A video AI platform’s ability to be deployed across several hundred sites is not judged by the performance of its models, but by four less obvious factors: the processing architecture and its marginal cost per site, the ability to support a heterogeneous camera fleet, the tools for remote operation and updates, and the integration model with existing systems.
A successful pilot at three sites does not necessarily indicate success in any of these four areas. Most projects that stall do not fail because of detection; they fail because of scaling. This article details what changes as the camera fleet grows, and the questions to ask before committing.
A pilot takes place under favorable conditions: carefully selected sites, properly positioned cameras, committed teams, available vendor support, and manually adjusted settings. Each of these elements disappears when the system is rolled out on a larger scale.
In a large-scale deployment, cameras come from multiple generations and multiple manufacturers. The field of view angles were not designed for measurement. Local networks differ from one site to another. On-site personnel change. The time available to configure each camera is reduced to just a few minutes, compared to several hours during the pilot phase.
The question to ask a vendor, therefore, is not about the performance of its model, but about the time and expertise required to bring the 100th site online.
Several effects emerge in a nonlinear manner.
Three main options coexist, each with a different cost profile.
Analysis is performed on equipment installed on-site. The required bandwidth remains low, since only aggregated data is sent back. On the other hand, each site requires hardware to be purchased, installed, and maintained, which creates a fixed cost per site and involves replacement logistics.
Data flows are sent to a shared infrastructure. Connecting a new site is quick, and updates to models immediately benefit the entire network. Constraints relate to the network and the volume of data transmitted, as well as hosting location requirements.
Part of the processing is done on-site, while monitoring and aggregation are centralized. This is the most common compromise for large fleets, but it requires strict version control: heterogeneous local hardware that is poorly maintained quickly becomes an obstacle again.
None of these options is superior in an absolute sense. The choice depends on the quality of network links, data location constraints, and tolerance for on-site intervention.
A model trained on a set of scenarios degrades when conditions change: a new layout, new work attire, the season, construction work, or the replacement of a camera. This drift is normal. What sets a vendor apart is its ability to detect and correct this drift without requiring human intervention at every site.
The factors to consider are concrete: the existence of health indicators by data stream, automatic detection of outliers, the ability to remotely readjust or retrain the model, traceability of deployed model versions, and a rollback procedure in case of regression.
Without these elements, the system silently deteriorates: dashboards continue to display numbers, but those numbers gradually cease to reflect reality.
A metric that doesn’t reach the tool where decisions are made is useless. In a large-scale deployment, integration often accounts for a significant portion of the project workload: connecting to the existing video surveillance system, feeding data into a monitoring tool or hypervisor, exporting to a data warehouse, and supplying data to a staffing planning tool.
Useful questions focus on the existence of documented interfaces, the stability of their format over time, and the ability to retrieve measurement history in a usable format without the software vendor. This last point determines the project’s reversibility.
A few simple questions can effectively distinguish between candidates:
Precise answers to these five questions speak louder than a demonstration.
For a fleet located in Europe, the hosting location and the law applicable to the contract are not mere administrative details. They determine the ability to document data processing, respond to a request from a supervisory authority, and fulfill commitments made to employee representatives. The European Artificial Intelligence Regulation also imposes additional documentation requirements that vary depending on the system’s intended use and risk level.
XXII is a French computer vision software company whose CORE platform analyzes video feeds from existing cameras without using facial recognition. For management planning a rollout across a large number of sites, the value of working with a partner based in Europe lies primarily in contractual clarity and operational proximity during the commissioning phases—which remain the critical moments that determine the project’s success.