AI PM in Defense Tech: Breaking Into Anduril, Palantir, Shield AI, and the New Wave of National Security Companies

User research in restricted environments

Your users are military operators, intelligence analysts, or logistics coordinators with classified daily work. Research requires security clearances, site visits, and significant relationship investment with program offices. You cannot run a standard user interview over Zoom. The discovery process is slower and more relationship-dependent than commercial tech.

Spec writing under operational constraints

AI features in defense need to account for degraded environments — no GPS, no internet, low bandwidth, electromagnetic interference, and adversarial attempts to confuse your AI. Your specs need failure mode documentation that commercial PMs rarely think about: what does the system do when sensor data is corrupted? When comms drop for 20 minutes?

Evaluation and testing requirements

Defense AI systems go through formal government evaluation and acceptance testing before deployment. The PM owns the acceptance criteria that the government evaluators test against. A missed acceptance criterion costs months. This makes specification precision a higher-stakes skill than in commercial products.

Cross-program office coordination

Your customer is a program office — a government bureaucracy with its own requirements, budget cycles, and political constraints. The PM in defense tech is often the translator between engineering capability and program office requirements, which is a more structured and slower process than commercial GTM.

Compliance and export control

Defense tech products are subject to ITAR (International Traffic in Arms Regulations) and EAR (Export Administration Regulations). The PM owns the product-level compliance decisions: which features can be shown to which countries, which data can be shared with which partners, and which capabilities require additional government approval.

High mission stakes, long feedback loops

The feedback loop in defense tech is longer than in commercial tech. You may not see your product deployed in the field for 12 to 24 months after development. When it is deployed, the stakes are higher than in commercial products — AI decision errors have real consequences. This attracts PMs who want significance over speed.

The case for defense tech AI PM work

The argument made by most people who work at these companies: AI-enabled defense systems reduce casualties on both sides by reducing collateral damage, enabling more precise strikes, and shortening conflicts. The alternative is not no autonomous weapons — it is autonomous weapons built by adversaries with fewer safety constraints. Building these systems with safety and human oversight baked in from the product spec level may be better than leaving the space to others.

The case against

The counterargument: AI in weapons systems lowers the threshold for military action by reducing human cost, normalizes autonomous lethal decision-making, and creates capabilities that will eventually proliferate beyond controlled use. A PM working on targeting AI is making ethical tradeoffs that extend beyond any individual product decision.

What you can verify before deciding

Before joining, ask directly: Does the company have an ethics review process for new product capabilities? Who has final authority on human-in-the-loop requirements? What products or capabilities has the company declined to build? Strong answers to these questions indicate institutional seriousness about the ethical dimension.

Dual-use work as a middle path

Scale AI, Palantir AIP commercial, and similar roles involve defense customers but focus on data infrastructure, analysis tools, and decision support rather than weapons systems. These roles are meaningfully different ethically from autonomous weapons development and are a reasonable middle ground for PMs who want defense tech exposure without the weapons system dimension.