[Retail](https://bfcs.ai/industry/retail/) · [Cheaper](https://bfcs.ai/cheaper/)

# Can AI Reduce Shrinkage Costs?

Retail crime now costs Australian retailers billions a year, and self-checkout has opened new ways to lose stock. Camera analytics that watch the checkout and the door promise to win some of it back. The savings figures mostly come from vendors and retailers themselves, and the privacy exposure is real: the Privacy Commissioner has already found two Australian retailers in breach of the Privacy Act over facial recognition.

Updated **October 2026** · 7 min read · Evidence-graded · 8 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Low confidence | Catches unscanned items; matching people is contested · Video analytics are built to spot non-scans and barcode switching at self-checkout. The Privacy Commissioner found Kmart's facial recognition system was of limited utility against refund fraud; law firms report that a tribunal found Bunnings' system, used with other controls, effective and suitable to identify known offenders. |
| Faster | Partly · Low confidence | Real-time prompts, no independent timing evidence · Alerts fire at the checkout rather than at the next stocktake. As The Grocer reported, customers corrected themselves in more than 80% of cases after an on-screen prompt, but the report does not say what generated the prompts. |
| Cheaper | Partly · Low confidence | Plausible at self-checkout; savings figures are self-reported · A vendor-commissioned Forrester model puts recovered revenue at 0.14% of the sales passing through equipped lanes for a composite retailer. Coles reports a 25 basis point improvement in its loss rate in FY25 but does not attribute it to any single tool. We found no independent trial isolating the effect of AI. |
| Safer | Conditional · Governance-gated | Lawful only with notice, governance and proportionality · Bunnings breached notice and governance obligations even after the tribunal accepted its limited purpose; the Kmart determination was still under review at the tribunal as at July 2026. The Privacy Commissioner says facial recognition must meet a high bar to be lawful. |

## The problem, in numbers

Shrinkage is the gap between the stock a retailer pays for and the stock it sells. Theft is a large part of it. Griffith University's 2024 ANZ Retail Crime Study estimates that Australian retailers lost almost 2% of turnover to crime in the 2024 financial year, $7.79 billion, with customer theft the largest driver of external losses. (Source 1) That is a research estimate, and the university's release does not set out its method.

Self-checkout has changed where the losses happen. In a 2022 survey of 93 retailers across 25 countries for the ECR Retail Loss Group, respondents estimated that self-checkout accounted for 20% of their unknown store losses, and that about half of those losses (48%) were deliberate rather than mistakes. (Source 2) These are the retailers' own estimates, and the study was part-funded by a grant from Everseen, a vendor of checkout vision systems. (Source 2) A 2026 ECR study drawing on data from 39 retailers in 11 countries, including Australia, found, as The Grocer reported, that theft losses rose by an average of 22% in the year after self-checkouts went in. (Source 3)

### The evidence, in four numbers

$7.79bn · lost to retail crime in Australia in FY2024, almost 2% of turnover · Griffith University estimate (2024 ANZ Retail Crime Study); method not stated in the release

20% · of unknown store losses that retailers attribute to self-checkout · self-estimate by 93 retailers in 25 countries; part-funded by a vision vendor (ECR, 2022)

0.14% · of revenue through equipped lanes recovered by camera analytics in a modelled composite retailer · vendor-commissioned Forrester study, four customer interviews; a model, not a measurement

25 bps · improvement in Coles' total loss rate in FY25 · company-reported; Coles cites loss technology among its investments but does not attribute the gain to it

## What improves, and why

Two kinds of AI are sold against shrink, and they carry very different risks. The first watches the transaction: cameras over the self-checkout detect an item that crossed the scanner without registering or a barcode that does not match the product. In the 2022 ECR survey these video and AI analytics were the least deployed controls, but more than half of grocers (53%) said they were trialling or planning non-scan analytics. (Source 2) In our reading, this is a narrow, well-defined task: the system compares what it sees with what the till recorded.

The second watches people. Facial recognition compares every shopper who walks in against a list of people the retailer has flagged. That is where Australia's privacy law has bitten. Bunnings ran such a system in 63 stores in Victoria and New South Wales from 2018 to 2021, capturing the faces of likely hundreds of thousands of people. (Source 6) Kmart used one in 28 stores from 2020 to 2022, scanning everyone who entered and everyone at returns counters. (Source 7)

> Checking the scan and checking the shopper are different tools. Only one of them puts every customer's face in the system.

## What the evidence actually says

**Savings figures are vendor or company numbers.** A Forrester study commissioned by Everseen modelled a composite grocery retailer, built from interviews with four customers, and estimated the system recovered revenue equal to 0.14% of the sales passing through equipped lanes, a figure taken from those interviews. (Source 4) Forrester makes no claim that other retailers would see the same result, and says the value varies with the balance a retailer strikes between false positives and loss prevention. (Source 4) Coles reports that it improved its total loss rate by 25 basis points in FY25 and kept investing in loss technology, at a lower level than the year before, but the company does not say how much of the gain came from any one tool. (Source 5) In our reading, both come from the vendor or the company, and neither isolates AI.

**Prompts work, but the source is unclear.** As The Grocer reported, the 2026 ECR study found customers corrected themselves in more than 80% of cases after an on-screen prompt, and one retailer saw walkaways fall 28% after installing exit gates, a physical control. (Source 3) The report as described does not say what generated the prompts. We found no independent trial measuring shrink before and after checkout vision alone.

**Facial recognition has a mixed legal record.** The Privacy Commissioner found Kmart's system "of limited utility" against refund fraud and that less intrusive options existed, and she weighed the estimated value of fraudulent returns against Kmart's total operations and profits. (Source 7) The Administrative Review Tribunal reached a different view on Bunnings on 4 February 2026. We could not open the decision, so this is as the OAIC and law firms report it. The tribunal affirmed the findings on notice, privacy policy and privacy governance, but set aside the finding that Bunnings collected sensitive information without consent, holding that a "permitted general situation" exception applied. A law-firm summary cites the tribunal finding the system, with other security controls, "effective and suitable to identify known offenders". (Source 8) On 5 March 2026 the Commissioner said she had not filed an appeal, and told retailers to treat the decision as a case study "rather than a green light". (Source 8) The OAIC said in July 2026 that the Kmart determination remains under review at the tribunal, with hearings scheduled for early 2027. (Source 8)

## Where it falls short

-   **No independent baseline.** The savings figures we found are vendor-commissioned models or company results that bundle many measures together. (Source 4), (Source 5)
    
-   **Physical controls compete.** Exit gates produce reported gains of their own, which in our reading makes the camera's share hard to separate. (Source 3)
    
-   **Many losses are mistakes.** Retailers estimate about half of self-checkout loss is not malicious, so in our reading treating every alert as theft risks accusing honest customers. (Source 2)
    
-   **Scale of capture.** Facial recognition scans everyone to find a few, which the Commissioner weighed against the value at stake in the Kmart case. (Source 7)
    

## Where governance decides the outcome

The Bunnings outcome is not a green light. Even after the tribunal accepted its purpose, the findings against Bunnings on notice and governance stood. (Source 8) The Privacy Commissioner says facial recognition "must meet a high bar to be considered lawful", that entities will need a detailed risk assessment before deploying it, and that the Privacy Act applies even to systems that hold a face for only milliseconds. (Source 8) The Commissioner's July 2026 guidance release says "a precautionary approach" is required under Australian law. (Source 8) In our reading, checkout vision that looks at products carries far less exposure than systems that identify people, but both need clear signage, a privacy policy that names them, and a human review before any customer is confronted. For the wider question of handling customer data with AI, see our review of [client confidentiality](https://bfcs.ai/articles/can-ai-protect-client-confidentiality).

## How to think about adoption

Measure shrink by cause before buying anything: how much is self-checkout, how much is organised theft, how much is administrative error. Start with tools that check the transaction rather than the person, and test them in some stores against comparable stores without them, so the result is not mixed up with new gates or staffing. Set alert sensitivity deliberately, because a vendor's own study says recovery depends on the trade-off with false positives. (Source 4) Treat facial recognition as a last resort that needs a documented necessity case, proper notice and a privacy impact assessment, and watch the Kmart review in 2027. Finally, report the saving in basis points of loss against a baseline, not as a vendor's return on investment.

## How we read this evidence

BFCS.ai does not fabricate figures. The Australian crime cost is a university estimate, the self-checkout loss shares are retailers' own estimates, the savings figures are vendor-commissioned or company-reported, the 2026 ECR findings are as reported by The Grocer, and the tribunal''s Bunnings findings are as reported by the OAIC and law firms because we could not open the decision. We found no independent trial that isolated the effect of AI on shrink, and we say so where it matters.

## Sources

1.  Griffith University, "Cost of retail crime up 40 per cent over two years," media release on the 2024 ANZ Retail Crime Study by Professor Michael Townsley, 16 December 2024. [news.griffith.edu.au](https://news.griffith.edu.au/2024/12/16/cost-of-retail-crime-up-40-per-cent-over-two-years/)
    
    **What we took from it:** Almost two per cent of turnover was lost in the 2024 financial year, $7.79 billion in Australia alone; customer theft remained the largest driver of external losses; employee theft cost millions annually; customer aggression reached record levels. The release does not state the sample or method.
    
2.  Beck, A., "Global Study on Self-checkout in Retail: Use, Impact and Control," ECR Retail Loss Group research paper, University of Leicester, March 2022. [rgis.co.uk](https://www.rgis.co.uk/wp-content/uploads/2023/04/ECR_Global_Study_on_Self-checkout_in_Retail.pdf)
    
    **What we took from it:** Online survey of 93 retailers from 25 countries (51 European, 29 North American, 11 Australasian), nearly three-quarters grocery. The foreword thanks Everseen Ltd for an additional research grant supporting the study. Overall, respondents estimated 20% of unknown store losses were due to self-checkout (23% among the 32% who said they had researched the figure); malicious behaviour was estimated to account for 48% of all self-checkout losses; an earlier ECR report estimated each 1% of sales through fixed self-checkout adds 1 basis point of unknown loss. The report lists video/AI analytics for non-scanning, product recognition and barcode switching among available controls. Copy accessed via a hosted PDF:
    
3.  The Grocer (George Nott), "Self-checkouts increase store losses by 22% in first year, major study finds," 17 June 2026, reporting the ECR Retail Loss self-checkout loss study led by Professor Matt Hopkins, University of Leicester. [thegrocer.co.uk](https://www.thegrocer.co.uk/news/self-checkouts-increase-store-losses-by-22-in-first-year-major-study-finds/720246.article)
    
    **What we took from it:** As reported: data from 39 retailers in 11 countries including Australia; theft losses rose by an average of 22% in the year after self-checkouts were installed; losses were 33% higher in stores with self-checkouts than without; mis-scanning was reported in 1% to 4.8% of transactions; customers self-corrected in more than 80% of cases after an on-screen prompt; one retailer reported a 28% drop in walkaways after installing exit gates. We could not open the report itself.
    
4.  Forrester Consulting, "The Total Economic Impact of Everseen Evercheck," commissioned by Everseen, September 2024. [tei.forrester.com](https://tei.forrester.com/go/everseen/evercheck/)
    
    **What we took from it:** Vendor-commissioned study (not peer-reviewed). Based on interviews with four representatives of customer organisations, combined into a composite grocery-focused retailer with $20 billion revenue and 200 locations. Modelled results: shrink recovered equal to 0.14% of revenues (about $88,000 per store a year); ROI 374%; payback under six months. In the study's own table the 0.14% comes from the interviews and is applied to the revenue running through Evercheck lanes (65% of the composite's $20 billion), before a 15% risk adjustment. Everseen reviewed the study and gave feedback; Forrester says it kept editorial control. Forrester "makes no assumptions as to the potential ROI that other organizations will receive", and says the value of reduced losses will vary with "the desired balance of false positives and loss prevention."
    
5.  Coles Group Limited, "2025 Full Year Results Release," ASX announcement, 26 August 2025. [announcements.asx.com.au](https://announcements.asx.com.au/asxpdf/20250826/pdf/06nb8z0vs5n4gj.pdf)
    
    **What we took from it:** Company-reported: "improved total loss rate by 25 basis points"; gross margin expansion included a 25 basis point improvement in loss (39 basis points in 1H25 and 10 in 2H25); investments in loss technology continued, at a lower level than the prior year. Loss here is total loss, not theft alone. The release does not attribute the improvement to any specific technology.
    
6.  Office of the Australian Information Commissioner, "Bunnings breached Australians' privacy with facial recognition tool," media release, 19 November 2024, and "Bunnings determination: use of facial recognition technology" factsheet. [oaic.gov.au](https://www.oaic.gov.au/news/media-centre/bunnings-breached-australians-privacy-with-facial-recognition-tool)
    
    and [oaic.gov.au](https://oaic.gov.au/__data/assets/pdf_file/0027/243936/Bunnings-determination-factsheet.pdf)
    
    **What we took from it:** Bunnings operated the system in 63 stores in Victoria and New South Wales between November 2018 and November 2021; it captured the faces of every person who entered, likely hundreds of thousands of individuals, and compared them against a database of people Bunnings had identified as posing a risk. The Commissioner found collection without consent and failures of notice, governance and privacy policy (APPs 3.3, 5.1, 1.2 and 1.3). The OAIC's July 2026 guidance release refers to 62 stores.
    
7.  Office of the Australian Information Commissioner, "Kmart's use of facial recognition to tackle refund fraud unlawful, Privacy Commissioner finds," media release, 18 September 2025. [oaic.gov.au](https://www.oaic.gov.au/news/media-centre/18-kmarts-use-of-facial-recognition-to-tackle-refund-fraud-unlawful,-privacy-commissioner-finds)
    
    **What we took from it:** Kmart used facial recognition in 28 stores from June 2020 to July 2022, capturing everyone who entered and everyone at returns counters, including many thousands not suspected of refund fraud. The Commissioner found the system of limited utility in preventing fraud, that less privacy-intrusive methods were available, and weighed "the estimated value of fraudulent returns against the respondent's total operations and profits" (the release gives no figures). The OAIC's July 2026 release dates the determination August 2025.
    
8.  Bunnings Group Limited and Privacy Commissioner (Guidance and Appeals Panel) \[2026\] ARTA 130, decided 4 February 2026. [claytonutz.com](https://www.claytonutz.com/insights/2026/february/bunnings-wins-appeal-on-facial-recognition-technology-key-takeaways-for-businesses-and-privacy-law)
    
    HWL Ebsworth, "Smile, you're on in-store camera! Tribunal decision leaves scope for facial recognition CCTV" (dash in title replaced), 5 March 2026: [hwlebsworth.com.au](https://hwlebsworth.com.au/smile-youre-on-in-store-camera-tribunal-decision-leaves-scope-for-facial-recognition-cctv/)
    
    Pointon Partners, "To Scan or Not to Scan: Administrative Review Tribunal decision on Bunnings' use of facial recognition tech" (dash in title replaced), 8 March 2026, citing paragraph \[156\] of the decision for the finding that the collection, with the other security controls, "was effective and suitable to identify known offenders". [legal500.com](https://www.legal500.com/intelligence/australia/privacy/to-scan-or-not-to-scan-%E2%80%93-administrative-review-tribunal-decision-on-bunnings%27-use-of-facial-recognition-tech)
    
    OAIC, "Privacy Commissioner statement on Administrative Review Tribunal's Bunnings decision," 5 March 2026: [oaic.gov.au](https://www.oaic.gov.au/news/media-centre/privacy-commissioner-statement-on-administrative-review-tribunals-bunnings-decision)
    
    OAIC, "Privacy Commissioner publishes updated guidance on facial recognition in retail spaces," 29 July 2026: [oaic.gov.au](https://www.oaic.gov.au/news/media-centre/privacy-commissioner-publishes-updated-guidance-on-facial-recognition-in-retail-spaces)
    
    **What we took from it:** We could not open the decision itself (AustLII refused automated access), so the outcome is as reported by the sources below. Clayton Utz, "Bunnings 'wins' appeal on Facial Recognition Technology: Key takeaways for businesses and privacy law," 9 February 2026: the Tribunal affirmed the APP 1 and APP 5 findings and set aside the APP 3 finding (collection of sensitive information without consent), holding that a permitted general situation under APP 3.4 applied; it accepted the use as a "reasonable and proportionate response to legitimate business concerns."
    
    APP 3.3 set aside; APP 5.1 (notice), APP 1.3 (privacy policy) and APP 1.2 (no documented privacy impact assessment) affirmed.
    
    Bunnings "was entitled to use FRT for the limited purpose of combatting very significant retail crime"; the findings on notice and policies and procedures were not disturbed; "I have not filed an appeal of this decision"; entities "will need to conduct a detailed risk assessment specific to their circumstances before deploying the technology"; facial recognition "must meet a high bar to be considered lawful under the Privacy Act"; Privacy Act safeguards apply even to technologies that keep data for only milliseconds.
    
    the Kmart determination remains under review at the Tribunal, with hearings scheduled for early 2027; "A precautionary approach to the deployment of FRT is required under Australian law."
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Make Recommendations Safer?](https://bfcs.ai/articles/can-ai-make-recommendations-safer/): More in Retail · RETAIL · SAFER · 7 min · Read →
- [Can AI Speed Up Fulfilment?](https://bfcs.ai/articles/can-ai-speed-up-fulfilment/): More in Retail · RETAIL · FASTER · 8 min · Read →
- [Can AI Reduce Discovery Costs?](https://bfcs.ai/articles/can-ai-reduce-discovery-costs/): Also Cheaper · LEGAL · CHEAPER · 3 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-reduce-shrinkage-costs/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml