Can AI Keep Pace Behind the Wire? Human vs Machine Speed in Producing Defence Training On Premises
Generative AI is reshaping how training content is produced, yet most published evidence comes from cloud-hosted services that Defence training organisations cannot use with sensitive, classified or export-controlled material. If AI is to deliver capability at pace for Defence, it must run where the data already lives: on premises, on sovereign infrastructure and, increasingly, in air-gapped or disconnected environments. This raises practical questions the training community has yet to answer with evidence. Which classes of generative model can now run effectively on local hardware? What do they realistically contribute to the production of training assets? And how does their output compare, in speed and quality, with that of experienced human developers? This paper addresses these questions from the perspective of a team that has deployed generative AI entirely within a UK Prime's security boundaries. It first surveys the open-weight models that can be hosted locally across six modalities. Large language models for instructional text, assessment items and scripts. Diffusion models for illustrations and technical imagery. Code models for interactive exercises. Text-to-speech models for narration. Video generation models for short demonstrations, and image-to-3D and text-to-3D models for equipment representations. It then outlines how these models were combined in a production workflow in which subject matter experts review, correct and approve every output, and sets out the hardware, model governance and update considerations of operating without internet connectivity. The core of the paper is an operational evaluation conducted with a UK Defence prime contractor, using its live training programmes for testing, metrics and benchmarking. Three aspects are examined. The first is capability, which asset types the AI-assisted workflow can now produce to an acceptable standard, as judged by the prime's own subject matter experts against their existing acceptance criteria. The second is time saving, production effort for each asset type is benchmarked against equivalent assets created by subject matter experts alone, with savings reported as percentage reductions. The third is workforce deployment, how roles, skills and team structures were reorganised so that experts spend less time authoring and more time directing, validating and approving content, and which approaches did most to embed the new way of working. The paper also reports where AI output fell short, including factual errors in technical content, inaccuracies in equipment imagery and the review burden this placed on experts. The implications are twofold. First, on-premises generative AI is now mature enough to accelerate Defence training production without compromising data sovereignty, provided human review remains central. Second, the gains are realised only when the workforce is redeployed around the technology rather than the technology being added to existing processes, so organisations should plan role changes, upskilling and review capacity alongside any deployment.