On December 9, the Department of Defense launched GenAI.mil, a platform that delivers commercial generative AI tools to nearly its entire workforce. The first tools on it are Google Cloud’s Gemini for Government products. Pentagon CTO Emil Michael, unveiling the platform, said “3 million employees, warfighters, and contractors are going to have AI on their desktop, every single one.”
Defense Secretary Pete Hegseth was blunter. “The future of American warfare is here, and it’s spelled AI,” he said, as quoted by Breaking Defense.
/ 01What was launched
According to DefenseScoop, citing Defense Department and Google releases, all tools on GenAI.mil are certified for Controlled Unclassified Information and Impact Level 5 data, the category Breaking Defense describes as sensitive but unclassified. Google described its tools as offering an edge through “natural language conversation, retrieval-augmented generation (RAG)”.
Gemini is the first model, not the only one. The Chief Digital and AI Office has contracts with four frontier AI companies, Anthropic, xAI, OpenAI and Google, and Michael confirmed other industry models will be added later. Breaking Defense reported that Michael also wants generative AI for classified data. He identified three near-term priority areas: organisational, intelligence and warfighting activities.
The launch carried a pointed message about the past. “For the past five years, the Department has had very little to show in the way of AI,” Michael said, according to Breaking Defense.
/ 02Why it matters
Putting a capable assistant in front of three million people is a different kind of AI programme from the pilots that preceded it. Most of the use will be ordinary: drafting, summarising, searching policy. That is where the time savings are, and it is reasonable to start there.
Scale also changes the risk profile. When a few analysts use an experimental tool, errors are caught by people who know they are using an experiment. When the tool sits on every desktop, its outputs flow into memos, briefings and decisions written by people who will not always check. The question shifts from “can the model do this” to “how does a reader know which parts of this output to trust.”
/ 03The traceability angle
Retrieval-augmented generation helps. It ties answers to documents rather than to a model’s memory. But RAG retrieves passages by similarity, and a model still composes the answer. For a summary of a travel regulation, that is usually fine. For intelligence and warfighting work, two of the three priority areas Michael named, analysts need more: which sources were used, how they were combined, and whether two reports that look similar actually describe the same unit, place or event.
Those are entity and relationship questions, and they are what knowledge graphs answer. Fusing sources into a graph, with each statement tied to the document that produced it, lets an analyst follow an answer back through its evidence and lets a reviewer reproduce it. As the platform moves toward agents and, eventually, classified data, that kind of lineage will matter more, not less.
A multi-vendor platform adds another dimension. With several frontier models available side by side, users will sometimes get different answers to the same question. That is healthy if they can see which sources each answer used and resolve the disagreement on the evidence. It is confusing, and potentially dangerous, if the only way to choose between answers is which one sounds more confident. A shared, inspectable layer of facts underneath the models is what turns model diversity into a strength.
The same layer protects the department against lock-in. If the knowledge an organisation relies on lives in a governed graph rather than inside any one vendor’s context window or fine-tuned weights, models can be swapped as better ones arrive without rebuilding what the organisation knows.
Access to AI is the first step. Knowing where each answer came from is what lets people act on it.
/ 04What to watch
- Which models follow Gemini, and whether users can compare their answers on the same sources.
- The move to agentic tools and to higher classification levels, which will raise the bar for audit trails.
- Training and guidance for a workforce that is being asked to adopt AI quickly.
Three million users will expose the weak point of retrieval quickly: it returns passages that look relevant, and nobody downstream can see whether two of them describe the same convoy or two different ones.
For intelligence work, the step that matters sits before generation. Fuse reports into a graph in which each observed entity, whether a vessel, unit or facility, is a node, each sighting links to the report and collection source behind it, and a match between two sightings records its basis. An analyst can then ask which reports place this vessel in a given port over the last 48 hours, and how many of them are independent, and get an answer that names documents instead of paraphrasing them. In the regional-conflict build on our site, six sources, including sanctions notices and a video transcript, became 720 records, each traceable to its document. With that layer underneath, answers from two different models can be compared on the evidence they cite.
How we’d model it
Sighting—observes→VesselSighting—reportedIn→Intelligence reportIntelligence report—collectedBy→Collection sourceModel answer—cites→Intelligence report
- Pentagon rolls out GenAI platform to all personnel, using Google's Gemini — Breaking Defense breakingdefense.com ↗
- DOD initiates large-scale rollout of commercial AI models and emerging agentic tools — DefenseScoop defensescoop.com ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.