AI Video Analysis: What It Can Watch For Now

Google reported cutting video analysis costs by around 66 percent with its latest model. Video has been technically possible and economically silly for small businesses for years, because analysing hours of footage cost more than the insight was worth. Drop the unit cost by two thirds and a lot of projects move from obviously not worth it to worth a two-week trial. The harder question arrives immediately behind that one.

What it does well, and what it does not

Give it footage and it will summarise the sequence, locate a specific moment, flag frames that differ from a reference, and count occurrences over time. That is genuinely useful and it is all description.

What it does badly is intent. It can tell you a person stood at a shelf for four minutes and left without buying anything. It cannot tell you whether they were confused, uninterested, or waiting for someone. The description is reliable, the inference is yours, and confusing the two is the most common way this goes wrong.

Four applications that hold up

Three of those four are about watching less video, not more. That is the framing worth keeping, because it points at the applications that pay off and away from the ones that quietly become surveillance.

Privacy arrives faster than you expect

Footage of identifiable people is personal information under Canadian privacy law, and analysing it is a use that needs a purpose you can state out loud. If staff appear in it, employment and provincial monitoring rules apply as well.

Get advice for your province before pointing analysis at anything containing staff or customers. Treat signage, a written purpose, and a retention limit as the starting point rather than the finish line. The general approach in keeping AI use PIPEDA-compliant applies, and the employment overlay is the same one we set out in time tracking versus surveillance.

The line between analysis and surveillance

It is the same distinction as with time tracking, and it sits in the same place: the unit of analysis.

Queue length peaks at eleven on Saturdays is operational information that changes a staffing decision. This employee was away from their station for six minutes is monitoring, and it changes the relationship whether or not you ever act on it. The technical capability is identical in both cases, which is why the boundary has to be a decision rather than a default.

Cheap removes the discipline

Here is the risk created by the price cut specifically. Cost used to force restraint. Nobody analysed footage speculatively because it was expensive, so the only projects that happened were ones somebody had thought about.

With that constraint gone it becomes easy to build continuous monitoring nobody reads and to accumulate analysis nobody asked for. Start from a decision you make repeatedly and badly, work back to what would inform it, and analyse only that. It is the same argument as building a dashboard from decisions rather than from available data, which we made in building a dashboard you will actually look at.

How to trial it

Pick one question you currently answer by guessing. How long does the third step of the install actually take. How many people walk past the display without stopping. Whether the finish on Tuesday's batch matched Monday's.

Analyse a sample of existing footage to answer that one question, and compare the answer against what you assumed. Two weeks, one question. If the answer changes a decision, you have found something worth building. If it merely confirms what you thought, you have saved yourself a project, which counts as a good outcome.

Frequently Asked Questions