Improved but incomplete: GenAI.mil and the last-mile problem

When I first tested GenAI.mil a few months ago, the main achievement was establishing an approved and secure Generative AI platform for the entire DoW to use. That was a massive first step. The early version provided a good starting place but lacked some of the features it needed to fully realize its potential for the Department of War.

Since then, the platform has evolved. The development team has clearly listened to user feedback, rolling out new features, including model upgrades to Gemini 3.1 Pro and 3 Flash, API access, the introduction of Nano Banana for image generation, and custom agent creation. It is encouraging to see the DoW pushing updates to try and keep pace with the commercial sector. However, as these new capabilities hit the operational environment, a familiar logistical challenge has emerged: the last-mile delivery problem.

In enterprise technology, the hardest part of any deployment is often the final stretch — getting the tool seamlessly integrated into the daily workflows of its users. GenAI.mil successfully does the heavy lifting of processing data securely, but its execution at the user level often requires clunky workarounds that slow its users down.

The promise and limits of automation

One of the most promising additions to GenAI.mil is the new Agent creator. Functioning largely as a customizable chatbot builder, it allows users to specify instructions, upload reference documents, and prescribe prompts. Users can then start a chat session with an “agent” that will deliver much more tailored and specific responses. When used for specific, scheduled tasks, it shows promise. For example, I created an agent to automatically generate a brief on personnel recovery news every morning.

However, current limitations prevent it from acting as a true “agent.” While it can generate a brief, it cannot deliver it — I cannot set it to automatically email the product to my team or even to myself. Furthermore, it can currently only draw on public sources or uploaded files; it lacks the ability to reach into internal environments like SharePoint or email to perform complex tasks.

Additionally, sharing these created agents is not easy. Users must be added individually, and there is no simple, shareable link; colleagues must navigate into the agent menu and know exactly where to look. Finally, if an operational ask is too large, the system quickly hits its context window, requiring the user to prompt the agent to continue. It is a step in the right direction, but it is not yet the autonomous staff officer many were hoping for.

The reality of collaboration and memory

While collaborative tools are essential for modern staff work, GenAI.mil largely remains a solitary experience. A newly introduced “chat sharing” feature allows users to share a snapshot of their prompts and history from a particular chat. While helpful for passing along a useful query, it is not a true collaborative workspace. New outputs remain visible only to the individual typing them, preventing a staff from seamlessly teaming up on a complex problem.

Compounding this isolation is a technical limitation: the platform lacks persistent memory across sessions. Each new chat is treated as a separate environment, and the AI cannot reference a user’s previous conversations. While this siloing may be done for security purposes, it prevents the tool from learning a user’s specific workflow, formatting preferences, or operational context over time.

The last-mile challenge: Documents, data generation and APIs

The lifeblood of any enterprise tool is its ability to ingest data and output usable products. Here, GenAI.mil struggles significantly with last-mile delivery.

Outputting files presents a primary bottleneck. The platform can produce documents through a Python-based generation tool, but it fails regularly. While it handles basic text requests well, the system frequently stumbles over requests to format tables or edit specific font features.

Furthermore, while the platform analyzes data quickly and meaningfully, generating that data into a usable file is a challenge. For example, if a user uploads a PDF of survey results and asks the AI to organize them into an Excel document, the environment currently cannot generate downloadable spreadsheet files such as .xlsx or .csv. Instead of a seamless file generation function, users are often forced to copy raw text and manually build their own spreadsheets. This creates unnecessary friction that slows down daily workflows.

The recent release of an API holds massive promise for integrating GenAI.mil into existing developer tools. Unfortunately, the rollout has been met with frustration by some due to frequent connection failures and errors when trying to link with various developer applications.

Image generation and system responsiveness

Finally, there is the issue of basic system responsiveness and generating visual products. While speed fluctuates — likely depending on network traffic — GenAI.mil still is not as fast as the almost instant responses users experience from commercial home environments.

On the visual front, the DoW deserves credit for recently releasing Nano Banana into the platform for image generation. In a military environment where visualizing intelligence, mapping terrain, and building slide decks are paramount, this is a welcome addition. However, like many of the newer features, it is experiencing growing pains. The tool currently fails to generate images a noticeable portion of the time. It is a step in the right direction, but it highlights the gap between deploying a feature and perfecting its reliability.

Delivering the last mile

GenAI.mil has proven it can deploy a secure AI platform and continuously iterate to add new features, which is commendable. The primary challenge moving forward is making it faster and more reliable.

To evolve into a true force multiplier, developers must focus on reducing user friction in addition to releasing new capabilities. An API product that connects reliably to developer tools, improved file generation with lower failure rates, the ability to directly generate downloadable spreadsheets, and smoother collaborative features will accelerate the DoW’s efficiency significantly. The heavy lifting of establishing a secure platform is complete; now it is time to improve the last mile delivery.

The views expressed are those of the author and do not reflect the official policy or position of the Joint Personnel Recovery Agency, the Joint Chiefs of Staff, the Department of the Air Force, the Department of War, or the U.S. Government.