Two weeks ago, a P-8 aviator asked if I could make a tool to help plan flight schedules when conducting operations against maneuvering submarines. The daily planning cycle for these kinds of operations usually ties down a couple of officers in a squadron every day. It’s the kind of problem that is easy to throw manpower at but would cost millions and take several years to solve by procuring software through the acquisition system. Given those options, commanders default to throwing people at the problem.
Over the last decade, it’s usually taken me a couple of weeks to develop the basic version of an application like this to be shared with users for feedback and iteration. These initial products usually lacked key features, but served as a canvas onto which sailors could sketch what they really needed to solve their problem. Over the course of a couple months, we could deliver a product that really moved the needle.
Today, AI tools reduce this to mere hours. The aviator made that request at 9:04 a.m., and by 12:40 p.m., I had built and sent the initial application back to him, using a combination of OpenAI’s GPT-4.1 and Google’s Gemini 2.5 Pro. Instead of a simple mock-up, it was fully functional. The code was well-documented, with complex math working flawlessly. The AI had even drafted a series of additional features for customers to consider and questions to help tease out their real needs.
The application comprised more than 5,000 lines of well-crafted code, and a human didn’t write a single one of them.
Even without dramatic advances in AI capabilities, these tools will fundamentally change how program offices acquire capabilities and how combat units solve problems. But this will happen only if the Department of Defense thinks systematically about changing how it understands problems, builds solutions, and fields secure final products.
Make More Bets
Traditional acquisition processes force program offices into high-stakes, one-shot decisions. Consider an aircraft program manager having trouble clearing maintenance backlogs, keeping aircraft on the ground. Given two million dollars to build something that solves the problem, that program manager would typically hold lengthy stakeholder meetings, debate endlessly about the right solution, and spend months in contract negotiations to get started. By the time engineers begin coding, operational needs may have changed entirely.
With AI-powered software development, though, that same program manager can immediately attack the problem with a portfolio of possible solutions. He could direct a small team to prototype three different applications in 72 hours: a maintenance work scheduler, a defect root-cause analysis tool, and a simulation-based training tool for maintainers. After two weeks of testing and iteration, real data would replace conjecture.
Perhaps the root-cause analysis tool creates more work than it solves. The training tool shows promise but needs industry refinement to ensure accuracy. And the scheduler demonstrates measurable improvements in aircraft availability while receiving positive feedback from maintainers. Armed with actual performance data, the program manager can make confident decisions about which solutions to pursue. They can scale the successful scheduler prototype in-house, flip the training tool to industry for refinement, and redirect resources away from the less promising root cause analysis tool.
Prototyping isn’t a nice-to-have. It’s the only practical way to uncover the subtle, context-driven frameworks that good software must work within. AI can slash the marginal cost of these experiments. Early user data dictates what to scale, refine, or eliminate. Program managers become portfolio managers, quickly reallocating resources based on measured value rather than on optimistic PowerPoint pitches.
This portfolio approach to development works best when program offices control their own products, but the government often does not own the rights to modify its own systems.
Overcoming Platform Constraints
Program offices fielding flexible, government-owned software should begin using AI to prototype today. The barriers are low, and the tools are available. But not every program office enjoys the luxury of open, modifiable software systems. Many manage platforms dominated by vendor-locked code that runs safety-critical systems with long certification processes.
These constraints might seem to preclude rapid AI-powered prototyping, but they shouldn’t.
Every major weapon system has problems that don’t touch safety-critical code: problematic supply chains, inadequate training resources, and poor post-mission analytics tools. These adjacent problems offer fertile ground for rapid prototyping. A fighter aircraft program office might not be able to modify flight control software quickly, but they can build tools to predict component failures or streamline pilot training schedules. These kinds of applications deliver immediate value while building institutional competency in AI-assisted development.
In the longer term, program offices need to invest in secure, containerized software enclaves within their platforms. These isolated environments should have read-only access to real-time mission system data that lets engineers prototype applications without compromising security or safety. They can quickly deploy prototype algorithms, visualization tools, and decision aids against real operational data, iterating at AI speed. Once proven in the enclave, promising capabilities can enter the formal integration pipeline with confidence in their value.
Failure to build enclaves like this will stop manned platforms from directly benefiting from AI-accelerated software development. Over time, small and unmanned platforms with less onerous assurance processes will compound gains from more rapid development and make those legacy platforms obsolete.
Security as the Critical Constraint
As the military starts developing hundreds of rapid prototypes, cybersecurity will become the primary bottleneck to deploying software quickly. Today’s generative AI models likely introduce more cyber vulnerabilities than the average human. Moving fast while maintaining security requires new approaches to threat assessment and certification.
The Department of Defense’s Software Fast Track Initiative offers a promising foundation for re-architecting security approvals. AI tools excel at generating compliance documentation, but the Department of Defense needs intelligent code review systems that can identify vulnerabilities better than traditional tools without slowing development cycles.
At least one cyber researcher has used a frontier AI model to find an undiscovered zero-day vulnerability in a major operating system. This is the future of cyber defense and because it is also the future of cyber offense, the Department of Defense cannot wait to adopt these technologies.
The military should also develop standardized “system prompts” that guide AI models toward cybersecurity best practices. Academic research indicates that security-focused prompts significantly reduce vulnerabilities in generated code.
Emerging threats require attention as well. Adversaries will likely attempt to poison public AI training data, inducing models to output malicious code or repeat false information. As public data continues training commercial AI models, defense research dollars should support efforts to counter these attacks.
The Strategic Imperative
It is hard to use frontier AI models every day to develop software and not believe we stand at the beginning of the largest revolution in how militaries develop capabilities in human history. The U.S. military was 15 years late to adopt commercial Agile software development methodologies. Because that delay happened in the wake of America winning the Cold War, the consequences were mostly experienced in terms of wasted resources, not lost wars. But if America is late to this AI-centered paradigm shift while facing off against China, the consequences could be far more severe for the free world.
Success requires three key changes: program offices should start using AI to build dramatically more prototype software, should build secure enclaves on their platforms in which they can prototype quickly, and should build cyber testing infrastructure that enables speed while increasing security.
Prototyping non-safety-critical software is the right place to start learning how to build organizations that can effectively and responsibly use imperfect AI tools.
Without experience, using AI can slow development due to increased time fixing bugs. I write significantly more automated tests to validate performance when building with AI. Without AI, the flight planning application described earlier would have required five end-to-end test cases to validate calculations. With AI-generated code, I needed thirty tests to feel confident it was correct. Bouncing code off multiple models and using standard code quality assessment tools also helps minimize problems and maximize speed.
The development of these best practices is a major component of the value that will accrue to program offices using AI to build products. Future AI models will likely achieve in days what now takes years and millions of dollars: integrating weapons onto legacy platforms, developing autonomous drone software, and refactoring safety-critical systems. Program offices that don’t use AI in low-risk settings today won’t have the experience needed to safely use future capabilities.
The window for adaptation is open, but it won’t remain so indefinitely. The military that masters AI-powered development first will gain a decisive advantage in tomorrow’s conflicts.
Sean Lavelle is a Navy aerospace engineering duty officer with master’s degrees in finance and machine learning from Johns Hopkins University and Georgia Tech. He is the founder of the first all active-duty software development team in the Navy and has built and deployed more than 60 software applications to units across the Navy. The views in this article are the author’s and not those of the U.S. Navy, the Defense Department, or any part of the U.S. government. The appearance of, or reference to, any commercial products or services does not constitute Navy or Defense Department endorsement of those products or services.
Image: Midjourney
Jesus Christ is King
Comments are closed.