A decision framework from a detective
Rish Gupta says every big call at Spot AI runs through one line of Sherlock Holmes. The same four steps are a sensible way to decide what a video agent is allowed to do on its own.
On September 29, 2025, Rish Gupta, the CEO of Spot AI, posted that "every major decision we make at Spot AI gets filtered through a quote from a detective novel published in 1892." The quote is a familiar one.
When you have eliminated the impossible, whatever remains, however improbable, must be the truth.
"I've read Sherlock Holmes four times," he wrote. "This line haunts my decision-making."
The framework in his words
Gupta reduces the line to four steps, quoted here exactly because the ordering matters:
1. Define the outcome you MUST achieve
2. List the fastest possible path to get there
3. Eliminate the paths that are physically impossible
4. Take whatever path remains - no matter how crazy it sounds
The reframing he attaches to it: "Stop asking 'What's most likely to work?' Start asking 'What's the only path to the outcome I need?'"
He gives two examples. In 2020, "rather than launching a camera company like several competitors," Spot AI took "the improbable path but the only path to build a video AI company - focusing on the video and not the cameras." His note on how that aged: "Every entrant including incumbents had to launch a camera agnostic play eventually." In 2024, "when everyone was busy launching peripheral accessories like sensors and access control, the improbable path was to bet on humans won’t watch videos - AI will."
The post is about corporate strategy. I want to borrow it for a narrower question that comes up in every video agent deployment: what should the agent be allowed to do without asking a person first?
Step one: start from the outcome
Most agent rollouts start from the vendor's catalogue: a loitering agent, a PPE agent, pick a few. Gupta's first step inverts that. Start from the outcome, and write it down as a result. "No dumpster fires on the back lot" is an outcome. "Alert on loitering" is a path, and one of several.
The outcome also fixes the tolerance. If it is no forklift injuries this year, a false alarm costs a minute of a supervisor's time and a missed event costs a hospital visit, so you will let the agent do more on its own than you would if the outcome were fewer managers sitting in the back office.
Step two: the fastest path, honestly listed
The fastest path to no dumpster fires at 2 a.m. is that someone tells the person by the dumpster to leave, at 2 a.m., before the fire. Written down like that it is obvious, and it is also something a motion alert cannot deliver, since a motion alert only notifies someone who may or may not be awake. The exercise is to write the fastest path without regard to whether your current system can do it, and prune from there.
Step three: eliminate the physically impossible
This is the step a security director tends to skip, and Gupta's version is strict. Physically impossible is the test; expensive and unfamiliar don't count. A person watching every camera all night is impossible for the team a district AP manager described at a Convenience Store News webinar this August as "three guys" covering 120 stores. Verifying dozens of motion alerts a night in time to act is impossible if, as Spot AI's own webinar material claims, nine in ten are false. Reviewing footage the next morning prevents nothing.
Once those are gone, what remains is a machine that looks at the clip, decides whether the person by the dumpster is the one you care about, and says something to them. In 2020 that would have sounded absurd.
Step four: take it, with limits
Gupta's fourth step says take the remaining path "no matter how crazy it sounds." For an operator, that means being precise about the action. An approved first action is bounded: a spoken warning, a locked door, a text to the manager on shift, a logged clip. Everything past that first action goes to a person.
The published Staccato case study shows where those limits sit in practice. Mike Tiller, the firearms manufacturer's director of technology, said the agents "were smart enough to recognize known staff members and adjust their monitoring and actions accordingly," because range safety officers and the general public needed different PPE rules. The outcome was fixed, the path that treated everyone the same was eliminated because it would have produced alarms the team learned to ignore, and what remained was an agent that knew the difference.
Gupta closes his post with a sentence written as advice for founders. It reads just as well as a rule for configuring an agent: "Then take that path. However improbable it seems."