Defense AI from Sensor to Shooter, Part 1: The Loop

The phrase "sensor-to-shooter" is shorthand for the loop that begins with a sensor detecting an object of interest and ends with an effector applying force against it. The defining property of modern defense AI is that it compresses this loop — sometimes from minutes to seconds, sometimes from hours to minutes — by automating the cognitive steps a human operator would otherwise perform. This four-part series walks through where AI fits inside the loop, where it does not, and how to engineer the boundaries between the two. Part 1 establishes the loop itself.

This series complements the architectural Complete Guide to AI and Edge AI in Defense Software and the broader C2 Systems pillar. Those guides survey the field; this series goes operational.

What the Sensor-to-Shooter Loop Actually Is

The loop has stages. Different services and doctrines name them slightly differently; the canonical stages, in order, are find, fix, track, target, engage, assess — sometimes shortened to F2T2EA. The OODA-loop framing (Observe, Orient, Decide, Act) maps onto these stages, with Observe-Orient covering find/fix/track, Decide covering target, and Act covering engage. Assess is the feedback that closes the loop.

Each stage answers a specific question:

• Find — did a sensor pick something up?
• Fix — where exactly is it, with how much certainty?
• Track — how is it moving, and what is the next-position prediction?
• Target — is it the right object, of high enough priority, in a permissible engagement context?
• Engage — which effector applies, with what timing, with what authorization?
• Assess — what was the outcome, and what do we learn for the next cycle?

Pre-AI, every stage was human-mediated, with software acting as display and message-routing infrastructure. AI changes this. The question is no longer "can AI participate in the loop" — it can — but "at which stages, with what guardrails, and crossing which thresholds requires explicit human confirmation."

Where AI Compresses the Loop

The operationally honest mapping of AI capabilities to loop stages in 2026:

Find — heavy AI deployment, low controversy. Computer vision on UAV full-motion video. Acoustic detection. Radar plot classification. SIGINT signal-of-interest extraction. The compression here is dramatic: a 12-hour UAV flight that produced 90 seconds of operationally relevant footage now arrives at the analyst pre-marked, with the 90 seconds ranked. The engineering pattern is treated in AI for ISR Data Triage and the computer-vision specifics in Computer Vision in Defense Systems.

Fix — moderate AI deployment, well-bounded. Track localization from multi-sensor inputs, geo-rectification of FMV detection boxes, identity-resolution across overlapping sensors. The compression is real but bounded by the fundamentals of the underlying sensor physics — AI cannot improve a radar's measurement accuracy, only its association with other reports.

Track — moderate AI deployment, hybrid with classical methods. Track-to-track correlation has classical statistical methods (JPDA, MHT) that remain the operational baseline; ML-native trackers augment them rather than replace them. The hybrid pattern — ML proposes associations, statistical engine validates — is the operational compromise. See Military Data Fusion Explained for the underlying discipline.

Target — light AI deployment, heavy human-in-the-loop. Target prioritization aids exist. Recommended-engagement lists exist. Autonomous targeting decisions are rare and tightly bounded by doctrine. The engineering implication: AI here produces ranked candidate lists with explanations; humans confirm. The structural boundary cannot be implemented as policy alone — it must be coded into the platform.

Engage — minimal autonomous AI, mature decision-support. Effectors fire when humans authorize them. AI assists with engagement-zone deconfliction, friendly-force avoidance, and effector selection, but the trigger sits with a human. This boundary is enforced not by AI ethics review (though there is plenty of that) but by accreditation, doctrine, and procurement requirements that look identical across NATO-aligned forces.

Assess — heavy AI deployment. Damage assessment, second-look UAV tasking, automated battle damage reporting. The compression closes the loop and feeds back into the next cycle's find stage.

Latency Budgets: What "Machine Speed" Actually Means

"Machine speed" is the marketing phrase. The engineering reality is per-stage latency budgets that compose into total loop time.

Typical budgets for a tactical air-defense scenario, brigade level:

• Find — sensor to detection: ~100 ms for an edge-AI detector on radar/FMV; ~1 s for centralized batch processing.
• Fix and track — fusion engine combines: 500 ms 95th percentile, 1.5 s 99th percentile.
• Target — human review of candidate, decision: 5-30 seconds for routine, longer for high-stakes.
• Engage — order propagation to effector: 1-5 seconds depending on radio.
• Assess — initial bomb-damage reporting: 30 s to 5 min depending on sensor revisit.

Slow links and human review dominate the total. AI compression makes a meaningful difference at the AI-eligible stages, but the loop is gated by the slowest stage. A platform that compresses find from 5 minutes to 100 ms is impressive; if the target stage still takes 30 seconds of human review, the total loop time is dominated by human cognition, not AI inference.

The implication is procurement-relevant. "Sensor-to-shooter loop compression" claims should be evaluated on the slowest stage, not the marketed stage. For most defense applications, the slowest stage is intentional — human review at target — and compressing it further requires changing the doctrine, not the software.

Domain Considerations: The Loop Differs by Domain

The same architectural pattern instantiates differently across operational domains.

Land. Slower loop, more human review, complex urban environments confounding sensors. AI compression is concentrated at the find stage (UAV imagery, acoustic detection) and the assess stage (damage reporting).

Air defense. Faster loop, tighter latency budgets, well-understood sensor models. AI participation is broader — track classification, threat-priority ranking — but still gated by human engagement authorization. Link 16 integration carries the tactical data; see Link 16 Tactical Data Links.

Maritime. Slow loop, large search areas, AIS-spoofing concerns. AI compression at the find stage (AIS/ADS-B integration, see AIS and ADS-B Integration) and pattern-of-life detection (Pattern-of-Life Analysis).

Cyber. Adjacent rather than identical loop — detection, attribution, response. The find stage is AI-heavy (anomaly detection on network telemetry); the target/engage stages are doctrine- and law-heavy. See Cyber Situational Awareness Platforms.

Space. Emerging, slow loop, sparse sensor coverage. AI participation early at sensor-fusion stages; the engage stage remains nearly entirely human.

Multi-domain operations require harmonized loops across these domains. The JADC2 architecture pattern is the U.S. expression of this; European equivalents follow similar contours. See European JADC2 Vendors.

Where the Human-in-the-Loop Boundary Lives

The structural boundary between AI and human action is not negotiable. It is enshrined in NATO AI strategy principles, in national doctrine, in international humanitarian law as applied to autonomous systems. The engineering implication: the boundary is implemented as code, not as policy.

The patterns that work:

Per-decision authorization. Every action with operational consequence — an engagement order, a track-classification change, a tasking instruction — requires explicit human confirmation. The platform refuses to proceed without it. The confirmation is logged with operator identity for after-action review.

Graduated autonomy. Within a single workflow, different decisions have different authorization gates. A track-creation event is autonomous; a track-engagement decision is human-confirmed; a track-cancellation revisits authorization. The boundary is decision-specific, not workflow-specific.

Failure-mode escalation. When the AI's confidence in a recommendation falls below threshold, the workflow escalates to a higher-authority human rather than proceeding with reduced confidence. This pattern matters most when the AI is operating in degraded conditions — sensor jamming, adversarial inputs, model drift.

The broader treatment of doctrine and ethics for defense AI is in NATO's AI Strategy for Defense Software. The LLM-specific guardrails for analyst-facing workflows are in LLMs in Intelligence Triage for Defense.

What This Series Covers

The remaining three parts walk through the loop in engineering depth.

Part 2: Sensor-Side AI goes into the find and fix stages. Edge-inference architecture, model deployment to UAV payloads and ground vehicles, the hardware choices, the data pipeline that supports the models. This is the AI-heavy end of the loop and where most engineering investment lands.

Part 3: Decision Support and Course-of-Action AI covers the target stage and the analyst-facing tooling around it. Recommended-engagement lists, course-of-action analysis aids, LLM-augmented briefing tools, and the structural patterns that keep humans in the loop.

Part 4: Effects, Effector Integration, and Human-in-the-Loop Boundaries closes the loop. How AI participates in engage and assess without crossing the autonomous-effects line. The accreditation, doctrine, and procurement realities that pin the boundary in place.

Each part assumes the loop framing from Part 1 and goes operational.