AI in border control

In June 2026 we presented alongside C4i Communication and PureTech Systems in a World Border Security Congress webinar: AI-Enhanced Border Surveillance — From Operational Challenges to Intelligent Solutions. It followed the Vienna edition of the Congress in May, where we showed OpCopilot running on both of its deployment tiers.

This is the argument we made there, written down.

Detection is mature, interpretation is not

Border programmes have invested heavily in sensors over the last ten years, and that investment has paid off. Radar, AIS, thermal and unattended ground sensors now produce reliable tracks in conditions that used to defeat them. The multi-sensor detection and tracking that PureTech Systems builds into PureActiv is mature technology. Detection is not where border operations are failing.

What has not scaled is the person in front of it.

A watch officer may be holding twelve feeds and three C2 screens while a radio net runs in the background and the procedure binder sits open on the desk. The track is there. Working out what it means is unaided human work under time pressure, and it has to be done against last week's pattern, against the ConOps, and against whatever the neighbouring sector saw two hours ago. So the situation report gets written by hand at the end of a long shift. The correlation that spanned three sectors never gets made. The knowledge leaves the room at handover.

The gap sits between a signal existing and a decision being made, and a fourteenth sensor does not close it.

What we build

OpCopilot is an intelligence layer. It sits over the sensors, systems, and teams already running the border, and it is sovereign by construction: models, doctrine, feeds, and operational memory all stay on the customer's infrastructure, inside the customer's accreditation boundary, air-gapped where policy requires it. Nothing depends on an external network, and no data leaves the country to be reasoned about.

It deploys in two positions, and the reasoning is identical in both. Edge runs at the surveillance station, next to the sensors. Server runs in the command centre, across every station at once.

Everything below is what that buys an operator.

Sense, reason, act with the operator

OpCopilot is a team of agents. It subscribes to live feeds, applies the organisation's own doctrine, correlates events across sources and hands structured insight to the operator. Operators generally expect either a chatbot or a dashboard, and it is neither. Three things happen, in order:

It replaces nothing. The C2 stays, the VTS stays, the sensors stay. The layer goes over them.

Edge at the station, Server in the command centre

Border agencies are not short of infrastructure. They run national command centres, regional operations rooms and accredited data centres. They also run forward surveillance stations, towers, checkpoints and patrol vehicles. Both tiers are staffed, both make decisions, and they are not asking the same question.

At the surveillance station: what is in my sector right now, is this track behaving normally, and what does the ConOps say I do about it. The link back to headquarters is often intermittent, contested, or absent by design.

At the command centre: what is happening across every sector, does it match the pattern from another sector three weeks ago, which agency owns the response, and what did the previous watch already decide.

OpCopilot Edge is a compact AI appliance that sits at the station beside the hardware. It carries the complete local model stack on a GB10 Grace Blackwell-class appliance with 128 GB of unified memory, and it can be ruggedised for on-scene, vehicle-borne and single-station deployment. It does not call back to a server. The station keeps reasoning at full capability when the link to the command centre degrades or was never there, at zero network latency, air-gapped where required. It is also how most teams start: one station, one watch, one workflow proven before anything scales.

OpCopilot Server is the command-centre tier. It collects from every connected station into one operational picture, correlates across sectors, and serves 50+ concurrent users across multiple watches and agencies. It is provisioned as infrastructure-as-code onto bare-metal Linux, a customer-controlled cloud tenancy, or a private-cloud equivalent, on data-centre accelerators including NVIDIA H20, H100, and H200. The customer selects and owns the substrate.

The two are complements. With Edge forward and Server at the centre, the station stays capable when the network does not, the centre sees what no single station can, and nothing crosses a boundary it is not allowed to cross.

Collection and reporting across every station

This is the part operators feel first.

Sector reporting today is a manual aggregation problem. Each station logs what it saw, in its own format, at its own tempo; the command centre chases the stragglers, reconciles the overlaps, and assembles a picture that is already hours old by the time it is briefed.

OpCopilot Server collects continuously instead. Events, sources, findings, and decisions from every station accumulate against the same operational picture, and reports are drafted from that live record in the formats and priorities the operation already uses — per station, per sector, or across the whole frontier. Rooms preserve the questions, findings, and decisions across watches, in more than twenty languages, so the shift changes and the operational memory does not.

5 minutes to review, instead of 2 hours to build by hand.

Speed is not what makes the output usable. What makes it usable is what the draft admits. In our demonstration the generated SITREP records its confidence, states CPA/TCPA uncertainty explicitly, lists the environmental and asset data it does not have, flags the urgent close-quarters risk, and separates recommended checks from decisions requiring operator authorisation. The assistant states its limits up front: it will not make command decisions, interpret rules of engagement, or task assets autonomously.

The gaps are part of the output.

An AI system that hides its gaps is worse than none, because it spends operator trust it has not earned.

Configured for real theatres

What we showed at the Congress was configured for three specific operating environments.

The same platform carries the adjacent missions: trafficking and poaching corridors, irregular-migration monitoring with automated reporting on group movement and resource allocation, and counter-terrorism work, which covers behavioural indicators, target tracking across checkpoint data and cross-border intelligence sharing inside an air-gapped boundary.

The same requirement, well beyond the border

Border control is a demanding proving ground. It has distributed sites, poor networks, multiple agencies and data that is not permitted to leave the country. But nothing in the architecture is specific to a frontier.

The requirement OpCopilot is built around is narrower and more portable than border security. It needs an operation running on live data, a body of doctrine governing how that data is acted on, and a hard boundary the data cannot cross. Ports and vessel traffic services, energy and utility control rooms, national emergency and search-and-rescue coordination centres, defence C2, and any critical-infrastructure operator whose regulator has an opinion about where inference happens. Same shape. Most of them cannot use a hosted model at all.

The answer is the same one we gave in the webinar. Run the models on your own infrastructure, inside your own accreditation boundary, air-gapped where policy requires. Encode your procedures instead of adopting someone else's. Keep the operator in the decision. Sovereignty here is the difference between a system that can be deployed and one that cannot.

How this gets proven

We work forward-deployed. An engineer embeds with the operational team from day one, on the customer's infrastructure, backed by the CTO, because encoding a ConOps is not a document exercise and cannot be done at a distance.

Engagements are staged, each stage funded and scoped on its own: proof of concept, proof of value, then deployment and adoption once results are demonstrated. It is the low-risk path government customers prefer, and it is the honest one.

Watch the webinar. AI-Enhanced Border Surveillance: From Operational Challenges to Intelligent Solutions, with C4i Communication, PureTech Systems, and Exolvo — hosted by the World Border Security Congress. Watch the recording →