AI in Defense: Navigating Procurement Challenges in an Era of Rapid Innovation
How Decades-Old Acquisition Processes Must Evolve to Counter Modern Threats
My first job out of school was at Lockheed Martin and I recently became curious about how defense is using AI. Defense has a number of challenges to utilize AI from procurement all the way through getting recommendations.
To understand why there are challenges, it is best to have some understanding some of the history in defense.
How We Got Here
The impacts of World War 2 influence everything from technology to geopolitics. But some of those impacts haven’t changed with the change in times. WW1 was the first time industries were mobilized for war, but it was most successfully implemented with WW2. During WW2, the Cost-Plus contract for companies producing goods for the war effort became standardized and one of the most prevalent procurement methods for defense contracts.
Soon after WW2, the National Security Act of 1947 established the the Department of Defense (DoD), centralizing procurement. This helped with coordination and alignment of different military branch’s goals - eliminating duplication of efforts and inefficiencies.
There was little reform in the 60’s and 70’s, but the procurement process was refined further to enable governement auditing of costs, qualifications for selecting contracts, and more. It really wasn’t until 1984 where major reforms were created. These reforms were in response to major government spending scandals - Pentagon’s $640 toilet seat, $7600 coffee maker, or $436 hammer.
These scandals resulted in the Federal Acquisition Regulation (FAR) in 1984. The overall goal was to ensure the government gets the best value for its money and to promote fairness, accountability and transparency with the procurement process. FAR defines the entire procurement process. You may have heard of companies requiring FAR Compliance to bid on government contracts - this ensures that contractors follow federal laws and regulations.
In the 1990’s, many companies found how difficult it was to bid on federal contracts. In many cases, companies actually refused to sell to government because of all the paperwork and compliance (see FAR Compliance above). So the government introduced the Federal Acquisition Streamlining Act (FASA) focused on encouraging commercial/private companies to bid on government contracts, simplifying procurement procedures.
FASA also encouraged fixed-price performance based contracts to simplify the estimation and charging procedures for commercial customers from the Cost-Plus contracts. At Lockheed, I was working on Cost-Plus contracts and I can only imagine the accounting needed to track time spent on each program for all the employees and figuring out associated overhead to charge accordingly.
In 2010’s and 2020’s, the military realized that commercial technologies were catching up or even surpassing military technologies and there needed to be more ways to acquire new and innovative technologies, so the National Defense Authorization Act (NDAA) was implemented. NDAA expanded Other Transaction Authorities and reformed DoD acquisition systems. This made procurement/funding easier for new innovative ideas and promoted rapid prototyping to demonstrate these new capabilities. Every year since 2016, NDAA has been expanding to encompass the militaries evolving needs.
This is not an all encompassing set of “Acts”, but some of the key aspects during each decade that can provide some insights into how and why defense procurement processes have changed over the years.
With AI, we are entering a new era where the military needs to become even more agile to counter our adversaries capabilities.
Key Challenges to Deploying AI
There are a number of challenges to utilize AI with defense. In the next few sections I’ll discuss the challenges from a contractual and implementation level.
Defense Contracts
AI Startups in the Defense Ecosystem
Most innovations in AI are coming from startups funded by Venture Capital (VC) who usually expect 10x growth every year (post-seed funding). DoD has made excellent strides to promote innovation through Small Business Innovation Research (SBIR) and these programs were further expanded with Defense Innovation Unit (DIU) in 2015 and Air Force Work Project (AFWERX) in 2017 and other similar programs (SDA, NSIN, AAL, etc).
SBIR is a great tool to research & test new technologies and get to a demo, but getting through a budget approval cycle to get a Program of Record (PoR) still takes several years. The time it takes to get a full program is beyond what a normal VC would invest in. To make matters worse, an AI program is likely not domain specific where they could fall under an existing PoR. Making approval even more difficult.
Defense Contracting Focused on Hardware
Defense contracting has been designed for hardware contracts. Today's defense procurement process involves:
Requirements definition by military branches
Request for proposals (RFPs) from potential contractors
Competitive bidding (though with exceptions for national security)
Contract awards with various structures (fixed-price, cost-plus, etc.)
Oversight through the Defense Contract Management Agency and Defense Contract Audit Agency
Usually this means that requirements for the entire project may be defined months or years before the contract is awarded. The requirements may be outdated before development has even started.
This isn’t usually as big of a concern with hardware because hardware isn’t flexible and hopefully the requirements defined shouldn’t change for the hardware. But with software where operational needs are constantly evolving, it would be better to have a more flexible procurement process. Commercial software development has shifted from a standard waterfall approach to Agile methodology over the last decade to create this flexibility to prioritize their target customer’s needs as they change.
Vendor Lockin
Many defense contracts were for complex weapons systems. These systems were customized, proprietary, and required specialized maintenance to support the systems for decades. This made it very difficult for newer entrants to compete if they couldn’t support legacy systems and resulted in just a few contractors (e.g. Prime Contractors: Lockheed, Raytheon, General Dynamics, etc) getting the majority of the deals and follow-on deals.
Technical Challenges of Military AI
Commercial AI development typically follows a pattern of: collect extensive data, train in cloud environments, deploy to standardized platforms, and continuously update based on user feedback. Defense AI must instead: work with limited classified data, train on air-gapped systems with restricted compute resources, deploy to heterogeneous and often legacy platforms, and update through rigorous verification processes that may take weeks rather than minutes - all with potentially degraded comms.
The military process for training new models is more difficult and the objectives are also more stringent. The military AI model must perform better with more fragmented data, limited connectivity (and compute), and with near-perfect accuracy.
But the main reason AI in the military is so challenging is that our adversaries are always creating new innovative solutions changing how we fight. Unlike most AI systems where the data is very consistent for the recommendations. Being able to change the model at the edge provides us with new methods of countering our adversaries, but with a number of challenges.
With respect to data, this can come with a number of challenges:
Data quality: the data may need to be filtered for bad sensors, incorrect operational information (human entered), or comms interference.
Data augmentation: creating syntehtic data that accurately represents adversarial tactics, environmental conditions, and system responses requires both AI knowledge and tactical expertise.
Data fusion: integrated multi-modal data can be challenging because of data quality, timestamps, variable frequencies, etc.
Sharing data / data siloing: data may have different classification levels, legacy systems may have proprietary formats, or a domain may just not be willing to share.
Retraining the model has a number of constraints:
Computational constraints with the assumption comms are degraded (no cloud)
Slower hardware: radiation hardened, MIL-SPEC compliant systems typically lag commercial performance capabilities by several generations.
Orchestration for Training: coordinating distributed training across tacitcal networks with limited bandwidth, high latency, and intermittent connectivity requires specialized solutions.
Hardware designed for inference: edge devices in defense are typically optimized for low-power inference rather than training.
Training of personnel
Validation of model: military operators need efficient procedures and tools to validate updated models before deployment in critical missions. This requires accessible metrics and testing protocols that non-AI experts can confidently execute under time pressure.
Transparency in the model decisions: operators and commanders must understand AI recommendations sufficiently to maintain appropriate trust and recognize potential failures.
Understanding AI model: models trained on historical combat data may inherit biases from previous doctrine, tactics, or sensor limitations. Identifying and mitigating these biases requires careful assessment of training data and model outputs against current operational realities and strategic objectives.
Deploying the new models have security constraints:
Security to deploy: model updates must be cryptographically signed and verified to prevent adversarial tampering. The deployment pipeline must maintain integrity across air-gapped networks and zero-trust environments while preventing unauthorized access or exfiltration of model weights that could reveal capabilities or vulnerabilities.
Sharing models: different classification levels and coalition operations create challenges in sharing models. Systems must support versioning and compatibility across diverse platforms while maintaining appropriate access controls.
Mission assurance with AI failure: operators need methods for ignoring - or marking systems to be ignored - AI systems providing incorrect recommendations. Systems must be designed with appropriate fallback modes that maintain minimal mission capability.
Test and Evaluation: traditional military T&E procedures must be adapted for AI systems that exhibit probabilistic rather than deterministic behaviors. This requires new frameworks for assessing reliability, robustness, and operational effectiveness across the full spectrum of potential scenarios.
Beyond the technical, ethical boundaries and constraints are some of the most important problems to discuss. This includes in what situations must there be a human-in-the-loop vs full automated. When providing recommendations, it is possible multiple sources of data need to be consulted before a decision can be made.
But there is also the consideration that our adversaries may not be taking the same approach (or valued life differently) and have an edge on autonomous solutions. What happens if their autonomous systems respond faster than our human-in-the-loop systems do?
Potential Solutions: Technical and Organizational
At a technical level, many of these challenges can be resolved with new software entries, interoperability across various hardware solutions, and cross-training personnel with baseline AI training knowledge.
At a contractual level, the challenges can be addressed with:
Phase based contracts
Sprint based contract structures (every 2-4 weeks contract is redefined)
Rapid experimentation contracts for “fail fast” approaches
Modular contracting breaking AI programs into smaller specific functions
Push for more interoperable solutions focused on:
Common data standards to be able to utilize data/information across all domains
Cross-domain solutions to easily move data between networks
"Open” APIs to provide data and results across applilcations
I can imagine a future where software designed for AI in defense that can be used across domains. Members of the various domains and coalitions act like a consortium to define the standards for interoperability. Similar to how LLVM or other Open Source projects have corporate members that can help drive the direction of the project.
Many of the suggestions I have written about are aligned with Joint All Domain Command and Control (JADC2) vision. JADC2 represents the Pentagon's vision for connecting sensors, shooters, and decision-makers across all military branches in a single network where AI will play a crucial role in processing and interpreting the massive resulting data flows.
Conclusion
The challenges of AI in defense represent perhaps the most consequential test case for how we integrate software and hardware in the age of AI.
China's approach to military AI differs fundamentally from the U.S. model. Their Military-Civil Fusion strategy deliberately blurs lines between commercial and defense sectors, allowing rapid transfer of commercial AI advances to military applications without the procurement hurdles faced by U.S. contractors.
There are new threats coming out every day, every war will be different, every adversary will have new innovations. We need to be agile to combat different scenarios. War in Iraq was very different than Ukraine. The usage of Drones, AI, and Cyber, have changed the face of war. It’s not just technology that needs to be concerned with the ethics of AI or social media, but the ethics of war and how AI is an active component within. This means having a human in the loop, but giving them a way to make decisions quickly.
For technologists working in AI, understanding defense needs represents an opportunity to solve some of the most challenging and consequential problems in the field. For defense professionals, embracing commercial AI practices—while adapting them to military realities—offers a path to maintaining technological superiority. The gap between these communities needs to close, and those who can translate between them will shape the future of defense technology.
Now is the time to spend smartly on the military.