When a user asks, “What enemy military units are in the area?” the AIP aide speculates, “Based on equipment patterns, it’s likely an armored assault battalion.” This prompted analysts to call in an MQ-9 Reaper drone to survey the scene. You then ask your AIP assistant to “generate three courses of action to target this enemy equipment,” and the assistant immediately suggests attacking your troops with either “air power,” “long-range artillery,” or “tactical teams.” The user instructs the assistant to send these options to a hypothetical commander, who ultimately selects a tactical team.
The final step is performed quickly. The analyst asks the AIP assistant to “analyze the battlefield”, then “generate routes” for troops to reach the enemy, and finally “assign jammers” to jam communications equipment. Within seconds, analysts finalize the battle plan and order troops to deploy.
In this scenario, Claude is the “voice” of the AIP Assistant and the “reasoning” that the AIP Assistant uses to generate responses. Other AIP demos show users interacting with large language models in much the same way. For example, in a blog published last week, Palantir detailed how Maven Smart Systems customer NATO can use AIP agents within its tools.
One diagram from Palantir shows how third-party defense contractors can choose from several built-in AI models in Palantir, including various versions of OpenAI’s ChatGPT and Meta’s Llama. The user selects OpenAI’s GPT 4.1, which apparently also means the soldier may have the option to select Claude instead.
The analyst then views a digital map showing the location of troops and weapons. Clicking a button on a panel labeled “COA” (course of action) launches a tool powered by GPT-4.1 that generates five possible military strategies, including one called “artillery support, penetration, impact, and destruction.”
Another example shows how the system helps interpret satellite images. The analyst selects the detection of three tanker trucks on the map and loads it into the AIP agent’s chat interface, asking it to “interpret” the image and suggest options for what to do next.
Claude could also be used by the military to create intelligence assessments that would inform future attack plans. In June 2025, WIRED watched a demonstration by Kunal Sharma, head of public sector at Anthropic, that showed how Claude’s Enterprise version could be used to generate an “advanced” report on an actual Ukrainian drone attack called Operation Spiderweb. In the demonstration, Sharma explained, Claude relied solely on publicly available information. But partnering with Palantir also allows the federal government to leverage internal data sets, he said.
“This usually involves sitting down for about five hours with a cup of coffee, reading Google, going into a think tank, writing a report, writing a quote,” Sharma said. “But I don’t have time for that.”
In the demo, Sharma asked Claude to create an “interactive dashboard” containing information about Operation Spider’s Nest and convert it into “object types” that could be analyzed with Foundry, one of Palantir’s off-the-shelf software products. He also asked Claude to write a detailed analysis of recent developments in Russia’s border areas and a 200-word summary of the “military and political implications” of the operation.
“Frankly, I’ve been reading these kinds of books for 20 years. I used to write and am an academic myself,” Sharma said. “This is actually pretty good.”
