Video and vision
How AI Detects Customer Behaviour in Stores, and What It Cannot Tell You
AI detects customer behaviour in stores by turning video into positions over time: a person appeared here, moved there, stopped for forty seconds, was joined by another, was approached by staff, left. Behaviour, in this sense, is movement and proximity and nothing more. Systems that claim to read interest, mood or intent from a face are still reading movement, and calling it something grander than it is.
What behaviour means to a camera
A detector finds a person in a frame. A tracker links that person across frames into a track: a start, a path, a set of stops, an end. Every behaviour signal is a rule applied to tracks. Browsing is a track that slows inside a zone. Waiting is a track that stands still near the counter. Group shopping is two tracks that arrive together and stay close. Attended is a staff track within reach of a shopper track for a few seconds. Abandoned is a shopper track that reaches the exit without ever being attended.
That is the whole vocabulary. It is useful because it is honest: it describes what bodies did, on a timeline, in a zone, without asking who they were.
The signals and how they are inferred
| Behaviour signal | Inferred from | Reliability on store cameras | What it cannot tell you |
|---|---|---|---|
| Passing versus browsing | Speed and time inside a zone | Good in open zones | Whether the browser wanted the product |
| Interest in a display | Dwell above a threshold, repeated returns | Fair; crowding shortens dwell | Interest in what, or why |
| Group shopping | Tracks arriving and staying together | Fair; pairs split | Who is deciding in the group |
| Waiting for help | Standing still, looking around, no staff nearby | Fair | What they wanted to ask |
| Attended | Staff track within reach for a few seconds | Good with staff exclusion | Whether the advisor said anything useful |
| Abandoned visit | Exit with no staff contact | Good with staff exclusion | Whether they would have bought |
| Picked up a product | Hand movement toward shelf | Poor on ordinary cameras | Almost everything |
| Frustration or delight | Claimed from facial expression | Not reliable, not run | Not a measurement |
What behaviour detection cannot tell you
- Intent. A long dwell by the premium range may be admiration, comparison, or waiting for a spouse in the next aisle.
- Reason. The camera sees the exit. It does not see the price quoted elsewhere, the delivery date that was too far, or the finance step nobody explained.
- Anything said. Cameras have no useful audio and the analytics does not listen.
- Whether the interaction was good. Staff proximity for three minutes could be a demo or an argument.
- Who the person was, and it should not try.
Why the face is the wrong place to look
Emotion recognition from faces at store distance, through a wide lens, in mixed light, is not a measurement. The published accuracy of such systems is poor even in laboratory conditions, and the categories themselves are contested. Age and gender estimation is more accurate and still unnecessary: knowing that a shopper was probably a man in his thirties does not change what the advisor should have said. Both process biometric data, which under India's DPDP Act needs a consent no store can honestly claim at the door. Behaviour signals from anonymous tracks give a store everything actionable with none of that exposure, and they can run on an edge box so no frame leaves the premises.
Behaviour plus conversation
A behaviour signal is a question. Eleven abandoned visits from the fridge zone on Saturday evening is a question about Saturday evening. The answer is in the consented conversations that were running on that floor in that hour: what was asked, what was skipped, what was offered. The camera describes the context of the interaction; the conversation is the interaction. Conversion has a cause only when both are read together, by hour and by zone, and the join needs no identity on either side.
Borentis captures consented in-store conversations on the advisor's phone today and scores them on the playbook. ShopperDNA, on the roadmap, would add behaviour signals from a store's existing cameras and join them to those conversations by time and zone.
Using behaviour signals in a store
- Report abandoned visits by hour and zone; staff the hours, not the store.
- Report time to first contact against the store's own standard, and coach the floor on the hours it slips.
- Read display dwell beside conversation topics: a display that draws dwell but no conversations is a merchandising win and a selling miss.
- Treat every signal as a count with a validation history, never as a fact about a person.
Frequently asked questions
Can AI tell if a customer is interested in a product?
It can tell that a shopper spent longer than usual near it and came back. Interest in the sense of wanting to buy is an inference, and a weak one.
Can AI read customer emotions in a store?
Not reliably, and not lawfully without biometric consent in India. The approach described here does not attempt it.
How does AI know who is staff and who is a customer?
By rules: a staff-only zone they pass through, a uniform colour, or a track that stays on the floor for hours. Plain-clothes promoters need their own rule.
Can behaviour detection tell me why customers walk out?
No. It can tell you how many walked out unattended and when. The why is in the conversation, which is a consented recording on the advisor's phone.
Related reading
- Can AI analyze CCTV footage?
- Measure customer movement using CCTV
- What is shopper intelligence?
- Walk-out recovery, the use case
Where Borentis applies this
- Walk-in Recovery: The customer who left is still yours.
- Objection Intelligence: The reason they did not buy, in their own words.
Borentis is the Agentic Operating System for Customer Interactions, built for Indian retail floors: consented one-tap capture on the advisor's phone, every conversation scored against your playbook with the evidence behind every number, leads created when a number is heard, and coaching from your own best conversations.