How to Document Independent Human Judgment When Using AI Tools in Contract Awards

When an agency uses AI to sort, summarize, or draft procurement materials, its records should show where the tool stopped and where human judgment began. A clear record protects the integrity of the award and gives evaluators a practical way to explain their work during a protest.

Federal agencies face real pressure to review long proposals quickly. AI tools can help organize large records, identify passages for closer review, and prepare draft summaries. The source selection decision still must rest on documented human analysis. Federal Acquisition Regulation 15.308 requires the source selection authority to exercise independent judgment and document the basis for the award. That requirement gives agencies a useful framework for governing AI use.

The recent Court of Federal Claims decision involving Trax International Corp. shows why this documentation matters. Trax argued that the Army had relied on an AI-generated evaluation containing apparent hallucinations and made-up references. The court ruled against Trax because the company did not connect the Army’s use of AI to a specific evaluation error or explain how the tool undermined the award decision. The court also required the government to produce AI-generated evaluations and later sought additional information about government representations concerning those materials.

The decision leaves agencies with a practical lesson: an AI tool can support procurement work, but the administrative record must show the reasoning that led to each material finding. I would build that record through the following steps.

Start by writing down exactly what the tool may do in the procurement. A short AI use plan can identify approved functions such as:

The plan should also identify functions that require additional controls. Examples include assigning strengths or weaknesses, making responsibility judgments, ranking offerors, assessing technical acceptability, making price or cost conclusions, and drafting final source selection findings.

The purpose of this exercise is to establish the tool’s place in the workflow before evaluators begin using it. A written scope gives the team a common reference point and helps the contracting officer explain the process later. It also makes it easier to identify the human decisions that require direct evaluator attention.

Create a responsibility chart for every significant evaluation step. The chart should identify the person or group responsible for reviewing the proposal, verifying the relevant facts, making the evaluative judgment, and approving the written conclusion.

For each factor and subfactor, record the assigned evaluator, the chair or team lead, the contracting officer, and the source selection authority. If a subject matter expert reviews an AI-generated summary, identify that reviewer by name and role. If a technical lead accepts or changes draft language, preserve that person’s approval in the working file.

This step matters because generic references to the evaluation team provide limited evidence of independent judgment. A record becomes more useful when it shows who reviewed a specific conclusion and what that person approved. Individual accountability also encourages evaluators to examine AI output carefully before using it.

Maintain an AI use log throughout the procurement. The log should capture enough information for someone outside the immediate team to understand how the tool affected the work.

At a minimum, include:

A material use includes any AI output that influences an evaluator’s research, draft language, finding, rating, recommendation, or source selection document. Routine administrative uses may need less detail, but the agency should define that category in advance. When in doubt, preserve the record.

The log can be maintained in a procurement file, a controlled collaboration system, or another approved records repository. The format matters less than consistency, access controls, and the ability to connect each AI output to the resulting human work product.

Save the original prompt and the complete response before anyone edits the material. Preserve the date, tool version, and any settings that could affect the result. Screenshots can supplement the record, but an export or system-generated audit file usually provides better evidence when available.

Keep the original output separate from the edited evaluation document. That separation allows the agency to show what the tool produced, what the evaluator changed, and what language appeared in the final record. It also helps identify whether an apparent hallucination entered the evaluation through an AI output or through a later human edit.

Access controls deserve attention. Proposal information may include proprietary data, controlled unclassified information, personal information, or other protected material. The agency should use approved tools, limit access, and follow applicable security and records requirements. A helpful workflow becomes a liability when personnel place protected proposal content into an unauthorized system.

AI systems can produce plausible statements that lack support in the proposal or solicitation. Build a verification step into the evaluation process so that every material factual statement has a source.

An evaluator should be able to trace a finding to a proposal page, solicitation provision, amendment, clarification, oral discussion record, or other authorized source. The evaluation file can use a citation table with columns for the finding, source location, reviewer, verification date, and disposition.

For example, a table might show that an AI tool identified a possible staffing gap, that the evaluator reviewed the proposed staffing plan on pages 42 through 47, and that the evaluator confirmed the gap after comparing the plan with the solicitation’s minimum staffing requirement. If the evaluator finds that the AI output misunderstood the proposal, the table should record the correction and explain the result.

This process gives evaluators a practical way to catch fabricated references, distorted quotations, and unsupported assumptions. It also shows a court how the agency reached a conclusion. The record should demonstrate the path from source material to evaluator judgment.

AI can help prepare a draft, but the evaluator should own the analysis. Require evaluators to explain the significance of the evidence in their own words and to identify the judgment they made after reviewing the source material.

A strong evaluation statement generally answers four questions:

The answer to each question should reflect the solicitation’s evaluation scheme. A generic statement that an offeror demonstrated a strong approach provides little support. A specific statement describing the proposed approach, the relevant requirement, and the resulting benefit gives the source selection authority a meaningful basis for review.

When AI helps draft language, the evaluator should mark the draft as reviewed, identify material edits, and confirm that the final statement reflects the evaluator’s independent assessment. A certification or signoff can state that the evaluator reviewed the cited source material, verified the factual assertions, and accepts responsibility for the analysis.

Some findings deserve a second human review because they can materially affect an award. These may include deficiencies, significant weaknesses, adverse past performance conclusions, noncompliance findings, major price or cost issues, and conclusions that eliminate an offeror from consideration.

The second reviewer should examine the underlying proposal material, the AI output if it influenced the work, the evaluator’s analysis, and the applicable solicitation language. The reviewer should record whether the finding is supported, whether the rating is consistent with the evaluation scheme, and whether any revision is needed.

A second review should involve substantive analysis. A signature alone provides limited value. The reviewer can use a short checklist that asks whether the factual references are accurate, whether the conclusion follows from those facts, whether the record contains contrary evidence, and whether the final language accurately states the solicitation standard.

The source selection authority’s decision should explain how the authority considered the evaluation results and reached the award decision. The document should identify the comparative advantages that drove the selection and connect those advantages to the solicitation’s stated factors.

If the authority departs from an evaluator recommendation, the record should explain the reason. If the authority relies on a finding that received additional review, the source selection document should identify the relevant record support. If the authority rejects an AI-generated suggestion after human review, that fact may be useful when it shows the safeguards operating as intended.

The decision document should stand on the contemporaneous record. After-the-fact explanations carry less weight in a protest because they can appear to reconstruct reasoning that the original file did not show. I would therefore schedule a source selection record review before award, with particular attention to unexplained conclusions and unsupported boilerplate.

Give evaluators a clear process for handling suspected AI errors. They should know how to flag the output, verify the underlying material, correct the work product, and report a recurring problem with the tool.

The process can require an evaluator to:

A recurring error may indicate that the tool is poorly suited to the procurement or that users need additional training. The agency should record the decision to continue, restrict, or discontinue the tool. That decision provides useful evidence of responsible oversight.

Assume that AI-related materials may become relevant to a bid protest. Identify the materials that belong in the administrative record and preserve them in an organized way. This can include prompts, outputs, use logs, evaluator notes, verification tables, review checklists, tool settings, training materials, and communications about material AI errors.

The Trax litigation demonstrates that a court may require the government to produce AI-generated evaluations even when the government initially characterizes them as outside the decision-making process. Agencies should therefore avoid informal systems that make relevant materials difficult to locate or characterize.

A records custodian should be able to answer basic questions quickly: Which tools did the team use? Who used them? What did they produce? Which outputs influenced the evaluation? Who reviewed the outputs? What corrections occurred? How did the final decision rely on the verified record?

Clear organization also helps the Department of Justice defend the agency’s position if a protest is filed. The agency’s litigation team should receive a complete and accurate file before preparing declarations or responding to discovery. Any statement about AI use should be checked against the preserved materials.

Training should use examples drawn from the work evaluators actually perform. Show how an AI tool can invent a proposal citation, merge two separate requirements, overlook a limitation, or produce a confident summary that changes the meaning of the offeror’s approach.

Then have evaluators practice the full control process. They should review the output, locate the source material, mark errors, revise the draft, cite the record, and write the reasoning supporting the final finding. Training should also cover protected information, approved tools, retention, access, and escalation procedures.

Managers should reinforce that speed is only one consideration. A fast draft that requires extensive correction can create more work and greater protest risk. A controlled workflow helps the team gain efficiency while preserving the quality of the award record.

Before award, ask the source selection authority and contracting officer to complete a focused review. The checklist can include the following questions:

The final question deserves special attention. A record filled with polished but generic language can raise concerns even when no single AI error caused a bad rating. Courts and oversight bodies expect the agency to show how evaluators reached their conclusions. Detailed citations and specific reasoning make that work visible.