Comparing French video AI providers boils down to answering a simple question: Which ones can scale from a pilot site to a network of dozens or hundreds of sites without operating costs, staff workload, and legal risk increasing proportionally? Three categories of criteria help determine the winner: compatibility with existing cameras, the ability to monitor an entire network from a single location, and compliance with the GDPR without the use of biometrics.

The rest comes down to the evaluation method. A useful comparison does not simply list logos; it compares real-world scenarios: the number of video streams processed per server, the time required to bring a new site online, disaster recovery procedures, the format of the data produced, and the terms for reversing the system at the end of the contract. XXII, a French computer vision software provider, designs its CORE platform according to this approach, relying on existing cameras rather than dedicated hardware.

What does the term “French video AI provider” mean?

The market brings together a wide variety of players under a single label. It includes camera manufacturers that embed algorithms in their products, video management software vendors that add an analytics layer, sensor-based counting specialists who do not process images, and analytics platform providers that integrate with a heterogeneous infrastructure.

This is the first distinction to make, as it determines what you’re actually buying. On-camera analysis requires replacing the entire infrastructure to upgrade processing capabilities. In contrast, a hardware-independent software platform allows models to be updated without altering the physical infrastructure. For a multi-site infrastructure that has been in place for several years, the cost difference between the two approaches quickly becomes a major factor.

The “French” criterion also warrants clarification. It may refer to the corporate headquarters, the location where the models are developed, the location of data hosting, or all three. Ask for the answers to these three questions separately: they do not always coincide.

What criteria distinguish a successful pilot from a large-scale deployment?

Compatibility with the existing infrastructure

A large-scale deployment rarely fails due to the quality of the model. It fails due to integration issues: cameras from different generations, unsuitable viewing angles, limited network bandwidth, and sites without a stable connection. Ask each vendor what their minimum requirements are for resolution, positioning, and bandwidth, then compare those requirements to a representative sample of your existing infrastructure—not just your best-equipped sites.

Operating Cost Per Site

The license price says nothing about the actual cost. The items to factor in are the computing hardware required on-site or at a central location, bandwidth consumption, installation and calibration time, annual maintenance time, and the operator’s workload in handling alerts. A solution that generates a lot of false positives is costly in terms of man-hours, even with a low-cost license.

Centralized monitoring

Once you have a few dozen sites, the question is no longer whether it works at a single site, but how to verify that it works everywhere. Look for a monitoring solution that reports the status of each stream, flags cameras that have been moved or obstructed, and allows you to deploy an update across the entire fleet without local intervention. This is what distinguishes a standardized product from a project rolled out site by site.

How does compliance affect the comparison?

The GDPR applies whenever processing involves personal data, and an image of an identifiable person falls under this category. The question, therefore, is not whether the regulation applies, but how the solution minimizes its scope.

There are two approaches. The first involves processing the image and then storing it. The second involves extracting non-personally identifiable data from the data stream—such as a count, a location within a zone, or the duration of presence—without storing the image or creating a template that could be used to recognize an individual. Facial recognition falls under biometric data, which is governed by Article 9 of the GDPR as a special category of data, subject to a general prohibition with limited exceptions. A solution that does not rely on facial recognition significantly simplifies the data protection impact assessment.

The CNIL has published a position paper on so-called “augmented” cameras, which distinguishes between uses based on their purpose and their intrusiveness. Ask each vendor for the documentation they provide to support your data protection impact assessment, as well as a detailed list of the data generated and their retention periods. A vendor that cannot summarize this list on a single page is a vendor that is shifting the compliance burden onto you.

What questions should you ask during the consultation?

Just a few questions are enough to weed out weak proposals. How many video streams does a single device process in your typical configuration, and with what latency? How long does it take from granting camera access to obtaining the first usable data at a new site? What is the expected behavior if a site’s network connection goes down for several hours? In what format can the data be exported to an existing management tool? What is the data deletion procedure, and what happens to the historical data?

Also ask for a reference site comparable to yours in size and constraints—not a demonstration under ideal conditions. A vendor’s ability to arrange this exchange speaks volumes about the maturity of its deployments.

How does XXII fit into this landscape?

XXII is a French computer vision provider. Its CORE platform analyzes footage from cameras already in place and generates operational metrics—such as counts, occupancy measurements, dwell times, and event-based alerts—without facial recognition. The approach targets multi-site installations in retail, logistics, manufacturing, transportation, and public-access facilities, where the camera infrastructure is already in place and replacing it makes no economic sense.

This positioning does not eliminate the need for comparison; rather, it structures it. If your need involves a single camera and a fixed use case, on-device analysis may suffice. If your needs involve an extensive, heterogeneous, and evolving camera infrastructure, the integration, monitoring, and compliance criteria described above should be the deciding factors when selecting a solution.

Where should you start a comparison?

Create an evaluation grid before meeting with vendors, not after. List your sites by type, identify priority use cases, set the target number of sites for twelve months and thirty-six months, and then evaluate each candidate based on these criteria. A comparison conducted in this manner yields a decision that can be justified to both the finance department and the data protection officer.