CrimeScience Episode 255: The Emerging Role of AI in Loss Prevention (feat. Sam Yeung, PhD)
00:00Hi everyone and welcome to crime science.In this podcast, we explore the science of crime and the practical application of this sciencefor loss prevention and asset protection practitioners as well as other professionals.Hello everyone and welcome to this edition of the crime science podcast here at LPRC.I am your host Bo Nutter.I'm a research scientist here at LPRC and today we have an episode with another research scientist,Dr. Sam Young, and we will be talking about his recent reports on AI in loss prevention from models to operational value.All right, Sam, let's get right into it.
00:39So for the listeners who may not be familiar with this topic, can you give a broad overview of this report and what it tells us about the role of AI in loss prevention?At a broad level, this report is really about closing the gap between what AI can do in a demo.and what actually improved laws prevention operations.The report grew out of a presentation by Rob Saker from Database to LPRC's Data Analytics Working Group.My goal was to take the idea from that presentation and translate them into practicalquestion that LP professional can use when evaluating an AI solution. I think the central messageis fairly simple. AI should help us identify a problem sooner, make a better decision,or take action while there is still time to make a difference.
01:31For example, that might mean identifying a missed scan at self-checkout before the transaction is completeor helping an investigator search through hundreds of incidents report to identify a connectionthat would otherwise take hours to find.The report focused on three things.First, start with an actual operation problem.It's not a product.Second, define what success look like.how you're going to measure it and third establish the data controls governance and human oversight
02:03needed to use the technology responsibly.Oh right. So one of the central ideas that you just mentioned was that LP team should startwith the problem rather than the AI tool. Why is this distinction so important?I think the probably is one of the most important points in an entire report.Right now we are seeing tremendous amount of excitement around AI from newer models to newer modelswrappers or AI products. New tools are coming out constantly. But there's a tendency to startwith the technology and then look for a problem to solve. For example, someone may say,we should implement an AI chatbot to our investigators. But my first question would be,
02:43what exactly are investigators having trouble with? Is it that they spend too much timesearching incident report? They're having difficulty connecting cases across multiple stores.Those are different problems and they may require different solutions. Once we identify the problem,we can establish a baseline. Let's say investigator currently spent two hours reviewing cases material.Can the proposed solution reduce that to one hour without sacrificing accuracy andcomplicit now we have something measurable? And importantly, AI may not even be the best answer,sometimes improving a process, changing how information is organized, or using an existing analytical toolmight solve the problem at a lower cost and without the learning curve.
03:35Right. Yeah, that makes sense to me. Start with your problem before you go and get your solution.So the report makes a distinction between AI applications that are more established and those that are still emerging.How should LP practitioners think about that difference when evaluatingpotential AI use cases.Thank you, Bo.I think it's important to distinguish between two things,how mature an AI application is and how strong evidence is that it actually works.Those are unnecessarily the same thing.
04:09Take self-checkout loss detection, for example.Among the use cases presented in the briefing,missed scan detection had the strongest LP-specific research example.The briefing pointed out an academicfield study evaluating missed scan problems. That gives us a more concretestarting point than a demonstration loan. But even then it doesn't mean that thesame solution will produce identical results in every retail environment. Storelayouts, customer traffic technology integration and operating procedures may all
04:45affect outcome. Now compare that with AI application for linking organizedretail crime cases acrossstores or retailers. Those capabilities are promising, but the briefing character characterise them as emerging. We still need more direct evidence about their accuracy and their impact on investigations. The report introduces four categories of evidence that I think are practically useful. The first one being confirmed evidence means that there's independent, reproducible, and clearly sourced support. The second, the same.Second being observable evidence means a result has actually occurred perhaps in a single deployment or an adjacent industry,but it hasn't necessarily been demonstrated in your specific LP setting.The third is projected evidence, referred to reasonable expectation that hasn't yet been demonstrated.Last but not least is the estimate evidence typically involved model figure or assumption such as projected cause,or return on investment.
05:55We might ask ourselves, why does that matter?Because of Vandum might demonstrate an AI systemthat is very good at identifying unusual patterns,but that's different from demonstrating that the systemwe do shrink, improve apprehension outcomes,or save investigation time.We shouldn't treat those claims as interchangeablebecause they really are not.
06:16Okay, I get what you're saying.So what were the most important AI use casesor capabilities that stood out in the reportreport for loss prevention practitioners.I will highlight three broad capabilities along withCERFO specific application. The first is making unstructuredinformation easier to use. LP organization have enormousamount of information in incident narrative,
06:39videos, reports, or other records. Traditionally, much of thatinformation has been difficult to search or analyzesystematically. AI create opportunities to summarize search andconnect information with other sources, such as transaction records, inventory, and case data.For example, an investigator could potentially use natural language to search casenarrative for incident involving similar behaviors such as a phone scam rather than manuallyreading every report. Of course, the summaries and search results still needs to be checkedfor omissions and error. The second capability is identifying patterns that traditional
07:22rules might miss. A traditional rule-based system might flag a transaction when it exceed a certaindollar amount or satisfy certain rule, but suspicious activity doesn't always follow a singlepredefined set of rules. Consider account takeover. An individual may change account information,use level delivery addresses, and show unusual purchasing behavior. Each activity by itselfmight not trigger an alert. But a model examiner,the combination of those signals could potentially identify something worth investigating.That doesn't eliminate false positive, but illustrate where AI might add value beyond fixed,exceptional rule-based type of system. The third capability is supporting the investigation
08:10workflow. This could involve summarizing cases, identify a potentially related incidentor helping investigators retrieve relevant information more efficiently.In terms of specific use case, the report discusses four areas.Self-checked out mis-scan detection is one, and it has the strongest LP-specific research example in the briefing.Fraud and return abuse is another. Here, AI might help detect unusual account or transactionpatents, although much of the supporting evidence come from a jacian fraud applicant,Then there's organized retail crime network language where systems might help identifyrelationship among incident people, stores, vehicle, or stolen goods.
08:58And finally, AI assists the case investigation, including document search and summarization.Okay.So the report argues that governance is not simply a restriction on AI, but something that makesAI deployable.What does good governance look like in an LP environment?I think governance is sometimes misunderstood as something that slow innovation down, but in laws prevention, we are frequently working with sensitive information or public perception, such as including a transaction record, customer data, employee information, and potentially PII.Without appropriate safe guard, even a technologically impressive AI system might not be suitable for operational use.The report identifies three main layers of governance.
09:48First is data governance.Where does the data come from?Who has permission to access it?And we all know that from the recent uproar to one of our partners,it really is a big deal.What information needs to be masked?How is it assessed?And what happens to the information when it's shared with a technology
10:13provider. Those are questions that need to be answered before deployment. The second is a consistentbusiness definition. This might sound less exciting than the AI itself, but it's extremely important.For example, two retailers might define shrink and incident or a link case differently. If we arefitting inconsistent definition into a system, the AI could generate output that appear consistent,but actually represent very different things to different organization.So data quality and common definitions are fundamental. The third layer is AI system oversight.We need to know which models we are using, which version is operating, which it was evaluated,and whether its performance remain acceptable over time, and whether a model that, say, you pull from
11:02hugging face is still serviceable. And importantly, there needs to be a way for a person toreview challenge or override consequential outputs. For example, if AI suggests that two incidentsinvolve the same individual, that shouldn't automatically establish identity. And investigators still needto review the supporting evidence. There's also an important principle from the briefing. We canrent the technology, but we should retain control of our data. Yeah, I agree with you. Government isextremely important now, especially nowadays with some of our partners know well aboutthat. So last thing here, what is the main takeaway you want LP practitioners to have when theybegin evaluating or testing AI? Where does this work need to go next? I would encourage LP
11:51professional to start small, but start with a clearly defined problem and a measurable outcomethat the organization want to go to. Before beginning, I'm pilot, asked a couple questions.What problem are we trying to solve? How will we know whether theAI system actually held and what evidence would justify expanding the deployment.I think the measurement question is especially important. We are often attempt to focus immediatelyon shrink reduction or return on investment. Those are definitely important outcomes, but they can bedifficult to attribute to a single technology. Sometimes the first useful measures are operational.The investigators identify a related case.
12:38faster or have a broader net to catch offenders that might have been missed if AI are not used,the force positive alert decline, etc. Those are measurable outcomes that help us understandwhether the technology is changing the process in a useful way. So looking ahead, I think we needmore real world evaluation that compares AI-supported processes against appropriate baselines. We also need tounderstand how these systems perform across different retail environments and how to maintain privacyand governance as information becomes more interconnected. The briefing I raised some interestingfuture possibilities including privacy preserving collaboration among retailers, but those ideasstill require research and validation. So I wouldn't encourage LP teams to chase every new
13:34AI capabilities or the latest model. I would encourage them,to be disciplined about identifying problems, testing solution, and documenting result.And I think the best way to summarize the report is with one question. Did AI help us know sooner,decide better, or act in time to matter? If we can demonstrate that, then we are moving beyondAI as a technology demonstration and taught AI as a source of real operational reading.All right. I think that brings us to the end of our podcast today.I know that I learned a lot just in this short time listening to you talk about AI.But thank you so much for being here today, Sam.
14:15If anybody wants to read this, it should be on the Knowledge Center, so you can read the full report there.And with that, thank you all for listening today.And we will see you next time on the Crime Science Podcast.Thanks for listening to the Crime Science Podcast, presented by the Lost Prevention Research Council.If you enjoyed today's episode, you can find more crime science episodes and valuable information.at LPRsearch.org.The content provided in the crime science podcastis for informational purposes only
14:45and is not a substitute for legal,financial, or other advice.Views expressed by guestsof the crime science podcastare those of the authorsand do not reflect the opinionsor positions of the Loss Prevention Research Council.
There aren't comments yet for this episode. Click any sentence in the transcript to leave one.