Government Use of AI to Evaluate Proposals Comes Under Scrutiny

If an agency used AI to evaluate your proposal, can you see that report? A recent Court of Federal Claims decision says yes, even if the agency claims it didn’t rely on it.

During an evaluation, the government used an AI platform to review offerors’ proposals and then argued that those AI findings did not belong in the administrative record. The Army claimed that the reports were not “useable” and “never entered into the contract file or [were] utilized in the decision-making process.” The U.S. Court of Federal Claims still rejected that position in TRAX International Corp. v. United States, No. 26-796 (publicly reissued on September 22, 2026).

The court explained why all AI-generated evaluations that were considered had to be included in the record. The distinction driving that ruling is important for contractors: An agency’s consideration of an AI evaluation is not the same thing as its reliance on that evaluation. Simply saying the AI did not influence the award did not justify excluding reports.

The dispute arose from an Army procurement for mission support services at White Sands Missile Range. TRAX initially protested the award to Southwest Range Services (SRS) at GAO, which denied the protest. The court noted that GAO did not discuss TRAX’s AI allegations.

During the procurement, an Army analyst was tasked with testing FAST TRACK, an AI platform, to determine whether it could be used in future source selections. The analyst generated evaluations of three proposals. The government initially represented that the Source Selection Evaluation Board (SSEB) had viewed only one of the AI reports and found it “not useable” and not a “reliable tool.”

However, supplemental declarations revealed that the AI evaluations of TRAX and SRS had also been distributed to three individuals involved in the procurement, including the contracting officer and two evaluation-board members. What the government characterized as a “test” for future source selections had involved actual proposals and officials participating in an ongoing procurement.

The court identified three reasons the reports belonged in the administrative record.

First, the AI evaluations were relevant to the procurement because they directly concerned the competing proposals. They were not generic demonstrations of what the platform could do. The court stated, “[t]he AI evaluations are ‘relevant to the process’ by which the Army made its decision[,] because they ‘were generated specifically to reflect the comparative merits of the proposals received in response to the RFP.’” Because they bore directly on the merits of actual proposals, they were directly relevant to TRAX’s claim that AI-assisted review introduced errors that affected the Army’s evaluation.

Second, the evaluations were generated during the agency’s decision-making process, not after it. The government claimed the reports were generated only after the evaluations were finalized, but the court found that was likely not the case given “the overlap between the AI evaluation for [one offeror] and the SSEB’s final evaluation,” which suggested the AI output may have informed the SSEB’s work.

Third, agency decision makers actually reviewed the reports. For one offeror, the entire evaluation board convened to view the AI output. Whether the SSEB liked what it saw was beside the point. As the court explained, “[w]hether the SSEB ultimately found that output ‘useable’ or ‘reliable’ has no bearing on whether it was considered” (emphasis added). For the TRAX and SRS reports, distribution to the contracting officer and two SSEB members was sufficient to establish consideration for purposes of completing the record.

This is the practical significance of the ruling. The court needed the material that was before agency decision-makers, not merely the documents the government identified as supporting the final award. The AI evaluations were the missing parts of that record. Describing them as an unsuccessful test did not remove them from it.

Once the reports were included, the court found multiple passages in the Army’s evaluation of the third offeror that were nearly identical, sentence for sentence, to that offeror’s AI-generated evaluation. That finding illustrates why the record issue mattered. Producing the reports allowed the parties and the court to compare the AI work against the official evaluation, rather than simply accept the government’s explanation that the outputs were not used.

The court nevertheless drew a limit. Overlap involving TRAX’s and SRS’s evaluations was minimal, and TRAX identified no actual error or inaccuracy in the challenged findings that it could trace to AI. Similar conclusions did not establish that AI had corrupted the evaluation. Nor had TRAX shown why any AI reliance independently undermined the Army’s findings. The court therefore rejected the request to give those findings less deference based on possible AI use.

The reports belonged in the record even though their inclusion did not establish the AI-related error TRAX alleged. Those were separate questions.

For contractors, the TRAX case suggests a more precise question than whether AI was generally used in a procurement: Did evaluators receive and “consider” AI-generated assessments of the proposals during the evaluation? If yes, they should be included in the record.

Now, an agency may describe a tool as experimental, advisory, or even unhelpful. But those descriptions do not answer whether its outputs were “considered” during the evaluation process. Contractors might then consider asking, during a debriefing, whether proposal-specific AI reports were generated, when they were generated, and who received or reviewed them. Where those reports are later omitted from the protest record, TRAX provides a basis for digging deeper into those questions.

We note that this decision does not establish a blanket entitlement to every AI-related document, nor does it require agencies to preserve or produce every AI-generated analysis for an entire procurement. It shows that AI reports the agency “considered” during evaluation may need to be produced if relevant to specific allegations.

Thus, TRAX makes clear that AI-generated assessments can be part of that record, even when the agency says it did not rely on them.