How much does computer vision development cost in Australia? (2026)

How much does a computer vision project cost?

Most Australian computer vision projects fall into three bands: a proof of concept on cloud APIs, a custom model at one site, and a multi site platform. We found no published Australian price guide for vision work, so these are worked estimates: typical effort multiplied by an assumed agency day rate of $1,100 to $1,500 ex GST. Re:Sourced’s 2026 guide puts senior Sydney contractors at $800 to $1,150 a day depending on discipline, and says specialist AI work currently prices above the software band; an agency day carries management and testing on top.

Project	Typical use	Typical effort	Worked estimate (AUD, ex GST)
Proof of concept on cloud vision APIs	PPE detection, people counting or text reading on existing camera feeds	12 to 30 days	$13,200 to $45,000
Custom trained model, one site, edge deployment	Defect inspection on a line, stock counting, a specific part or vehicle type	50 to 100 days	$55,000 to $150,000
Multi site platform	Several sites with alerts, dashboards, model monitoring and retraining	180 to 350+ days	$198,000 to $525,000+

These are not AWLabs prices. We quote a fixed price after discovery and don’t publish day rates.

Vision API or custom model?

Use an off the shelf API when the thing you need to see is common; train a custom model when it’s specific to your business. APIs from AWS, Azure and Google detect people, vehicles, text, faces and safety equipment out of the box. Amazon Rekognition, for example, includes a protective equipment detection API. A defect on your product, a particular pallet type or a weed in your crop needs your own labelled images.

	Vision API	Custom trained model
Upfront cost	Low: integration and testing only	Higher: data collection, labelling, training, evaluation
Time to first result	Days	Weeks to months
Recognises	General objects and scenes	What you trained it on
Running cost	Per image or per minute of video	Per hour of hosted inference, or edge hardware
Best for	Common objects, early proof of concept	Specific defects, parts or conditions

A sensible path is to try an API first. If it gets close enough, you have saved the cost of a custom model. If it doesn’t, the proof of concept has given you the camera positions, sample images and accuracy target the custom build needs.

Data collection and labelling: the hidden half of the budget

A custom model is only as good as the images it learns from, and gathering images that represent real conditions takes longer than training. Night shifts, rain, glare, dust, unusual angles and rare defects all need to be in the training set, or the model will fail on the day they happen.

A worked labelling estimate, with stated assumptions: 10,000 images, drawing boxes around objects at 30 to 90 seconds an image, at a fully loaded cost we assume at $45 an hour, is 83 to 250 hours, or about $3,750 to $11,250 for one pass. Add a second person checking a sample, and a further round when the first model shows what it gets wrong. If defects are rare, you may need to wait weeks to collect enough examples, which is a timeline cost as much as a dollar one.

Edge hardware or cloud inference?

For occasional images the cloud is cheapest; for always on camera feeds, processing on site usually wins. The arithmetic shows why. These figures use AWS’s Sydney price list at the time of writing and the RBA’s 6 November 2026 rate of US$0.69 per A$1.

Cloud, per image: 8 cameras, one frame every 5 seconds, 10 hours a day, 30 days is about 1.73 million images a month. At US$0.0012 per image for the first million and US$0.00096 above that, the bill is about US$1,900, or roughly A$2,750 a month.

Cloud, hosted custom model: Rekognition Custom Labels charges US$4 per inference hour in Sydney. Left running all month (730 hours) that is US$2,920, about A$4,230. Running it only for 10 hours on 22 business days is 220 hours, about US$880 or A$1,275.

Edge: NVIDIA prices its Jetson Orin Nano Super Developer Kit at US$249 at the time of writing. A developer kit isn’t a production unit; enclosures, power, mounting, networking and an industrial module add to that, so budget per site from your own installer’s quote. Once installed, there’s no per image fee, and video doesn’t need to leave the site.

Edge brings its own costs: remote updates, device monitoring, replacing failed units and someone who can attend site. For two or three cameras with intermittent use, the cloud is often simpler. For a line or yard watched all day, the figures above suggest edge hardware costs less than per image cloud fees well within the first year.

Privacy and surveillance law for cameras in Australia

Once a camera system captures people, privacy law applies, and once it identifies them by face, it handles sensitive information. In November 2024 the Privacy Commissioner found Bunnings breached Australians’ privacy with a facial recognition system that captured the faces of everyone entering 63 stores in Victoria and New South Wales. The OAIC’s announcement says the Privacy Act classes facial images and other biometric information as sensitive information, for which consent is generally required, and the determination also found failures in notice and in the privacy policy. Bunnings had the right to seek review.

Workplace cameras have their own rules. In NSW, the Workplace Surveillance Act 2005 requires written notice to employees at least 14 days before surveillance starts (section 10), unless an employee agrees to less. For camera surveillance, section 11 requires the cameras, or casings that indicate them, to be clearly visible and signs to be clearly visible at each entrance. Other states have their own surveillance laws.

Design choices that reduce both risk and cost:

• Detect objects, not identities, wherever the use case allows it.
• Process on the edge and send counts or alerts rather than video.
• Blur faces and number plates before storing frames for retraining.
• Keep retention short and write it down.
• Do a privacy impact assessment before installation, not after.

Our guide to the Privacy Act and AI covers the wider obligations. This is general information; get legal advice for your sites.

Accuracy, testing and false alarms

A vision system that cries wolf gets switched off, so agree how you’ll measure false alarms before the build. Set aside a test set of images the model never trains on, captured in the conditions that matter, and report two numbers: how many real events it missed, and how many alerts were wrong. Then decide which error is more expensive. Missing a forklift near a pedestrian is worse than an extra alert; flagging a good product as defective wastes stock.

Plan a supervised trial in production before alerts go live, with a person reviewing every alert for a few weeks. Our guide to moving an AI pilot into production covers what that phase needs.

Ongoing costs and a three year total

Vision systems drift as lighting, products and camera positions change, so budget for retraining from the start. A worked three year total for a custom model at one site, with stated assumptions:

Cost line	Assumption	Three year total (AUD, ex GST)
Build	50 to 100 days at $1,100 to $1,500	$55,000 to $150,000
Edge hardware	Two production devices with enclosures and installation, assumed $3,000 to $8,000	$3,000 to $8,000
Cloud for alerts, dashboard and storage	Assumed $100 to $400 a month	$3,600 to $14,400
Relabelling and retraining, years 2 and 3	One round a year: labelling as above plus 5 to 10 engineering days	$18,500 to $52,500
Total		about $80,000 to $225,000

The biggest swing factors are the number of conditions the model must handle, whether the site has reliable power and connectivity, and how many alert types feed other systems. For construction sites, our construction software and AI page covers the site specific issues.

How AWLabs approaches computer vision

We start with your cameras and your images, not a model. A short proof of concept on existing footage tests whether an API is good enough, and gives us the accuracy target and privacy design for anything custom. We prefer detecting objects over identifying people, and processing on site where feeds run all day. Senior engineers lead every build.

The first step is a Development Consult, a 15 minute call. A paid Discovery Audit then tests feasibility on your own footage, and you get a fixed price Schedule of Work with three options: The Minimum, The Recommended and The Complete. Our AI integration service and AI development cost guide cover the wider picture.