Automating Judgement - AI's Promise and Peril in Federal Procurement
Updated: Aug 11
Artificial Intelligence (AI) is being hailed as the fourth industrial revolution. While companies like Apple and the MAG 7 are leveraging astonishing valuations to fuel a relentless cycle of debt-driven capital expenditures, opinions on AI's future diverge sharply. On one side, Tech enthusiasts and Star Trek fans envision a world of unlimited abundance, where work becomes obsolete. On the other, the doom-focused Sci-Fi crowd warns of a dystopian future marked by mass unemployment. Predicting how new technology will shape our future is often a fool's errand. Human systems and politics are complex and unpredictable. The truth is, we won’t know the full impact for some time.
Current Landscape of AI and Revenue
What do we know right now? No one seems to be making any money. Investors, weary of riding a wave of hype-driven valuations, are beginning to ask, “Where is the revenue?” The use of GPUs with a limited lifespan as collateral appears risky. There are many questions surrounding the private credit market. Examining CDS spreads may indicate that holders of this debt are getting nervous and hedging against downside risk. Even some of the more famous AI researchers are making public pivots that could spook investors. But what does all this mean for competitive public sector work? One thing is certain: there will be an acceleration in integrating AI into every possible government workflow. Revenue needs to come from somewhere, fast. Lobbyists and governments have a vested interest in this integration. Historically, during debt-fueled manias, the public often ends up holding the bag.
The Complexity of Government Acquisition
State and federal acquisition processes are complex, high-effort, and heavily regulated. The United States Army is working on an AI approach to generate reverse RFIs without the typical solicitation/response cycles. Integrating AI into competitive source selection is another area actively discussed within government agencies. The incentives are palpable, even at the operational level within organizations. Requirement owners, engineers, and other professionals are often pulled from their daily tasks to evaluate proposals. This isn’t an enjoyable duty, and they would welcome a black box solution to streamline the process. It isn’t a question of if, but when we will see this happen.
Today’s article is a cursory exploration of existing regulations, technical pitfalls, and potential job growth for aspiring contract law attorneys and expert witness software engineers looking for a career change. Most importantly, we will explore what this means for your business. Perhaps we can find a silver lining!
Regulatory Landscape and AI Integration
The White House issued Executive Order 14275 – Restoring Common Sense to Federal Procurement in April 2025. This order kicked off a FAR overhaul with debatable results relative to its core intent. OMB-25-21 issued a memo urging agencies to promote ‘high-impact’ AI use cases. EO 14179 was signed shortly after a new administration took office, aiming to clear out regulatory obstacles. EO 14365 goes even further, banning states from issuing regulations that could hinder AI development. The Department of War has issued guidance to further accelerate AI integration into every workflow possible. The gist is that the Trump Administration has a maximalist, all-in view of AI. Debt bubbles and regulatory pain be damned; it’s full speed ahead!
Challenges with AI in Proposal Evaluation
The FAR amendments, to this point, haven’t changed enough to allow for a fully AI-driven source selection for future proposals. FAR 1.102 sets the basic rules for the entire process and requires human involvement in decision-making. The fundamental principle is maintaining public trust in the process through an accountable body. An LLM cannot be held accountable and lacks legal personhood or professional responsibility. If AI makes a flawed recommendation that leads to an improper award, who is accountable?
FAR 15.308 places the final decision-making authority with the Source Selection Authority (SSA). If the SSA is presented with a ranked list of offerors from AI and simply signs off on the top choice, they have not exercised "independent judgment." They have deferred their judgment to the machine, violating core principles outlined in the FAR.
FAR 15.305 establishes how the government evaluates proposals. This section mandates that the agency must evaluate proposals based solely on the factors and sub-factors specified in the solicitation. The evaluation must be consistent, equitable, and documented in sufficient detail to support the final decision. Herein lies a significant technical challenge with LLMs: hallucinations. LLMs can generate responses that sound plausible but are factually incorrect. Scaling was supposed to address hallucination issues, but it has not delivered as promised. Numerous studies show a wide range of hallucination probabilities. During any mania, the for/against divide produces highly variable data. However, we know that hallucinations occur and will pose legal challenges to the government’s approach. Simply prompting, “Hey Claude, review this proposal and make no mistakes,” won’t satisfy a risk-averse contracting officer.
FAR 15.506 entitles offerors to a debrief. However, it remains unclear what this debrief will look like given the broader FAR rebirth. The government must provide enough information to determine if a black box performed your evaluation. Outputs that are unexplainable and potentially incorrect pose significant risks to the government, undermining the entire concept of accountability. Currently, there is no case law regarding a protest based on an AI evaluation. However, this will be an interesting process to observe, likely favoring your company. The first court case will set a significant precedent for how contractors compete for work in the public sector.
Anticipating Legal Risks and Strategic Approaches
So, what is the likely near-term outcome? Government attorneys will almost certainly identify these massive legal risks. The initial, rational approach will likely be to limit LLMs to purely administrative tasks, such as checking page counts, font sizes, and basic compliance. However, given the current mania and immense pressure to show a return on AI investment, it’s hard to predict how much pain will be inflicted before a rational, compliant approach is universally adopted. In my opinion, the industry is about to become the guinea pig.
Instead of waiting to become a test case, proactive companies can adapt their proposal processes now to gain a significant competitive edge. Here’s how you can prepare for the AI evaluator.
A. Write for Machines and Humans
Your proposal’s structure and language will matter more than ever. The goal is to make your proposal both machine-readable and human-persuasive. Use clear headings that map directly to the RFP's section numbers (e.g., "Volume 1, Section 3.1.a: Response to Technical Subfactor 1 - System Architecture"). This allows AI to easily parse your document and align your content with the evaluation criteria.
Use simple, declarative sentences. Avoid complex, multi-clause sentences. Use clear, direct language. State your compliance and strengths explicitly. For example, instead of saying, "Our team is capable of...", write, "Our team meets this requirement. We will provide..."
B. Front-Load Your Compliance
Research on LLMs has revealed a phenomenon known as "lost in the middle," where the model pays more attention to the beginning and end of a document or section. AI is not a human evaluator who can read back and forth to connect disparate ideas. Assume it has a limited attention span. Therefore, state your most critical compliance points and win themes at the beginning of each section. Master the art of technical writing for AI.
C. Ask Strategic Questions During the Q&A Period
Use the official Q&A process to establish a written record. The government will likely disclose AI evaluation in the RFP. However, you should ask, "Will the government be using any automated tools, artificial intelligence, or machine learning algorithms to assist in the screening, scoring, or evaluation of proposals? If so, can the government describe what safeguards will be in place to ensure the evaluation is conducted in accordance with FAR 15.305 and that human evaluators will validate all outputs?"
D. Request a Thorough Debriefing
If you lose, the debriefing will be crucial in this landscape. Thanks to the "enhanced debriefing" rules, you have the right to ask follow-up questions in writing. Even if the agency discloses its use of AI in the RFP, you are still entitled to a rational explanation of the evaluation. If the agency’s rationale seems thin, nonsensical, or they can't explain why you received a certain rating, it’s a major red flag that they are simply parroting an AI's output without understanding it. This is critical information for potential protests.
Navigating the New Landscape
Navigating this new landscape requires more than just good proposal writing. It requires a new strategy. BWIT Solutions can help you compete in this ever-evolving landscape with:
A. AI-Resilient Proposal Architectures
We don't just write your proposals; we architect them. We structure your content to be easily parsed by AI evaluators while remaining compelling and persuasive to the human Source Selection Authority. We help you "write for the machine" without losing the narrative that makes you unique.
B. Strategic RFP Analysis and Q&A
We analyze solicitations for AI-related risks and opportunities. We help you craft strategic questions that protect your rights and put the agency on notice, setting the stage for a successful bid long before the proposal is even submitted.
C. Debriefing and Protest Support
We assist you in dissecting debriefings to identify the tell-tale signs of a flawed, AI-driven evaluation. We formulate critical follow-up questions designed to expose a lack of rationale.
Please visit us at our Contact page if you’d like to discuss further! Perhaps over a cup of Earl Grey from the BWIT replicator? Thanks for reading!
Comments