Saturday, fragrance bay. The pharmacist is counselling at the dispensary. A customer opens a tester, then a second bottle goes into a lined bag. Playback on Monday will show the event. By then the stock is gone and the only decision left is write-off versus insurance.
That is the gap AI theft detection is supposed to close: move from post-event evidence to in-aisle intervention. The market is crowded with claims. Australian pharmacies need a sharper filter - what the crime data actually shows, how the pipelines work, and which architecture choices create Privacy Act and latency risk.
> Quick answer: For Australian pharmacies, effective AI theft detection means real-time concealment analysis on existing CCTV, with alerts fast enough for floor staff to act, preferably with inference on-premises so customer video does not leave the building for routine processing.
At a glance
Skip global “retail theft is rising” headlines. Use regional numbers you can defend in a banner-group business case.
The 2022 Australia and New Zealand Retail Crime Study, produced by Griffith Criminology Institute researchers for the Profit Protection Future Forum (PPFF), surveyed companies operating more than 8,900 stores with ~AUD 136 billion turnover - roughly a third of the ANZ industry.
Key findings from that study:
Those figures are not pharmacy-only. They are the sector baseline. Pharmacies sit inside that distribution with a distinctive product and method profile (next section). For store-level cosmetics economics we use elsewhere, see the real cost of retail shrinkage in Australia.
Recruiting context matters too: 94% of respondents in the same study said hiring LP staff with relevant experience was somewhat or extremely difficult. Independents rarely have a dedicated LP analyst. Tools that create actionable alerts for floor staff matter more than tools that create more footage for an empty LP seat.
The PPFF study’s pharmacy hot product ranking for the period was:
| Pharmacy hot product (PPFF 2022) | Why it is targeted |
|---|---|
| Perfumes and fragrances | High resale value, small form factor, open-shelf display |
| Facial creams | Premium unit price, easy bag concealment |
| Make-up products | High volume SKUs, hard to spot missing units until stocktake |
That matches what pharmacy operators describe operationally: small, high-value, open-shelf SKUs with strong secondary markets. Industry explainers aimed at pharmacy also commonly list infant formula, vitamins, razors, and some OTC packs as frequent targets - useful for zone planning even when they sit outside the PPFF top-three list.
On methods, the study found most external theft used low-sophistication approaches (concealment, walking out unpaid). Planned techniques were category-specific. Booster bags were reported most commonly in pharmacies, then apparel - a signal that pharmacies should treat bag-lining and foil/metal-lined bags as first-class detection scenarios, not edge cases.
Distraction techniques also rose across categories versus the 2019 survey. AI that only watches a single camera crop without multi-person context will miss the classic “one distracts, one conceals” pattern.
Strip the marketing. A production pipeline looks like this:
Vendors differ on where steps 2-5 run (edge appliance vs cloud GPU), how many cameras are licensed, and whether alerts include face crops for staff approach. Those choices dominate privacy and latency more than the brand name on the box.
For a deeper walkthrough of a multi-model on-premises cascade, see how AI detects shoplifting in real time. For why false positives kill adoption, see what is a false positive and why it matters.
Australian pharmacies are health-adjacent environments. Customers expect discretion. That is why several commercial systems (including widely marketed gesture products) emphasise behaviour and body movement, not facial biometrics, for primary theft detection.
That distinction is load-bearing:
Both can appear in one product. Some stacks use face crops only as a staff visual for “who to approach”, stored briefly or as embeddings under staff-controlled policies. Others market chain-wide offender matching. Those are different legal and cultural products. Put the difference in the statement of work.
OAIC guidance is clear that security-camera footage can contain personal information when individuals are reasonably identifiable, and APP entities must notify, secure, and destroy or de-identify that information when it is no longer needed. Facial databases amplify that obligation. Gesture-first detection does not remove Privacy Act duties - it reduces the blast radius of what you collect for the primary purpose of loss prevention.
A 2026 Australian industrial comparison of edge vs cloud AI puts the practical split plainly: edge inference lands in milliseconds for real-time work; cloud round-trips often land in hundreds of milliseconds to seconds, depending on model and network. Loss-prevention vendors who stream frames for remote inference commonly quote multi-second alert paths once upload, queue, and push are included. One pharmacy-facing vendor publicly cites roughly seven seconds detection-to-alert on their stack - still useful versus Monday stocktake, but a different intervention window than sub-two-second local inference.
Australian operating realities push edge further than US metro demos imply:
| Decision | Cloud / hybrid upload | On-premises edge inference |
|---|---|---|
| Where frames are analysed | Vendor data centre (often overseas) | Appliance in your store |
| Typical alert path | Upload → queue → infer → push (seconds) | Infer locally → push (sub-second to ~2s) |
| APP 8 exposure | Cross-border disclosure of identifiable video may apply | No routine overseas disclosure of footage for inference |
| NBN outage behaviour | Detection often stops when upload fails | Core detection continues offline |
| Bandwidth cost at 8-15 cameras | Continuous HD upload is material | Only alerts / metadata leave the LAN if anything |
Hybrid designs exist (edge for inference, cloud for dashboards). The question to ask is whether raw pixels leave the site for routine inference, or only structured events (alert IDs, timestamps, optional short clips under your retention policy).
Side-by-side detail: edge AI vs cloud CCTV comparison and why on-premises AI beats the cloud for pharmacy security.
This is not legal advice. It is the diligence pharmacies should run with their adviser.
On-premises inference does not make you exempt from the Privacy Act. It removes a large class of overseas-disclosure and bandwidth failure modes that cloud-first video AI introduces by default.
EAS gates, locked cabinets, and CCTV playback still have roles. They fail the pharmacy pattern in predictable ways:
Organised and semi-organised crews exploit time: seconds in the bay, minutes to the carpark, days before stocktake. Real-time detection only helps if the alert arrives inside that first window and staff trust it enough to walk over.
Use this in demos. Score vendors in writing.
If a vendor cannot answer (1), (2), and (6) without marketing adjectives, you are buying a slide deck.
IntelliGuard is NeuraIQ’s managed, on-premises AI theft detection service for Australian retail and pharmacy. Inference runs on a NeuraIQ edge appliance in the store. Staff get fast face-photo alerts and optional floor display cues so the whole team - not only the person at the NVR - can respond. Video is not uploaded to cloud servers for routine inference.
Commercial shape: subscription pricing from $299-$369 AUD per store per month (excl. GST) depending on store count, appliance included, 30-day money-back guarantee. Pilot the highest-shrink site, measure incidents interrupted and dollars retained, then expand on evidence.
Book a walkthrough: request an IntelliGuard demo.
AI theft detection is not “smarter CCTV.” It is a real-time intervention system sitting on top of cameras you already own. Australian evidence says pharmacies lose margin on fragrance, skincare, and make-up to methods that include booster bags and distraction. Systems win when they (a) detect concealment gestures in seconds, (b) keep identifiable video processing local enough to satisfy operational and APP 8 reality, and (c) produce alerts staff will act on.
Demand architecture answers before brand stories. Then measure shrink on the floor - not accuracy on a demo reel.