Replacing Claude could take 18 months, Pentagon users warn

The push follows internal concerns over supply chain risk and vendor security, part of a broader review of which AI providers can be used in sensitive defence environments. But in this case, the technology under scrutiny is already widely deployed.

Anthropic secured a $200 million defence contract in July 2025 and quickly became embedded in military workflows.

Claude went on to become the first AI model approved to operate on classified networks, and officials familiar with its use say adoption was strong. Within the federal government, its models have been widely viewed as more capable than rival systems.

That footprint is now complicating efforts to remove it.

Users inside the Department of Defense say replacing Claude across existing workflows could take up to 18 months, driven less by procurement and more by the work required to unwind it, retraining tools, rebuilding integrations and revalidating systems that already depend on it.

Claude is used across software development, analysis and internal automation, often sitting inside existing systems rather than alongside them. That makes it difficult to isolate without affecting broader workflows.

Reuters had previously reported that the Pentagon used Claude tools to support US military operations during the conflict with Iran, and sources say the technology remains in use despite the blacklisting. One expert described that as “the clearest signal” of how highly the Pentagon values the tool.

Removing it will require systems to be reworked and recertified under strict security requirements, a process that is both slow and costly.

That creates a gap between policy and implementation. While the direction of travel is clear, the systems involved are not easily swapped out without disruption.

Similar dynamics have played out in telecoms and cloud, where removing vendors on security grounds has taken months or years once infrastructure is in place. AI is now following the same pattern, with models becoming tightly coupled to specific workflows and environments.

Anthropic to shoulder price hikes from its US data centres