Computer Vision ROI in Retail: A Model That Holds Up

What actually drives computer vision ROI in retail?

Computer vision ROI in retail comes from four line items, not from the model: on-shelf availability, labour scheduling, shrink, and safety incidents. Everything else a vendor demo shows you is a dashboard. If a proposed system does not move one of those four, it will not return its running cost.

Those four share the property that makes them financeable - each already has a number attached to it in your existing reporting:

• On-shelf availability - your out-of-stock rate, visible in POS gaps.
• Labour scheduling - your labour spend per store hour.
• Shrink - your shrink percentage.
• Safety incidents - your recorded incident count.

That matters because a vision system cannot create a new metric and then claim credit against it. It has to move a line that finance already tracks.

When a pilot reports "dwell time improved," nobody can price that. When it reports "phantom out-of-stocks on aisle 7 fell from 11 percent to 4 percent over six weeks," that converts directly into recovered sales.

Start the business case from the existing line item. Work backwards to the model.

How do you model the cost side before you build?

Cost per store per year, not total project cost. A 200-store chain is 200 copies of the same running cost, and the per-store line is the one that decides whether the rollout continues past the pilot.

The figures below are our own planning heuristics rather than published benchmarks - the ranges we use when scoping vision work, and they assume one use case, not three.

• Cameras: $0 to $4,000 per store per year, amortised. Zero where existing CCTV already sees the shelf face, the top of the range where every aisle needs a new angle.
• Edge compute: $600 to $1,800 per store per year, amortised across one or two boxes. Concurrent stream count and frame rate move it.
• Connectivity and cloud: $180 to $900 per store per year for metadata only. Uploading frames instead of detections moves it by an order of magnitude.
• Model maintenance and re-labelling: $1,200 to $3,600 per store per year, set by planogram churn and seasonal SKU turnover.
• Store staff time handling exceptions: usually the largest of the five, and governed entirely by alert precision.

That last line is the one that sinks projects. It is rarely in the vendor proposal, because it is your cost and not theirs.

Most retail computer vision deployments fail on camera placement and alert routing, not on model accuracy.

Which retail computer vision use cases pay back fastest?

On-shelf availability, by a wide margin. It has the shortest path from detection to a specific action a specific person can take in under five minutes.

Shelf monitoring works because the intervention is unambiguous. The system says aisle 7, bay 4, facing empty. An associate walks there with a cart. The loop closes the same shift.

Queue and labour scheduling is second. Counting people in a queue zone is a solved detection problem, and the output feeds a rota decision that a store manager already makes weekly. The value is real but smaller, because you are optimising an existing schedule rather than recovering lost sales.

Loss prevention is the one most often sold and least often measured honestly. Detecting a suspicious motion pattern is not the same as preventing a loss.

Attributing a shrink reduction to the camera rather than to the staffing change that accompanied it is genuinely hard. Treat any shrink claim from a pilot as unproven unless you ran a control group of stores.

Which metrics should you track?

Three model metrics and three business metrics. Track them separately, because a system can have excellent model metrics and zero business effect.

On the model side:

• Recall on the specific condition you care about.
• Precision at the alert threshold you actually ship.
• Latency from frame to alert.

Precision matters more than accuracy here. A model at 94 percent accuracy that fires 40 alerts a day with 60 percent precision is worse than useless, because the store learns to ignore it.

On the business side:

• Time from alert to resolution.
• Percentage of alerts acted on.
• The underlying line item - out-of-stock rate, labour hours, incident count.

The second one is your early warning. In the deployments we have scoped, once the acted-on rate falls below about half, the system is dying regardless of what the model scores. That threshold is our field heuristic, not a published benchmark.

Getting these numbers requires an event store, not screenshots. That is a data engineering job, and it should exist before the first camera goes up.

Where does computer vision in retail break?

Two places, consistently: the camera cannot see what you need, and the alert reaches nobody who can act.

Camera geometry is the first killer. Loss-prevention cameras are mounted high and wide to cover floor area, while shelf-level product recognition needs an oblique or near-frontal view from roughly three metres - the working distance we plan around, not a published spec.

Small objects in dense groups are a known weakness of single-stage detectors, documented since the original YOLO paper, and a packed shelf shot from eight metres away is exactly that case.

Infrared night mode is the second geometry problem. It strips colour, and a large share of SKU discrimination on a packed shelf is colour. Systems validated in daytime lighting quietly degrade after close.

The routing failure is more mundane. Alerts land in a web dashboard that a store manager checks twice a week. By then the shelf has been refilled or it has not, and either way the data is now noise. Route to the handheld device the associate already carries, or do not bother.

Does safety detection change the business case?

Yes, and it is the most underrated pool. Safety cases in retail computer vision - spill detection, blocked fire exits, unsafe ladder use, loading-bay proximity - carry both a direct cost and a liability tail.

The direct cost is straightforward: incident rate multiplied by average claim and lost-time cost, both numbers your operations team already holds. The liability side is harder to price but tends to be larger, particularly in jurisdictions with active occupational health enforcement.

Safety detection is also technically easier than shelf work. A spill is a large, high-contrast region on a uniform floor. A blocked exit is a static object in a defined polygon. These are reliable at camera distances where SKU recognition is not, which means your existing CCTV may already be adequate.

The trade-off: safety alerts are low-frequency, so a store may see two per month. That is good operationally and bad for building staff habit. Pair safety with a higher-frequency use case so the alert channel stays alive.

Edge or cloud: what does the architecture actually cost?

Edge inference for anything above four cameras per store. That threshold is our own scoping rule rather than a vendor benchmark: above it, streaming raw video to the cloud has consistently cost us more in bandwidth than the GPU we avoided buying, and it turns every store outage into a blind spot.

The workable pattern is inference at the edge, metadata to the cloud. A box in the back office runs detection on local streams and ships only structured events: timestamp, camera, zone, class, confidence. Bandwidth drops from megabits per second to kilobits per hour.

Runtimes worth evaluating include NVIDIA DeepStream for multi-stream pipelines, AWS IoT Greengrass for fleet deployment and over-the-air model updates, and Azure AI Vision if your estate is already on Microsoft.

The operational reality is that you are now running a fleet of edge devices with model versions that drift apart. That is an MLOps problem, and it is the part most retailers are not staffed for. Budget for it explicitly.

How should a Canadian retailer handle privacy here?

Avoid biometric identification unless there is a legal reason you cannot. Almost all of the value in retail computer vision comes from anonymous detection that never identifies a person.

Commercial collection of personal information in Canada falls under PIPEDA, and faceprints are sensitive personal information requiring meaningful consent. A shopper walking past a camera has not given that. Several Canadian retail operators have learned this through regulatory investigation rather than legal review.

There is an accuracy argument too. NIST's Face Recognition Vendor Test programme has documented meaningful differences in false match rates across demographic groups depending on the algorithm. Deploying that against your customers creates exposure that no shrink number justifies.

Design the pipeline so identification is impossible rather than merely disabled. Detect people as anonymous bounding boxes, discard frames at the edge, persist only counts and zone events. That is a defensible architecture in a privacy review, and it costs nothing extra to build that way from the start.

How do you prove retail computer vision ROI in a quarter?

One use case, three to five stores, one control group, ninety days. Anything broader produces a result nobody can attribute.

Pick the single highest-value pool from the four drivers above. Instrument the baseline for two weeks before you switch anything on, because a baseline reconstructed afterwards will be argued with. Run matched control stores with no system at all.

Write the kill number down before the first frame: the target movement on the line item, the minimum acted-on rate, and the precision floor below which you pull the alerts instead of retuning them. A pilot with no kill number becomes a permanent pilot.

This is the shape we have used across AI development work since 2007, now across 212 delivered projects for clients including Paytm, Snapdeal, IKEA, Nestlé, Shiprocket and Vikram Solar. The constraint is always the same: a pilot that cannot produce a defensible number is a demo.

Before you scope a pilot at all, find out which store workflows the cameras would actually improve. Our broken workflow assessment maps where the time goes today, so you are not automating a process that should be redesigned instead.

If ninety days is too slow for your board, a Moonship MVP puts a working single-aisle pipeline in front of you in 30 days from $2,999 - enough to test your camera angles before committing to hardware.

If you already know the use case, book a 30-minute call and we will size the pilot with you.