POS exception reporting vs AI video: what each one catches

What POS exception reporting is

POS exception reporting — also called exception-based reporting — is loss prevention software that reads the transaction log every register already produces and flags the entries and the people that stand out. It looks at the functions that can hide employee theft: voids and post-voids, refunds and no-receipt returns, no-sale drawer openings, price overrides, employee discounts, coupons and manager overrides. It then ranks cashiers, registers and stores by how far they are from the norm.

It is a mature, sensible tool, and it does several things well:

• It uses data you already have. The POS records every transaction; no new hardware is needed to read it.
• It covers every transaction. Nothing is sampled. Every void, refund and override is counted.
• It is inexpensive to run compared with anything that needs people or cameras at each register.
• It ranks where to look. A cashier with an unusual rate of voids, refunds, no-sales or discounts rises to the top of the list.
• It shows trends over time. A store whose refunds climb month after month is visible at head office.
• It works without cameras, so it covers registers that no camera sees.

Many setups go one step further with video-linked reporting: each flagged transaction carries a link to the recorded footage at that minute, or the transaction text is overlaid on the video. That makes manual review much faster, because the reviewer no longer has to search for the right camera and the right minute — but a person still has to watch each clip and decide.

What AI video checked against the POS log is

AI video checked against the POS log starts from the other end: the action at the register. Primarch builds on and enhances vision-language models (VLMs): models that read video and language together. A VLM reads a window of time, not a single frame, so it can recognize an action — an item passed around the scanner, cash handed over, a drawer opened with nobody at the counter — and each action is matched with what the POS recorded in the same second.

• It reads the action, not just the entry. It can recognize that a sale was paid and handed over before it was voided, or that a refund was entered with no one at the counter.
• It can see events that have no POS line. Cash handled while the POS shows nothing — an unrung sale, an unlogged payout — has no entry for a report to count, but it does happen in front of the camera.
• It writes the finding in plain language, with a clip. What happened, at which register and why it doesn't match the log, with a time-stamped evidence clip and the matching transaction line.
• It chains events into patterns. Findings are linked by store, register, cashier and hour into 30-day patterns, so one event is not treated as a verdict.

What each one catches

The two approaches overlap on schemes that leave a POS entry, and part ways on schemes that don't. Scheme by scheme:

Void after payment

Exception report: a cashier with more voids than their peers stands out, but an honest void and a void of a paid, handed-over sale look the same in the log. AI video + POS: can recognize on video that the sale was paid and handed over before it was voided. See void and refund fraud.

Fake refund with no customer

Exception report: flags unusual refund counts, no-receipt returns and refunds near closing. It can't tell whether anyone was at the counter. AI video + POS: checks whether a customer is at the counter when the refund is entered and whether anything came back. Customer-side return fraud is covered in return fraud detection.

No-sale drawer openings

Exception report: counts no-sales per cashier and hour. AI video + POS: during a no-sale, can tell cash going in from cash going out and whether it was handed across, and flags a drawer opened with a key or left open with no POS entry. See cash skimming and no-sale drawers.

Missed scans and sweethearting

Exception report: largely blind. An item that was never rung leaves no line to report; at most, unusually small baskets or short transactions hint at it. AI video + POS: recognizes an item passed around the scanner with no matching POS line, and compares the item on camera with the product on the POS line for under-rings. See sweethearting and missed scans.

Discounts and price overrides

Exception report: strong at spotting unusual discount, coupon and override rates by cashier. It can't see who the discount went to. AI video + POS: reads the moment the discount or override was entered alongside what happened at the counter. See discount, override and loyalty abuse.

Cash with no POS entry at all

Exception report: blind by design — there is no transaction to report. AI video + POS: flags cash handled while the POS shows nothing at all, such as unrung sales and unlogged payouts or pickups.

An exception report answers "where should we look?" AI video checked against the POS log answers "what happened at the register when this entry was made — or when no entry was made at all?"

Where each falls short

POS exception reporting

• Counts without context. It shows how many, how much and by whom, not what happened at the counter.
• The drawer still balances. When the POS record is changed to match the missing cash, over/short reports stay clean, so the theft doesn't show in the cash count.
• It can't see what never reached the POS. Unrung sales, missed scans and cash handled with no entry have no line to flag.
• Honest outliers look suspicious. The busiest cashier, the one who handles the returns desk or the shift that always processes online-order corrections can top the list without doing anything wrong.
• Someone still has to watch the footage. Even with video links, every flagged entry needs a person to find, watch and judge the clip, so review tends to happen only for the top of the list.

AI video checked against the POS log

• It needs a camera that covers the register. What the camera doesn't see, it can't read; registers without coverage stay with the reports.
• It needs a privacy setup. Analysis limited to the checkout zone, anonymization, and where footage is processed all have to be agreed before it runs.
• It needs the POS log. The integration has to be set up and validated, usually during a pilot.
• Findings still go to a person. A plain-language finding with a clip makes review faster and fairer, but the decision about what to do stays with people and the store's own rules.

Using them together

The two approaches are complementary rather than rivals. A practical way to combine them:

• Keep the reports for breadth and trends. They cover every register, including those no camera sees, and show which stores and cashiers drift over months.
• Use AI video at the registers where the risk sits. Each void, refund, no-sale and override there is checked against the moment it was entered, and cash handled with no entry becomes visible.
• Let each answer the other's question. When a report puts a cashier at the top of the list, the video findings show whether the entries behind it were honest. When video shows a pattern, the report shows whether it is part of a wider trend.

For an overview of the schemes themselves and how they are caught, see how to catch employees stealing from the cash register.

What to look for in a solution

Whichever vendor or combination you choose, these are fair questions to ask:

• Reads your POS transaction log. Which POS exports can it read, and how is the integration validated?
• Works on your existing cameras. If video is part of it, can it connect to the IP cameras you already have?
• Shows what happened, not just a count. Does a finding come with the transaction line and the moment on video, or only a score?
• Covers events with no POS entry. Ask directly how it handles unrung sales and cash handled with no transaction.
• Handles false alarms openly. How are findings reviewed by a person, and how are single events separated from repeating patterns?
• Produces evidence usable in HR processes. A time-stamped clip plus the matching transaction line, not a bare alert.
• Respects privacy. Is there an on-premise option where footage never leaves the site? Is anonymization applied, and does it meet GDPR and other data-protection requirements?
• Can be piloted on your own registers and data. Not a demo video.

How Primarch's Loss Prevention Expert helps

This section is about our product. The Loss Prevention Expert is the Primarch expert module for the checkout. It recognizes actions at the register — scan, skip, void and refund — and cross-checks each one against the POS record in the same second. Alongside the reports you already run, it:

• Recognizes on video that a sale was paid and handed over before it was voided.
• Checks whether a customer is at the counter when a refund is entered.
• Tells cash going in from cash going out during a no-sale, and flags a drawer opened with a key or left open with no POS entry.
• Flags cash handled when the POS shows nothing at all: unrung sales, unlogged payouts or pickups.
• Compares the item on camera with the product on the POS line to catch under-rings, and recognizes items passed without a matching line.
• Chains events by store, register, cashier and hour into 30-day patterns: single events are alerted for review, and the patterns show what repeats.
• Produces an evidence file for each finding: a time-stamped clip plus the POS record, usable in HR processes.

Putting it to work

1. 01

   Connect

   Connect your existing RTSP cameras and the POS transaction log export; no new cameras or special hardware. It is designed to work with any common POS system that can export transaction logs; the integration is validated together during the pilot.

2. 02

   Pilot on your own footage

   Run it on your own registers, camera angles and POS data, and review the findings together — next to the reports you already use.

3. 03

   Roll out

   Extend to more registers and stores, with central monitoring across the chain.

• Alerts, digests and dashboards. Instant notifications, periodic digests and role-based dashboards, so each finding reaches the person who acts on it.
• Ask the past in plain language. Findings are written into a queryable event memory: "Which no-sales were flagged at register 3 last week?"
• Privacy by design. With on-premise deployment, footage never leaves the facility. Analysis focuses on the checkout zone with privacy-compliant anonymization; the goal is action–transaction consistency, not tracking people.