AI in Loss Prevention: Whose Data Trained the Model That Scores Your Business?

By mid-2026, almost every loss prevention platform ships with AI in some form. Fraud scoring, alert prioritization, anomaly detection, forecasting on returns and voids: what differentiated a platform two years ago is close to table stakes today. The variance is no longer in whether a vendor has AI, but in how each one implements it, how the underlying data is structured, and which processes it touches.

What still separates them is whose data the models learned from. A fraud model is a compressed record of the transactions it trained on, so the training data decides what the model treats as normal and what it flags as risk.

Most platforms train their machine learning models on pooled historical data, drawn from a cross-retailer consortium. If your business sits close to the profile that pool describes, the scale is a real advantage, and for catching a consumer working the same return scheme across four different chains, nothing built on a single retailer's data will match it. But most companies do not sit at the center of that profile.

Agilence Analytics builds a separate model for every customer, trained only on that customer's data inside that customer's isolated environment. Nothing is pooled, and no retailer's model is shaped by another retailer's transactions.

Where a pooled model reaches its limit

An aggregate model performs well on patterns that repeat across its pool and worse the further your operation sits from the industry average. That gap is wide. A grocer's scale production waste in the deli has no equivalent in specialty apparel. A restaurant group's comp and void thresholds reflect manager discretion a mass merchant would never authorize, so routine behavior in one environment scores as suspicious against the other. A convenience operator sets its own tolerance for cash over and short as a business decision, not an industry constant.

An aggregate model reads that variation as noise. Alerts fire on legitimate local practice while behavior that is genuinely abnormal for your business, but ordinary across the industry, produces nothing. After a few months of both, investigators stop working the queue and start working from instinct.

Your definition of fraud, your rules, your model

With Agilence Analytics, your organization defines what counts as fraud, and that definition drives what the model learns. Your alert rules and thresholds stay yours, with AI adding prioritization on top of your logic rather than replacing it. When an analyst marks a transaction fraud or not fraud, that becomes labeled training signal, so the model converges on the risk definition your team actually applies rather than an industry average of it.

The governance answers are correspondingly short. Your data stays in your tenant, trains only your model, and is never pooled with another retailer's or used to build a foundation model.

Human-first AI, and what that actually means

Agilence AI is built to narrow the work rather than to do it. The models score transactions, rank alerts by likelihood instead of by rule trigger alone, and forecast where risk is building, so an investigator opens a queue that is already sorted by what deserves attention first. What happens after that is a person's call.

The boundary is deliberate. A fraud score expresses likelihood, not a finding of wrongdoing, so it sequences investigative work rather than concluding it. Prescriptive alerts carry a recommended action plan to the right person in the field, and your team writes that plan as part of the alert definition, which means the response reflects your policy rather than a model's guess at it. Where the platform automates, it automates the clerical layer: adding a comment, sending a notification, creating a task.

Two questions to ask any vendor

Ask whose data trained the model that will score your business and ask where your data goes once you hand it over. Both answers are architectural rather than configurable, so they will not change after you sign.

The Agilence AI overview covers how fraud scoring and alert prioritization work in practice.