Computer Vision Use Cases in 2026: What Works, the Data You Need and How It Fails

What changed, and what did not

Two kinds of model now sit behind most computer vision products. Trained detectors and classifiers (YOLO-family detectors, segmentation models, classifiers fine-tuned on your images) are fast, cheap per image and very accurate on the narrow task they were trained for. Vision-language models such as Claude, Gemini and GPT answer open questions about a photo in plain language with no training at all, at a higher cost per image and with less predictable accuracy. The new option makes pilots far cheaper. What has not changed is that a model only works as well as the images it sees, that rare events are hard, and that someone has to decide what happens when the model is wrong.

Below are five areas we are asked about most. For each: what works today, the data you need and how it fails. Where we have first-hand work we say so; where we do not, we say that too.

Security and video review

This is the area where we have the most first-hand experience. For a US security services company whose officers patrol client sites wearing body cameras, we built an AI-assisted body-camera operations platform. Uploads land in S3, an event router classifies each one and fans it out to specialist jobs for video stitching, audio extraction, frame and sprite-map generation and transcription, each with idempotency keys and retry policies. Google Vertex AI then reviews every incident and returns a proposed severity, detections, scene conditions, key phrases and recommended actions.

Two design decisions carry over to any security product. First, the AI review is stored beside the guard's own record and never overwrites it; the model proposes, a person decides. Second, every review records the provider, model, prompt and temperature, so a result can be reproduced and explained when a client asks why an incident was rated the way it was. The value for the operator is speed: an hour of patrol footage becomes reviewable in seconds through stitched video, frames and sprite maps, and live dashboards show patrols in progress, GPS positions on a map and the latest frame from each camera.

• Works: summarising incidents from video and transcripts, finding the segment with a person or vehicle, reading signs and plates in good light, building shift and incident reports.
• Data: footage from your own cameras in your own conditions, plus past incident reports to compare against.
• Fails: low light, rain on the lens and body-worn camera shake; false alarms that train staff to ignore alerts. Measure the false positive rate per camera per night, not just detection accuracy.
• Watch: face recognition is regulated in many places (for example Illinois's Biometric Information Privacy Act, and biometric data used for identification is special category data under the EU GDPR). Leave it out unless a lawyer has signed off.

Retail

• Works: detecting empty shelf space and checking a shelf against its planogram from fixed or handheld cameras; recognising products from a photo for search or self-checkout support; counting people at entrances and queues without identifying them.
• Data: labelled images of every product in your range, from the angles and lighting your cameras see. Product recognition needs updating every time packaging changes, so plan a pipeline for new product images, not a one-off dataset.
• Fails: shoppers blocking the view, look-alike products (different sizes of the same brand), seasonal packaging, reflections from fridge doors. Loss-prevention uses produce false accusations if no person reviews them.
• Start with: a vision-language model checking a few hundred shelf photos against a written rule ("are there gaps on the top two shelves?") to prove the value before training a detector.

Manufacturing inspection

• Works: detecting known surface defects, missing components, wrong labels and misalignment on a line where the camera, lighting and part position are fixed. This is the classic computer vision success story because the conditions are controlled.
• Data: plenty of images of good parts, and as many defect examples as you can collect, labelled by your quality team. When real defects are rare, anomaly detection trained on good parts only can flag "different from normal" for a person to check.
• Fails: rare defect types the model has never seen, gradual lighting or lens changes, new product variants, and false rejects that stop the line. Agree the cost of a missed defect and of a false reject with operations before choosing a threshold.
• Watch: line speed sets the latency budget. If parts pass every 200 milliseconds, the model probably has to run at the line on edge hardware, not in the cloud.

Healthcare imaging

Medical imaging is the most regulated use case on this list. In the US, software intended for a medical purpose such as diagnosis or treatment can be a medical device in its own right; the FDA calls this Software as a Medical Device and expects the device pathway, clinical evaluation and a quality system. The FDA publishes a list of AI-enabled medical devices authorised for marketing, most of them in radiology, and some entries carry a predetermined change control plan that sets out how the model may be updated after clearance.

• Works: triage that puts likely urgent scans at the top of a worklist, measurement aids, and detection aids that a clinician confirms. Outside diagnosis, vision helps with administrative images such as reading insurance cards and referral forms.
• Data: large, diverse, expertly labelled datasets from several sites and scanner makers, under data use agreements and HIPAA safeguards.
• Fails: a model trained at one hospital performs worse at another with different scanners and patients. Plan validation on data from sites the model has never seen.
• Honesty: we have not built an FDA-cleared imaging device. Our healthcare work is HIPAA-aware software, such as the Zeuss telehealth platform; an imaging AI project needs a regulatory consultant and the client's clinical team involved from the start. See how to make an app HIPAA compliant.

Logistics

• Works: reading barcodes, labels and shipping documents from phone photos; checking proof-of-delivery photos (is there a parcel, is it at a door); damage capture at receiving; dock door and yard occupancy from fixed cameras.
• Data: photos from the real handheld devices and sites, including the bad ones: glare on shrink-wrap, crumpled labels, night deliveries.
• Fails: blur from rushed photos, glare, labels partly covered by tape. A good capture screen that rejects blurry photos on the device often adds more than a better model.
• Related: reading shipping documents is document processing, not general vision; see our intelligent document processing guide and Edge OCR. For fleet and delivery software, see delivery management software.

How to scope a computer vision pilot

1. Write down the decision the model will support and what a wrong answer costs, in both directions.
2. Collect 200 to 500 images or clips from the real cameras and conditions, including the difficult ones, and label them carefully as a test set.
3. Try the cheapest approach first: a managed API or a vision-language model with a written rule. Measure it on the test set.
4. Only if that falls short, label more data and train a custom model; measure it on the same test set.
5. Release with a person reviewing every result, track the error rate in production for a few weeks, then decide what to automate.

For budgets, read computer vision development cost. For the service, see computer vision development; for prediction rather than images, see machine learning development. Our wider AI and data services page lists what else we build, and a computer vision MVP sits in our $25k to $60k band.