Computer vision
You already have cameras. The question is whether anyone watches them, and when: almost always after something has happened.
Service status
A line in development. It is not yet a product you can buy with a catalogue and a price. We publish this page to talk to companies willing to take part in the first projects — and so that nobody buys a promise.
In short
The computer vision is the use of artificial intelligence so a system understands what is happening in the footage rather than only recording it: counting people, detecting that someone entered a restricted area, checking that staff are wearing a helmet and vest, or linking each register void to the exact second of the recording. IstmoDigital is developing this line in Panama, where there is currently almost no local supply. It is real, bounded technology: it works for watching processes and raising alerts, not for reliably identifying people.
Every void or manual discount is linked to the second of the recording. Reviewing twenty voids goes from an afternoon to ten minutes. It is the use with the best return.
Checking helmets, vests and footwear at site or plant entrances, with a log. A supervisor cannot stand at the entrance all day; a camera can.
An alert when someone enters a warehouse out of hours or crosses into a restricted area. An alert, not an accusation: a person reviews it.
How many people come in, at what times and where they walk. Useful for deciding where to place product and how many staff per shift.
Detecting that a space has had no product for hours. The sale that never happens because the shelf was empty shows up in no report.
It does not reliably identify thieves, it does not replace staff and it does not fix badly installed cameras. A backlit camera does not improve because you added AI to it.
Three contexts where the return is clear and measurable:
Retail and minimarkets. The biggest problem is not the customer who walks off with something: it is unsupervised register voids, miscounted deliveries and expired product. Video analytics genuinely helps when cross-referenced with the point of sale, because it turns a review of hours into one of minutes. Before getting there it is worth having the basics sorted, which we explain in how to reduce shrinkage.
Construction and industry. Checking personal protective equipment at entrances is the cleanest case: the rule is objective — helmet or no helmet — the log is useful for the safety file, and it does not depend on a supervisor standing at the gate. It fits with what we already do in ConstructorIA.
Logistics and warehouses. Checking that what leaves matches what was dispatched, and detecting after-hours access.
This is a category where too much gets promised, so it is worth being explicit about what does not work well.
Identifying people is not reliable in real conditions: changing light, bad angles, masks, caps. A system that "recognises suspects" produces false positives, and a false positive here means accusing an innocent customer. That is why the correct design is to alert on events — a long stay with no matching sale, an after-hours entry — and let a person decide.
The camera's quality sets the ceiling. No artificial intelligence fixes a camera pointed at the ceiling, shooting into backlight or running at two frames per second. On most projects, the first real cost is repositioning or replacing cameras, not the software.
There is a legal limit, and in Panama it is specific. Footage of identifiable people and any biometric data are covered by Law 81 of 2019, which requires consent, a declared purpose, visible signage, retention periods and security measures, with a higher bar because this counts as sensitive data. That gets resolved before installing, not after a complaint.
The name physical AI is given to artificial intelligence systems that perceive and act on the real world — cameras, sensors, robotics — instead of working only with text. It is the category that will be talked about most in the coming years, and in Panama there is still almost no local supply.
What changed is not the idea, which is decades old, but three practical things: vision models now run on cheap equipment installed on site, without sending footage to the cloud; they can be tuned with a handful of examples instead of thousands; and the cost per camera has dropped enough to make sense for a mid-sized business.
Processing happening on site rather than in the cloud is not a technical detail: it is what makes data protection compliance workable and what avoids depending on a stable internet connection for the system to work.
Transparently: this line is in development. There is no catalogue and no price, and we are not going to sell a pretty demo of something not yet in production. We publish this page for two reasons: to talk to companies willing to take part in the first projects, and so that anyone evaluating the subject has an honest explanation of what is possible and what is not.
What is in production today and attacks the same problem is the systems side: point of sale with shifts, cash counts, a supervisor code per register and an inventory audit trail, inventory control with purchasing and warehouses, and AI agents that message you on WhatsApp when something goes out of range. In most cases that recovers more than a camera project would cost, and it is the step we recommend first.
If your operation already has the till under control and cycle counts running, and the numbers still do not add up, then it is worth talking about video analytics.
Frequently asked questions
Let's talk about your operation, with nothing to sign. If what you need today can be solved without cameras, we will say so.