The Race for Context: How to Measure What Happens Across Hundreds of Thousands of m²

E-commerce has twenty years of context: every click, every search and every abandoned cart is recorded. That is why an AI assistant can explain a drop in online sales. In physical spaces, the data usually stops at the edges: counts at the doors and sales at the registers. The same happens in a shopping center, a retail chain or an airport.

Capturing context at scale means measuring what happens inside, continuously and anonymously: flows between floors and zones in a 100,000 m² shopping center, visits and queues in hundreds of stores at once, journeys and waiting times in an airport terminal. The question is not whether that is useful. The question is which architecture can do it at a reasonable cost, and keep doing it as AI evolves.

AI already knows how much. It does not know why. Context is what turns an indicator into an explanation.

A real example from a shopping center in South America. The total traffic of the center held steady, but one floor started receiving fewer visits. The question to an AI assistant was the same in both cases: why does Level 2 receive fewer visits if total traffic held steady?

With entrance counts only, the assistant can only confirm that traffic held steady. With the camera data connected via MCP (the open standard that connects AI assistants to data sources), it sees the flow between floors week by week: after two escalators were reversed, 8.2% fewer people went up to Level 2. The answer arrives with the cause and the moment it started.

For years, analytics in physical spaces was bought as a closed package. The same vendor manufactured the equipment, installed it, built the network, developed the software and integrated it with the client. Changing one piece meant changing the vendor, and what could be measured depended on the device purchased.

Then the chain split into layers: camera manufacturers, installers, network and IT providers, AI software and integrators. The operator could choose the best option in each layer. KSI was born in that shift: AI software on cameras of any brand, without manufacturing or installing hardware.

Today AI agents are disrupting several of those layers at once. Hardware becomes a commodity: a camera only needs to record well. Dashboards and reports are absorbed by agents. Integration is standardized through MCP. And a new layer appears: analysis and action, where agents explain, recommend and alert.

In this transition, large operators are splitting into three groups. The difference is not budget or intent. It is where each one decided the AI should run: in a dedicated sensor, inside each camera, or as software on the cameras that already exist.

Those who stay with the old model

Dedicated counting sensors at the entrances, with or without AI. They solved the question of the last decade: how many people came in. Their limit is structural: high cost per analytics point, a second fleet of devices to maintain next to the CCTV, vendor lock-in and innovation that arrives only when the manufacturer renews its product line.

Those who fall into the embedded-AI trap

Replacing sensors with smart cameras looks like modernization. In practice it repeats the same model with a more expensive device: a camera with embedded AI can cost up to USD 1,000, becomes obsolete in about 5 years and ties the analysis to the hardware vendor again. At that price per point, covering a 50,000 m² shopping center, an airport terminal or a network of hundreds of stores becomes unviable, and the AI stays frozen at the version that came with the device.

What the leading operators are doing

The most advanced operators separate the infrastructure from the intelligence. Cameras of any brand, owned by the operator, that only record. A perception layer, delivered as software, that turns video into anonymous and auditable data. And agents connected through an open standard that query that data in natural language. Each layer can be replaced without rebuilding the rest.

Seen from first principles, measuring what happens inside an entire space, or an entire network of spaces, is a decision tree. Each wrong decision makes full coverage economically unviable.

The only branch that reaches the goal is AI as software on the existing cameras. Where coverage must be extended, cameras are available from USD 50. In real projects, KSI has validated a difference of 10x or more compared with dedicated hardware: lower infrastructure cost and more analytics coverage for the same budget.

The principle is the same in every vertical, but the unanswered question changes, and so does what each path looks like. Each vertical has its own article with published cases.

In the model that is consolidating, each layer will have a clear owner and the operator will be able to change any of them without rebuilding the others.

• Vendor competition: perception providers will compete on intelligence over the operator's infrastructure, not on hardware.
• Several AIs in parallel: different models and vendors can run on the same cameras at the same time.
• Evolving AI: improvements will arrive as software updates, without replacing devices.
• Agentic connection: Claude, ChatGPT, Gemini or KSI's own agents will query the data through MCP.
• Cost per analytics point, not per device, and coverage of the whole space.

The operator that owns its infrastructure and chooses its data layer will be able to adopt each new generation of AI on the day it appears.