It does not just answer: it does the work, and nothing leaves your site.

A mid-sized installation produces more information than anyone can watch: hundreds of cameras, thousands of detections a day, cases piling up and recordings nobody has time to review. The operator knows exactly what they want to ask. What they do not have is half an hour to walk through six different screens and remember where each thing is kept. So plenty of reasonable questions never get asked, and information that cost money to install sits there unused. The assistant removes that walk: you ask once, and it works out for itself what it needs to look at. If the answer takes three things, it does the three at the same time.

An ordinary chat answers with what it already knows. This one has 95 tools plugged into the real system, and it uses them. It finds cameras and tells you how many each site has, looks through one live, opens yesterday's three o'clock recording, compares this month's cases with last month's, works out the route between two points, draws the map and puts the report together as a PDF. And when the person asking has the right, it acts: it creates the case, assigns it, closes it, aims a moving camera, sends a view to the video wall or pushes the alert out by email. Putting an automation live from the chat always demands an explicit confirmation, because answering is one thing and touching the installation is another. And the catalogue can grow without getting in the way, because the assistant looks up the tool it needs instead of carrying all of them around.

The useful part is not asking the system: it is asking the picture. “What is on the platform right now?”, “which cameras have smoke?”, “is this the same van as in the other photo?”. It takes the frame from that moment, looks at it and answers. The same question can go to a whole group of cameras at once, returning only the ones that match, which is the only way a question is any use when there are three hundred. And you decide how rigorous you want it: there are six levels, from answering straight out to separating what is seen from what is inferred, listing what cannot be known from the evidence in front of it, or arguing against its own conclusion and stating how sure it is. Moving up a level does not make every answer longer: it just stops cutting off the ones that needed the room. There is one boundary worth being clear about, though. Asking is how you search. Counting is how you measure. And they are not done the same way: a line cannot be described in a sentence, it has to be drawn on the image.

Here is the argument that settles a lot of purchases, and it deserves saying without hedging: everything runs inside the customer's installation. The intelligence behind the assistant is our own and runs on the customer's server; the mapping is ours too and works offline; speech is transcribed and read aloud in the same place. Not one image, not one piece of data and not one query from the conversation travels to a third party, and it all keeps working with the internet cable unplugged. For a hospital, a prison or a critical infrastructure operator that tends to be the whole conversation, and it is the reason this product gets in where others do not even turn up. Everything also travels over a single secure port, 443, so there is nothing else to ask the network department to open. And the assistant is not a back door: it only sees the tools the person asking is entitled to use, so the same chat gives an operator and an administrator genuinely different powers. Every tool call is logged with who asked for it, what they asked, what came back and how long it took; conversations are private by default, and when an administrator is reading someone else's, the screen says so.

In a room that runs on shifts, yesterday's conversation matters as much as today's. Any thread can be renamed, starred, filed into a folder, pinned to the top or archived without being deleted, and the search looks inside the messages, not just at the titles. From any point in a thread you can open a branch: a new conversation that inherits everything up to there, so you can try another line without spoiling the original. The whole conversation can be exported, or turned into a report with its text, its images and its times, which is what the incident write-up ends up needing. And it can be shared: one thread with colleagues from the same organisation, up to fifty of them, with every invitation, every moderation and every departure recorded one by one, and the messages of anyone who leaves are kept. The limit is worth stating plainly: this is asynchronous collaboration, messages appear when the page is refreshed. It is not a live chat room and it is not sold as one.

An assistant that invents the answer when it does not have one is worse than no assistant at all, so the work here went the other way. If there is no recording at the moment you asked for, it says so instead of showing you something else. If it runs out of room to think or to write, it says so and suggests moving up a level, rather than handing over a truncated answer without warning. At the higher levels it lists what cannot be determined from the evidence in front of it and what would be needed to know, and when it cannot get there it answers that it cannot be determined. The reasoning is shown separately, folded away and marked as internal process, so nobody mistakes it for the conclusion. And errors are explained with their actual cause — a missing permission, an expired session — instead of failing in silence. Every answer also arrives with the tools it used and what they returned: an answer like that can be checked, not just believed.