The mandate widened, the team didn't
The argument Rish Gupta made to loss-prevention audiences this summer: retail asked its security teams to cover more stores, more footage and more kinds of incident, and gave them cameras that only watch.
Twice this summer Rish Gupta, the CEO of Spot AI, made the same argument to retail loss-prevention audiences, first at a Loss Prevention Magazine webinar on June 17 and then at a Convenience Store News webinar on August 11. The line that stayed with me from the first one is six words long and sits on a slide with no image at all.
The mandate widened. The team didn't.
What widened
The LPM deck tells the history of cameras as a series of jobs: see in the 1940s, record in the 1970s, reach from anywhere in the 2000s, understand in 2024 and 2025, and, the claim for 2026, act. "Cameras used to watch. Now they act. Capture got easier. Then access. Then search. Now AI makes it easier to act."
The catch, in Gupta's telling, is that every step before the last one made more work for the same people. "More cameras than ever. A person still does everything that comes next. Footage doesn't resolve anything." The deck puts numbers on that: 90 percent of motion alerts are false positives, and 40 percent of in-store security tasks still need a human watching.
The CSNews deck lets the audience say it themselves. On a slide titled "What LP/AP at a c-store sounds like right now," a director of global security and loss prevention is quoted: "Everyone I talk to — we're two, three people short. Everyone's working at 125% capacity." A district asset-protection manager: "Three guys go through 120 stores trying to stop shoplifters." A director of AP: "Average internal theft investigation is going to take 40 hours of man hours." The same slide notes that roughly one in five overnight alarms turns out to be real.
The 2:13 a.m. problem
The LPM deck's central scenario is deliberately ordinary.
It's 2:13 a.m. A truck pulls into the far side of your parking lot. It's outside the guard's line of sight. Motion alerts trigger — but there are dozens of alerts every night. No one verifies it in time. By morning, you're reviewing footage. Catalytic converters are gone. Or the dumpster's been set on fire.
Gupta's contrast is between two loops. The traditional one runs detect, alert, wait for a person. The agentic one runs detect, read context, take an approved first action, check the outcome, then escalate or document. The deck compresses it to three verbs: see (old computer vision, a person or a car), understand ("a customer browsing vs. someone casing") and act, which "takes the first step: talk down, notify, document."
In the dumpster example the first step is a voice. The deck's sample talk-down script reads, in full: "Person in the black hoodie - please step away from our dumpster. Leave the area now." The outcome claim attached to it, from a $5 billion retailer the deck does not name, is $20,000 saved per incident per store on dumpster fires and illegal dumping.
Detect, deter, resolve
Spot AI's 2026 messaging organises the pitch as detect, deter, resolve, and both webinars follow that order. Detect is what the industry has had for years, done badly by motion. Deter is the new part, an approved automatic response before a person is involved. Resolve is where LP teams spend their 40 hours, and the argument is that an agent that already found the clip, logged the event and issued the warning has done most of the paperwork before anyone opens a case. As one LPM slide puts it, "The difference is not merely better detection. It's in the follow-through."
The CSNews version carries the sharper customer proof because it can name one. Blackmon Oil, an 11-location convenience chain, is the published case, and the deck's line on it is that Blackmon Oil "didn't buy security. They bought answers." Kelsey Blackmon, the company's VP of marketing and technology, has described how the chain uses it: "We've set up a Spot AI vehicle loitering filter to keep our parking lots clean and safe, especially for our 24/7 stores. Even with a smaller overnight crew, we want eyes on the lot at night."
The caveat I would add
Both decks are vendor decks, and the outcome figures in them are the vendor's. What holds up regardless, I think, is the diagnosis. Retailers added stores, cameras, footage and incident categories faster than they added people to watch them, and a camera that only records shifts the burden onto the morning after. Gupta's closing slide at LPM was an instruction: "Stop letting cameras collect dust. Let your team of agents deliver outcomes." Whether the agents deliver is something an LP director should test on their own lot. The dust I am prepared to take on the evidence of the 40-hour investigations and the one-in-five overnight alarms.