This paper investigates offline multimodal large language models as decision‑support tools for air operations. In air‑operation settings connectivity is limited and security constraints are strict, so analysts must fuse textual doctrine with imagery under tight time pressure. We describe a modular retrieval‑augmented architecture that runs without Internet access, accepts both text and image inputs from technical manuals, and lets users trace knowledge back to its original source via natural‑language interaction. To evaluate the design we conducted a pilot study with four image analysts from the Brazilian Air Force. The study comprised two parts: (i) a doctrinal knowledge test based on the electronic target‑identification doctrine, comparing human and system performance on the same items; (ii) a measurement of the cognitive workload required to produce a reconnaissance target report (REMIR) manually, without AI assistance. Findings show the manual task imposes high mental demand (6.0/7) and effort (5.0/7), taking on average 26.5 minutes, whereas the proposed system achieved an 8/10 score and completed the assessment in 7.1 minutes, matching human accuracy while dramatically reducing time. This establishes a baseline for future AI‑assisted evaluation. Finally, we outline a protocol for systematically comparing manual and AI‑augmented workflows.
Review