's technology before any model sees it. The City of Genoa in Italy is named as a recent contributor. NVIDIA's Cosmos foundation models are the base, with Milestone's data layered on top.
The release offers one benchmark for that approach. A traffic vision-language model tuned on the library scored 19.4 percent better on flow and direction accuracy and 4.4 percent better on alert verification than the untuned base model. Those are the company's own figures, and the second number is modest.
Andrew Burnett, Milestone's chief technology officer, framed the whole program around a single word. "Trust that the data is compliant, trust that the AI is accurate, and trust that these tools will genuinely make operators' jobs easier," he said.
The framing is aimed at European buyers, for whom the provenance of training data has become a procurement question rather than a curiosity. It is also a quiet concession that a general-purpose model watching a car park needs to have seen a lot of car parks. Milestone avoids the word "agent" throughout the release. What it is shipping is the documentation half of what agents are supposed to do, with the search half held for later.
Read the original at Milestone Systems.