When Your AI Should Stay in the Building: A Plain-English Guide to On-Premise Computer Vision
A client once asked us to add cameras that could recognize faces at their front desk. The interesting part wasn't the AI. It was deciding where the AI should live.
thehsquares Team
Mar 3, 2026 · thehsquares
Picture a security guard who never looks up
Imagine you hire a guard for your front desk, but every time someone walks in, he can't decide anything himself. Instead, he snaps a photo, phones a stranger in another city, describes the person, and waits for that stranger to call back and say "yes, let them in." It works. But it's slow, you pay for every single call, and now a stranger has a photo of everyone who visits your building. That phone-a-stranger setup is basically what happens when your cameras send every frame to an AI service in the cloud. In this article we'll walk through a real kind of project we do, and why we sometimes keep the "brain" inside your building instead. No engineering degree required.
First, what we mean by "computer vision"
Computer vision is just software that sees. You point a camera at the world, and instead of only recording video, the software understands what's in the picture. It can spot a face and recognize whose it is, count how many boxes are left on a shelf, or follow a forklift as it moves across a warehouse. Think of it as giving your existing cameras a pair of eyes and a bit of judgment. The part that does the actual recognizing is called the "model" — that's the trained piece of software that learned, from thousands of examples, what a face or a box or a forklift looks like. Hold on to that word "model," because a common and expensive misunderstanding is thinking the model is the whole job. It isn't, and we'll get to that.
Cloud or on-premise: the choice underneath the project
There are two places the "brain" can live. The first is the cloud — a big AI service run by someone else, out on the internet, that your cameras phone for every decision. It's easy to start with and someone else owns the machines. The second is on-premise, which simply means "on your own servers, inside your own building." It's the guard who works on-site and makes the call himself, right there. Neither is automatically better. For a chatbot that answers a question now and then, the cloud is usually fine. But for cameras watching continuously, all day, we often find on-premise is the smarter call — and it comes down to three plain reasons: speed, cost, and privacy. Let's take them one at a time, using a face-recognition system we built for a private client and a warehouse camera setup as our examples.
Reason one and two: it's faster, and it's cheaper at scale
Speed first. Every time a camera phones the cloud, the request travels out to the internet and the answer travels back. That round trip takes time — engineers call the delay "latency," but you can just think of it as the pause before the answer comes. For a door that should unlock the instant someone is recognized, or a system tracking a moving object, that pause is the difference between smooth and clunky. A brain inside the building answers in a blink because nothing has to leave. Now cost. Cloud AI services usually charge you a little for every image they look at. One camera, occasionally, is cheap. But a warehouse with forty cameras running every second of every day is millions of little charges a month — the meter never stops. Buy the server once and run it in your building, and that meter simply isn't there. Past a surprisingly small number of cameras, owning the hardware is the cheaper path.
Reason three: privacy you can actually promise
This is the one that decides many projects. When faces or sensitive footage leave your building for the cloud, you now have to trust — and be legally responsible for — what happens to that data out there. For a lot of our clients, that's a non-starter. Sometimes it's a company rule; sometimes the law simply won't allow it. Here's the beautiful thing about keeping the brain on-premise: if the footage never leaves the building, a whole category of "what if the data leaks" problems can't happen at all. It's not a promise written in a privacy policy that you're hoping holds up. It's a fact of how the system is built. You can hand your own security team the whole setup and say, "look — nothing goes out." That structural guarantee is often the difference between a project that gets approved and one that dies in a legal review.
The model is only a quarter of the work — plan for the rest
Here's the honest part most people don't expect. That clever "model" that does the recognizing? It's maybe a quarter of the actual work. The other three-quarters is everything around it — what we call the pipeline: catching the video, cleaning up bad lighting and weird angles, deciding how sure is sure enough before it says "that's a match," and connecting it all to the door or the alert or the dashboard your staff actually use. A brilliant model wired up carelessly gives you false alarms nobody trusts. And one more thing to budget for from day one: the world changes. Cameras get moved, lighting shifts with the seasons, you start stocking a new product the system has never seen. So a good system is built to be re-taught — engineers call that "retraining" — so your own team can keep it sharp over time instead of it slowly going stale. The takeaway: when you're weighing an AI-vision project, don't just ask "is the AI good?" Ask where it should live, and who keeps it smart next year. Get those two right and the technology quietly does its job — like a good guard who's simply always there.
Have a project like this in mind?
thehsquares turns ideas like these into production software. Let's talk about what you're building.
Start a Project