Computer vision for business: practical uses and how to start

What computer vision can do for businesses today, from document reading to quality control, plus accuracy, privacy and cost considerations.

By the UK Technology Hub teamPublished 4 minute read

The short answer

Computer vision is AI that interprets images and video. Businesses use it to read documents, check product quality, count stock, verify deliveries, analyse damage photos and power camera features in apps. Start with a clear, measurable task, test general multimodal models first, move to custom models when volume or accuracy demands it, and handle images of people with strong privacy safeguards.

Computer vision used to require specialist teams and large budgets. Today, general AI models can describe images and read documents out of the box, and custom models can be trained for specific tasks with far less effort than before. That brings practical visual AI within reach of many businesses.

What computer vision can do

  • Read text in photos and scans (optical character recognition).
  • Classify images: is this product damaged or intact?
  • Detect objects: find and count items in a picture.
  • Segment images: outline specific regions, such as a defect.
  • Compare images: spot differences between before and after.
  • Track movement in video.
  • Recognise patterns such as barcodes and QR codes.

Practical uses by industry

Retail and e-commerce: check shelf stock, tag product photos, power visual search ("find similar items"), generate alt text.

Logistics: read labels, verify parcels, photograph proof of delivery, detect damaged packages.

Manufacturing: spot visible defects on production lines, check assembly steps, read gauges.

Property and insurance: analyse photos of damage to speed up initial assessment, with human review.

Field services: technicians photograph equipment; AI reads serial numbers and model details.

Hospitality and facilities: monitor occupancy levels using anonymised counts.

Finance and admin: extract data from receipts, invoices and forms.

General models versus custom models

General multimodal models can describe and reason about images with no training. They're ideal for prototypes and varied tasks. Our guide to multimodal AI explains what they can do.

Custom models are trained on your own labelled images for one specific task. They're usually more accurate, faster and cheaper per image at high volume, and can run on edge devices like cameras.

A common path: prove value with a general model, then train a custom model once the task and volume are clear.

Data and labelling

Custom models need examples: images labelled with the correct answer, such as "defect" or "no defect," or boxes around objects. Quality matters more than quantity. Include varied lighting, angles, backgrounds and edge cases. Our guide to preparing data for AI covers how to collect and label data.

Accuracy and testing

  • Define what "good enough" means, such as catching 98% of defects with few false alarms.
  • Test on images from real conditions, not just ideal samples.
  • Understand the cost of errors: a missed defect may matter more than a false alarm.
  • Keep humans reviewing uncertain cases.
  • Monitor accuracy over time, as lighting, products and cameras change.

Privacy and ethics

Images often show people. Be careful:

  • Avoid identifying individuals unless it's essential and lawful.
  • Facial recognition is heavily regulated or restricted in many jurisdictions; take legal advice before considering it.
  • Blur or crop faces and licence plates where they're not needed.
  • Tell people when cameras and AI are in use.
  • Limit retention and access.
  • Assess risks formally for higher-risk uses.

Costs

Costs include image capture (cameras or app features), model usage or training, integration with your systems and ongoing monitoring. General model API costs scale with image volume and resolution; custom models cost more to build but less per image at scale.

Edge devices versus the cloud

Some vision tasks run in the cloud, where powerful models analyse uploaded images. Others run on edge devices, such as smart cameras or phones, close to where images are captured. Edge processing reduces bandwidth, speeds up results and keeps images local, which can help privacy. Cloud processing is easier to update and can use larger models. The right choice depends on volume, speed and privacy needs.

Running a pilot

  1. Choose one task with a clear business value.
  2. Collect a few hundred real images and label a test set.
  3. Test a general model and measure accuracy.
  4. Design the workflow: what happens to results, and who reviews uncertain ones?
  5. Run in parallel with the current process for a few weeks.
  6. Decide: scale up, train a custom model or stop.

How we help

Our AI solutions team designs computer vision pilots and production systems. Our app development team builds camera features into mobile apps, and our web app developers integrate vision results into dashboards and business systems.

Frequently asked questions

Do I need my own cameras for computer vision?

Not always. Many uses work with photos taken on phones or images already in your systems.

How many images do I need to train a custom model?

It depends on the task. Simple classification may need a few hundred good examples; complex detection can need thousands.

Is facial recognition legal for businesses?

Rules vary widely and are often strict. Take legal advice before using any form of facial recognition.

This guide is general information, not legal, tax or financial advice, and details such as prices, features and rules can change. Product and company names mentioned are trademarks of their respective owners and are used only to identify those products; UK Technology Hub is not affiliated with or endorsed by them.

Related guides

All AI & automation guides