The Pentagon has begun plans to roll Google’s Gemini AI agents out to roughly three million civilian and military users, marking one of the largest enterprise-grade deployments of generative AI inside a defence ministry. Emil Michael, the Department of Defense’s deputy undersecretary for research and engineering, told officials that the agents will start on unclassified networks and could be extended into the department’s classified cloud after further testing and security checks.
Google has already added an "Agent Designer" feature to its Gemini for Government offering on the GenAI.mil platform, allowing non‑technical staff to configure digital assistants that automate repetitive, multi‑step administrative tasks without writing code. The company plans to ship eight off‑the‑shelf agents for tasks such as meeting summaries, budgeting, action‑plan checks and other routine workflows; Pentagon officials say some of these tools may also support operational planning and resource estimation.
The move comes as the U.S. government accelerates the integration of generative AI across federal agencies, even as private companies diverge over the commercial and military uses of large language models. Anthropic’s refusal to lift two safety constraints demanded by the Defense Department — on domestic mass surveillance and fully autonomous weaponisation — has already produced a high‑profile standoff: the department has designated Anthropic a supply‑chain risk and banned it from receiving contracts, and the company has pushed back legally and rhetorically.
Google’s pathway into the Pentagon is not without precedent or controversy. In 2018, Google staff protested the company’s participation in Project Maven, an AI project to analyse drone footage, and the company initially declined to renew that contract amid employee opposition. Google has since relaxed some internal limits on defence work, and Pentagon officials now describe the firm as a trusted partner for delivering AI capabilities at scale.
Officials stress that human oversight remains crucial: agents are intended to save time, not replace judgement. Michael emphasised that department employees will need to train and review agents to prevent "hallucinations" and other errors, and that the Defence Department is still building the basic AI capacities many civilians take for granted.
The Google rollout highlights both practical benefits and systemic risks. Lowering the barrier to build bespoke agents across a sprawling bureaucracy could greatly accelerate administrative efficiency and planning, but it also creates new vectors for data leakage, governance failures and vendor lock‑in. The debate over Anthropic illustrates the political leverage that companies can exert by choosing which constraints to accept, and it exposes discrete trade‑offs between operational agility and security assurances.
For other governments and defence establishments watching closely, the Pentagon’s embrace of commercial AI agents will serve as both a playbook and a warning: scaling generative AI inside a military requires not only technical validation on classified infrastructure but also robust procurement rules, continuous oversight and clearly defined limits on autonomous effects. How the department governs, audits and diversifies its suppliers in the next 12–24 months will shape the operational and ethical contours of military AI for years to come.