Flarepoint

TRAX v United States: Army AI Bid Evaluations Ruled Part of the Record

2026-10-10

On 22 September 2026, the US Court of Federal Claims published a ruling in TRAX International Corp. v. United States. It ordered AI-generated evaluations of all three bidders for a $449 million Army contract into the court record, even though the government had told the court the outputs were unusable and never entered into the contract file. TRAX lost its challenge to the award, but on 5 October the court said it needed more information to decide whether to sanction the government over its misstatements about who had seen the AI evaluations. For anyone whose client is assessed by an AI system, the case turns on one point: an AI assessment that reaches a decision-maker is part of the record, whether or not anyone admits relying on it.

What happened

The Army awarded a mission support contract at White Sands Missile Range in New Mexico to Southwest Range Services at a price of $449,372,096, according to the court's opinion. TRAX, the losing bidder, had offered a lower price; Engineering News-Record reported on 6 August that the winning bid was about $29.4 million higher.

TRAX first protested to the Government Accountability Office, which denied the protest on 14 May. In that process the Army conceded that one weakness it had assigned to TRAX's bid “was improper”. GAO found the error caused no prejudice, because the final source selection decision did not discuss that weakness, and the court later agreed. TRAX then filed a sealed complaint in the Court of Federal Claims on 29 May, assigned to Judge Carolyn N. Lerner.

What did the AI produce, and who saw it?

The AI tool was FAST TRACK, an internal Army system. A procurement analyst ran it on the White Sands bids while assessing whether it “could be used in future source selections,” the opinion records. ENR reported that the analyst worked under instructions that the results would not inform any real source selection, and that the Army's position was that the output “played no role” in a formal evaluation carried out by human reviewers.

TRAX's case was that the conceded weakness looked like machine error. Its suit, Defense One reported on 6 August, called it a “classic AI hallucination”, with made-up references to TRAX's proposal that no one on the evaluation board checked. The Army declined to comment to Defense One because the litigation was ongoing.

The record question became the centre of the case. According to the opinion, the government initially told the court that the evaluation board had viewed only one bidder's AI report, and that “none of the AI reports were deemed usable, they were never entered into the contract file.” A later supplemental declaration acknowledged, for the first time, that the AI summaries had been shared with three people: two evaluation board members and the Contracting Officer.

What has the court decided so far?

Judge Lerner granted TRAX's motion to add all three AI evaluations to the administrative record. The opinion states that “Agency decision makers considered all three documents during the evaluation process,” and that distribution was enough “regardless of whether the SSEB relied upon or used the outputs further.” The opinion was first filed under seal on 3 September and reissued with redactions on 22 September.

On the merits, TRAX lost. The court found the overlap between the AI outputs and the human evaluations was minimal and that TRAX had not identified actual errors traceable to AI. It denied TRAX's motion and directed judgment for the government and Southwest Range Services. Bloomberg Law reported on 23 September that the ruling lets the contract proceed.

The opinion also said the court would issue a separate order requiring the government to explain “the errors and misrepresentations in its briefing.” Bloomberg Law reported on 8 October that in an order dated 5 October the court said it needed more information to decide whether sanctions are appropriate, and that the government's response is under seal.

Why this matters beyond US procurement

This is a US ruling on a US administrative record, and it does not say how an English or European court would treat similar material. Its value for UK and European readers is the pattern, which is not specific to defence contracts: a named company is assessed by an AI system, the organisation running the system says the output did not count, and only litigation reveals who actually saw it.

Lawyers at Bradley Arant Boult Cummings, writing in The National Law Review on 6 October, put the core reasoning plainly: the court treated an agency's consideration of an AI evaluation as different from its reliance on it. That distinction is the one that matters whenever a bank, acquirer or counterparty runs an AI screen on a company or person, a practice covered in the automated screens companies never see and the diligence memo nobody wrote.

Three practical points follow for reputation lawyers, crisis advisers and corporate affairs teams:

  • Ask the question directly. Whether AI was used to assess your client, and who received the output, came out here only in a supplemental declaration filed during the litigation. If a screening, vetting or evaluation went against a client, ask in writing whether an AI system produced any part of it and who saw the result.
  • The AI output is the evidence. The dispute was not about what an AI system tends to say; it was about three specific documents, when they were generated and who received them. An account of what an AI system said that cannot be reproduced and dated is a weak foundation for any later argument.
  • A conceded error can still be harmless in law. The Army accepted that a weakness assigned to TRAX was improper, and it still made no legal difference to the award. A mistaken AI assessment of a named party may never be corrected in a way that counts. Having your own timestamped record of what was said, and when, is what keeps that question open.

Sources