Modern tactical operations rarely fail due to a lack of data, but they often struggle under the weight of too much of it. ISR analysts are routinely inundated with simultaneous feeds including radar tracks, RF signals, electro-optical/infrared (EO/IR) streams, and a continuous cascade of AI detections. At the other extreme, an FPV pilot may have one poor analogue feed to fly one platform toward a high-value target.
In both cases, the problem is not simply whether the system can detect something. It is whether it can surface the information that matters, at the moment it matters, without creating another task for the operator.
Adding Automatic Target Recognition (ATR) directly to these systems does not automatically reduce cognitive burden or enhance mission effectiveness. In many cases, more detections simply mean more information for a human to process, what we are all too familiar with calling “AI slop.”
The real engineering challenge at the tactical edge is defining the systemic division of labor between algorithm and operator. The next measure of defense AI maturity is not simply how accurately an algorithm can recognize objects, but how effectively the broader system turns those detections into information an operator can use.
To build true operational capability, defense architectures must treat human attention as a limited system resource, designing processing, filtering, and escalation logic so operators enter the workflow when human context, authority, or judgment are actually required.
Autonomous Edge Processing vs. Human Decision-Making
To offload cognitive strain effectively, systems must separate routine machine-level processing from high-consequence human decisions. Traditionally, “human-in-the-loop” has sometimes been implemented as a blanket requirement where an operator manually validates every step. While preserving human touchpoints technically satisfies control requirements, it can leave the operator trapped in continuous processing.
That becomes increasingly untenable as militaries move from operating individual unmanned systems to supervising fleets of them. A model that requires a human to continually monitor each sensor or platform does not scale simply because the platforms themselves have become more autonomous.
Algorithms excel at persistent, high-frequency surveillance tasks. Onboard edge compute can continuously fuse sensor data, maintain visual custody, apply mission-defined geometric boundaries, compute confidence metrics, reassociate lost tracks, and log behavioral trajectories without requiring constant operator intervention.
Human cognitive bandwidth can then be reserved for tasks that software cannot resolve: interpreting intent, evaluating complex tactical environments, understanding novel targets, navigating sensor ambiguity, and assessing operational risk.
Edge processing should act as an intelligent filter, maintaining the routine state of the battlespace while escalating the exceptions that warrant human attention.
Consider the payload specialist historically required on larger UAS. Maintaining a gimbal lock, following a target and continuously analyzing imagery are mentally exhausting but increasingly automatable tasks. If AI can assume that routine workload, the value is not simply that a platform has become “more autonomous.” The value is that the human can devote attention to the decisions for which human involvement actually matters.
Human Control Matters Most Where Consequences Are Highest
The division between machine processing and human decision-making becomes most consequential when workflows touch kinetic engagements. ATR models can rapidly detect, classify, and track potential targets, but an algorithmic identification is not in itself an authorization to strike.
“Human-in-the-loop” should not mean placing a human at every stage of machine processing. It should mean preserving meaningful human authority where judgment and accountability are required.
AI can assist with distinguishing objects, presenting target candidate packages and highlighting changes in track behavior. But determining operational intent, interpreting ambiguous context and making decisions about the use of force are fundamentally different tasks from recognizing an object in an image.
System architectures therefore need clear boundaries between machine-generated understanding and human authorization. Otherwise, operators can become overwhelmed by routine machine outputs while simultaneously becoming too dependent on automation when consequential decisions arise.
From Continuous Monitoring to Management by Exception
Preventing operator overload also depends on the user interface and underlying escalation logic. If threshold rules are overly sensitive, the HMI becomes a source of false-alarm fatigue. If rules are too permissive, critical tactical threats may slip past unnoticed.
The goal should be to shift the operator from continuous monitoring to management by exception and escalation.
Confidence should become a functional part of that workflow, not merely an algorithm score buried somewhere in the system. Rather than displaying a raw bounding box labelled “Vehicle: 87%,” the interface should explain why an operator is being brought into the workflow: “Track crossed geo-fence; confidence dropped 30% due to EO/IR clutter.”
That enables the operator to begin at the point of decision rather than reconstructing historical data.
This becomes even more important as forces deploy larger numbers of low-cost, attritable autonomous platforms. The question is no longer simply whether one operator can control one drone. It is how many platforms, sensors and detections one operator can effectively supervise.
Systems therefore need to cluster and prioritize information at the level at which decisions are being made. Ten individual vehicle detections may be useful to a tactical operator as ten tracks, but to a higher command echelon they may need to be presented as a formation.
More AI detections are not necessarily more intelligence.
The Bandwidth Problem Is Also an Information Problem
Tactical environments are frequently contested, congested, and bandwidth constrained. Transmitting continuous, high-definition full-motion video from dozens of distributed edge sensors to centralized command nodes is both technically difficult and operationally counterproductive.
The same principle that should govern human attention should govern network traffic: move what matters, not everything that can be collected.
Raw multi-sensor feeds and high-rate full-motion video can be processed locally at the tactical edge. Higher command may instead need filtered telemetry, metadata tracks, target coordinates, thumbnails, alerts and full-motion video on demand. Tracking and local threat triage can happen onboard while higher echelons receive the information required for broader situational understanding.
Edge AI therefore does more than accelerate inference. It helps determine what information is important enough to consume scarce bandwidth and scarce human attention.
Stop Measuring Defense AI Only by the Model
Developing viable defense AI requires changing how engineers evaluate system performance. Traditional metrics such as Mean Average Precision (mAP) or inference latency tell us how accurately or quickly an algorithm performs. They do not tell us whether the overall human-machine system functions effectively under operational stress.
A model can achieve excellent benchmark performance and still contribute to a poor operational system if it constantly interrupts the operator, produces redundant alerts or requires humans to reconstruct context before acting.
System-level evaluations should therefore measure operator intervention rates, time required to comprehend an escalation alert, false-alarm frequency, and performance under degraded conditions.
They should also ask a more basic question: did the AI reduce the number of things a human had to think about while preserving the things a human actually needed to decide?
The goal of tactical AI is not maximum automation for its own sake, but appropriate automation: assigning repetitive, high-frequency surveillance tasks to edge processing while reserving human cognition for authority, context, and accountability.
The industry has become very good at asking whether an AI model can detect the target. The harder and increasingly more important question is what happens after it does.
Reducing tactical overload isn’t achieved by removing the human from the loop. It is achieved by engineering a smarter, context-aware loop where machine and operator perform the tasks they are best built to execute. That is the benchmark defense AI now needs to meet.
Stephen Bornstein is the Chief Product Officer at Sightline Intelligence.
