In This Story
Background Information: Baroni Center On-Demand Video: AI in Proposal Evaluations: What Could Go Wrong? May 19, 2026
Federal acquisition leaders are under growing pressure to move faster. Agencies are being asked to modernize procurement operations, reduce procurement administrative lead time, and process increasingly complex acquisitions with a workforce that is often stretched thin. At the same time, proposal volumes continue to increase, technical submissions are becoming more sophisticated, and evaluators are expected to absorb massive amounts of information under compressed timelines.
Against that backdrop, it is not surprising that agencies are beginning to explore how Artificial Intelligence (AI) could assist the source selection process.
The potential advantages are easy to understand. AI-enabled tools may help evaluators identify compliance gaps, organize proposal content, compare submissions against solicitation requirements, summarize technical approaches, flag inconsistencies, and accelerate the administrative aspects of evaluation. For agencies managing large procurement workloads, even modest efficiency gains could significantly reduce acquisition cycle times.
But source selection is not a document processing exercise. It is one of the most sensitive and scrutinized functions in federal procurement. Federal Acquisition Regulation (FAR) Part 15 evaluations require agencies to conduct source selections in accordance with solicitation criteria, document evaluation conclusions, and preserve records capable of supporting award decisions under protest scrutiny. AI-assisted evaluations therefore introduce not only operational questions, but governance and defensibility questions as well.
The central challenge is whether agencies can use these tools in a way that preserves trust, transparency, accountability, and the integrity of the acquisition process.
The Appeal of AI-Assisted Evaluations
There are legitimate reasons agencies are considering AI-enabled support during proposal evaluations.
Most evaluation teams face familiar operational problems:
- Large proposal volumes
- Limited evaluator availability
- Compressed acquisition schedules
- Inconsistent documentation quality
- Significant administrative burden
- Difficulty maintaining evaluation continuity across long procurements
AI tools may help address several of these challenges. For example, agencies could potentially use AI to:
- Cross-reference proposals against solicitation instructions
- Identify missing or incomplete submission elements
- Organize technical narratives by evaluation factor
- Highlight inconsistencies within proposals
- Summarize lengthy technical approaches for evaluator review
- Support document comparison across offerors
- Reduce repetitive administrative review tasks
Used appropriately, these capabilities could allow evaluation teams to spend more time exercising judgment and less time managing paperwork.
There is also a broader acquisition workforce issue driving interest in automation. Many agencies are operating with constrained staffing levels while simultaneously managing increasingly complex technical procurements involving cloud computing, cybersecurity, artificial intelligence, digital modernization, and advanced analytics. AI-assisted evaluation tools may be viewed by some organizations as a way to help acquisition personnel manage rising complexity without proportionally increasing staffing requirements.
None of these motivations are unreasonable. In fact, many mirror earlier procurement modernization efforts involving electronic procurement systems, automated compliance tools, and digital acquisition platforms.
The difference is that modern AI systems may participate analytically rather than merely administratively.
Not All "AI" Carries the Same Risk
One challenge in discussing AI-enabled source selection is that the term “AI” covers very different categories of technology with very different risk profiles.
Some acquisition support tools are essentially deterministic automation systems. These tools apply predefined rules to perform structured tasks such as checking page counts, validating attachment presence, or identifying missing representations and certifications. These functions are generally easier to explain, validate, and defend because the logic is relatively transparent.
Other tools rely on machine learning models that identify patterns or classify information based on training data. These systems may help organize proposal content, categorize technical narratives, or identify similarities across submissions.
Generative AI tools introduce a different level of complexity. These systems may summarize proposals, draft evaluation narratives, recommend strengths or weaknesses, or produce analytical observations that are not fully traceable to predefined rules.
The distinction is important because explainability and oversight requirements increase as AI systems move closer to subjective analysis rather than procedural support. A rules-based compliance checker presents a very different governance challenge than a generative AI system recommending technical discriminators between offerors.
As agencies evaluate potential use cases, they will likely need to differentiate carefully between automation that supports process execution and AI systems that may influence evaluative reasoning.
The Defensibility Problem
The most significant risk surrounding AI-assisted source selection is not necessarily technical failure. It is the possibility that agencies may adopt tools faster than they develop governance frameworks capable of defending their use.
Federal source selections operate within a highly scrutinized environment. Every rating, discriminator, weakness, and tradeoff decision may ultimately become part of a protest record. Agencies are expected to demonstrate that evaluations were conducted fairly, consistently, and in accordance with the solicitation. That expectation becomes more complicated when AI participates in the evaluation process.
Companies that lose competitions may reasonably ask:
-
How was the AI tool used?
- What role did it play in evaluator conclusions?
- Can the agency explain how the tool generated its outputs?
- Were all offerors evaluated consistently?
- Did evaluators independently validate AI-generated findings?
- Can the government reconstruct the evaluation rationale?
These questions become especially important if agencies cannot clearly distinguish between AI supporting evaluators and AI effectively shaping evaluative judgment.
The issue is not simply whether a tool produced an inaccurate output. Even a technically capable system may create protest exposure if the agency cannot explain how conclusions were reached or demonstrate that government officials exercised independent judgment. In federal procurement, defensibility often matters just as much as, or often more than, efficiency.
A Practical Scenario
Consider a hypothetical cybersecurity services procurement under a large multiple-award BPA competition.
An agency receives twenty highly technical proposals averaging several hundred pages each. To reduce evaluation timelines, the agency deploys an AI-enabled summarization platform to organize technical content by evaluation factor and generate concise summaries for evaluator review.
At first glance, the approach appears low risk. Evaluators still read proposals and retain responsibility for ratings. But complications may quickly emerge.
One offeror later argues during a protest that the AI summaries omitted key technical differentiators discussed deep within its proposal narrative. Another questions whether the AI tool interpreted all offerors consistently. Evaluators themselves struggle to explain exactly how the tool prioritized certain proposal elements over others because the underlying model behavior is not fully transparent.
Even if the agency ultimately prevails, the process introduces additional scrutiny regarding documentation, consistency, and evaluator independence.
Whether the agency can clearly demonstrate that human evaluators exercised independent judgment and that the procurement record remains explainable and defensible is the primary issue.
The “Human in the Loop” Question
Much of the debate around AI in source selection ultimately centers on the role of the human evaluator. Most acquisition professionals would likely view administrative support functions as relatively low risk. For example, using AI to organize proposal sections, identify missing attachments, or summarize technical content may not materially differ from other advanced software tools already used in procurement operations. The risk profile changes, however, when AI begins influencing subjective evaluation judgments.
Technical evaluations frequently require nuanced assessments involving:
- Risk
- Feasibility
- Management approach credibility
- Technical tradeoffs
- Past performance relevance
- Innovation merit
- Value judgments across competing strengths and weaknesses
These are not purely mechanical determinations. They involve contextual judgment, experience, and interpretation.
If agencies become overly reliant on AI-generated scoring, rankings, or qualitative conclusions, they may create the perception that inherently governmental judgment has been delegated to a software platform. FAR Subpart 7.5 already establishes limitations surrounding inherently governmental functions and closely associated support activities. Whether AI-assisted evaluations ultimately fall within those boundaries will likely depend heavily on how agencies structure oversight and retain decision authority.
The most defensible model is probably one in which:
- AI supports administrative and analytical tasks
- Human evaluators independently assess conclusions
- Government officials retain accountability for all judgments and ratings
- Evaluation records clearly document human concurrence and rationale
In other words, AI may help evaluators work faster, but it cannot replace evaluator responsibility.
Is AI-Assisted Evaluation Inherently Governmental?
As agencies experiment with AI-enabled acquisition tools, questions will inevitably arise regarding inherently governmental functions.
This issue deserves careful treatment because the answer is unlikely to be absolute.
Federal procurement already relies on substantial support infrastructure. Agencies routinely use software systems, evaluation support contractors, pricing tools, data analytics platforms, and advisory support throughout the acquisition lifecycle. The existence of support mechanisms alone does not necessarily create inherently governmental concerns.
The more important issue is where decision authority resides.
If AI tools are being used to organize information, support analysis, identify patterns, or assist evaluators administratively, agencies may argue that the technology is functioning as a decision-support mechanism rather than a decision-maker.
The concern becomes more substantial if:
- AI-generated conclusions are adopted without meaningful review
- Evaluators cannot explain or defend outputs
- Tool recommendations effectively drive award decisions
- Agencies rely on opaque scoring methodologies
- Human participation becomes largely procedural
At that point, critics may argue that the government is delegating evaluative judgment rather than merely augmenting evaluator capability.
This is one reason agencies should approach AI-assisted source selection incrementally rather than attempting immediate end-to-end automation of complex evaluations.
AI Adoption Will Likely Occur in Stages
For many agencies, AI adoption in source selection will probably evolve through maturity stages rather than through immediate full-scale automation.
Early implementation will likely focus on lower-risk administrative augmentation functions such as:
- Compliance verification
- Proposal organization
- Data extraction
- Document summarization
- Workflow coordination
As confidence and governance mature, agencies may experiment with analytical support functions such as identifying patterns, highlighting potential inconsistencies, or assisting evaluators with comparative review preparation.
The most sensitive stage would involve AI systems influencing qualitative assessments, discriminator analysis, or tradeoff reasoning. That is where oversight, explainability, and defensibility concerns become significantly more consequential.
This phased approach matters because agencies do not need to resolve the hardest governance problems before pursuing lower-risk modernization opportunities. Incremental adoption may allow acquisition organizations to build internal familiarity, governance discipline, and workforce competency before introducing more complex AI-assisted functions.
Procurement Data Security Cannot Be Treated as Secondary
Any discussion of AI-assisted source selection must also account for procurement-sensitive information handling.
Source selections routinely involve proprietary technical data, pricing information, controlled unclassified information (CUI), source selection sensitive information, and contractor intellectual property. Agencies evaluating AI-enabled platforms will therefore face questions extending well beyond evaluation methodology alone.
Important concerns include:
- Whether proposal data is retained or used for model training
- Where procurement information is stored and processed
- Whether AI platforms operate within authorized FedRAMP environments
- How agencies prevent cross-customer data exposure
- Whether audit logs preserve chain-of-custody visibility
- How access controls are enforced throughout evaluations
These issues become especially important for classified, national security, or highly sensitive acquisitions. Even highly capable AI tools may struggle to gain adoption if agencies cannot confidently address information security, procurement integrity, and data governance requirements.
Building Guardrails Before Problems Emerge
If agencies intend to incorporate AI into source selection activities, governance will matter as much as technology capability. Strong guardrails can help agencies pursue modernization while preserving procurement integrity and defensibility.
Maintain Human Validation Requirements
AI-generated findings should not become final evaluation conclusions without documented human review. Evaluators should independently assess outputs, validate recommendations, and retain responsibility for all ratings and tradeoff decisions.
Preserve Auditability
Agencies should ensure that AI-assisted evaluations generate traceable records showing:
- How tools were used
- What outputs were generated
- What evaluators accepted or rejected
- How final conclusions were reached
If an agency cannot reconstruct the evaluation process during a protest, it may struggle to defend the procurement regardless of whether the underlying evaluation was reasonable.
Start with Lower-Risk Functions
Agencies may reduce implementation risk by initially limiting AI use to:
- Compliance checks
- Document organization
- Administrative review support
- Data extraction
- Proposal summarization
These functions generally present fewer concerns than qualitative scoring or comparative tradeoff analysis. While there is certainly tremendous value in the latter, it may make sense for agencies to build up to that level of sophisticated AI use.
Establish Governance Policies Early
AI implementation should not begin with tool deployment alone. Agencies will likely need:
- Internal governance frameworks
- Usage policies
- Evaluator training standards
- Approval authorities
- Documentation requirements
- Legal and procurement oversight involvement
Waiting to develop governance until after protests emerge would be a far more difficult position and it may deter the acquisition workforce from exploring the use of AI.
Train the Workforce
Acquisition personnel must understand both the strengths and limitations of AI-assisted evaluation tools. Evaluators should be trained to question/challenge outputs, identify potential inaccuracies, and avoid overreliance on automated recommendations. Also having a basic level of AI literacy about how LLMs are developed is beneficial for successful adoption.
The Responsibility of AI Vendors
Technology providers also have an important role to play in reducing procurement risk. Many commercial AI products are designed primarily for speed and user convenience. Federal source selection environments demand something different: explainability, traceability, accountability, and procedural defensibility.
AI vendors seeking success in the federal acquisition space may need to prioritize features such as:
- Detailed audit logging
- Transparent output traceability
- Version control
- Evaluator override functionality
- Configurable governance settings
- Procurement-specific workflows
- Bias monitoring capabilities
- Record preservation support
Tool providers should also recognize that procurement environments are not typical enterprise automation settings. Source selections involve formal procedures, legal scrutiny, documentation standards, and potential protest review. An AI tool that produces fast outputs but cannot help agencies explain how conclusions were reached may struggle to gain long-term adoption in federal procurement.
Vendors that understand the procedural realities of acquisition defensibility will likely be better positioned than those focused solely on automation speed.
The Path Forward
AI-assisted source selection is unlikely to remain theoretical for long. The operational pressures driving acquisition modernization are real, and agencies will continue exploring technologies that can improve procurement speed and workforce efficiency.
The question is not whether AI will enter the acquisition process. In many ways, it already has.
The more important question is whether agencies can integrate these tools without weakening the transparency, accountability, and defensibility that federal procurement requires.
Used carefully, AI may help evaluators manage growing workloads and reduce administrative friction. Alternatively, it could create new protest vulnerabilities, reduce confidence in procurement outcomes, and complicate already sensitive acquisition decisions.
The agencies that succeed will likely be the ones that treat AI not as a replacement for procurement judgment, but as a governed support capability operating within clearly defined human oversight.
In federal acquisition, faster decisions only matter if they remain defensible decisions.
Read this article and others from the Greg and Camille Baroni Center for Government Contracting at our website.