Every business development lead who has chased federal work knows the ritual. You lose, you request a debrief, and three weeks later you get a letter with all the warmth of a parking ticket: your proposal was evaluated as acceptable, award was made to the offeror representing the best value, no further information will be provided. The whole team crowds around that paragraph like it is a locked safe, trying to reverse engineer which line cost them the job, half suspecting the real answer was a coin flip nobody let them watch.

That black box just got a very specific new suspect. A contractor is now telling a federal court that the thing grading its bid may have made part of its reasoning up.

TRAX International Corporation filed a complaint in the U.S. Court of Federal Claims in late July, asking a judge to force the Army to redo its evaluation of a roughly $449.4 million mission support services contract at White Sands Missile Range, which the Army awarded to a competitor, Southwest Range Services. TRAX had already lost a bid protest over the same award at the Government Accountability Office, which denied the protest on May 14 even after the Army admitted it had assigned TRAX an erroneous "weakness" during evaluation. GAO called it harmless error and moved on. TRAX's new lawsuit argues it was not harmless, and that it was not really a human error either: the complaint describes the weakness as "a classic AI hallucination, with made-up references to TRAX's proposal, that no one on the Source Selection Evaluation Board checked." TRAX also says the Army would not clearly confirm whether some of the strengths credited to the winning bid came from a human evaluator or from the same tool. Given the roughly $29.4 million price premium in Southwest Range's proposal, TRAX argues that erasing one unsupported weakness could plausibly have flipped who got the award.

Engineering News-Record reported on court filings made public July 31 that name the tool in question: an internal Army system called FAST TRACK, built to speed up source selection by crosswalking solicitations against proposals and flagging strengths and weaknesses for evaluators. The Army's position, per those filings, is that the Source Selection Authority who actually made the award decision did not rely on FAST TRACK's output, so whatever the tool got wrong never touched the final call. The case is still active. As of the same July 31 filings, TRAX, the government, and Southwest Range had all filed competing motions asking the court to rule on the existing record, and a judge had not yet decided the merits. Southwest Range's filing noted, pointedly, that TRAX had not actually pressed the hallucination claim in its own motion, which suggests this fight is far from settled even on TRAX's side of the table.

This is not the first time an agency has faced this exact accusation. Federal News Network's coverage of the TRAX case notes that Salient CRGT previously alleged a different Defense Department sub agency used AI instead of human evaluators to grade its bid, a claim GAO treated as abandoned and dismissed in a decision dated January 5. Two protests in eight months raising the same underlying question, whatever machinery is reading your proposal, is a pattern, not a fluke, and it lands in a genuine regulatory gap. There is currently no federal acquisition rule that specifically governs how, or whether, agencies can lean on AI during bid evaluation, which means the disclosure obligations TRAX is trying to establish through litigation do not yet exist as a matter of written policy anywhere else.

Set this next to what has otherwise been a good year for AI on the bid preparation side of construction and government contracting, tools that read solicitations, draft compliance matrices, and flag scope gaps faster than any human team could manage on a compressed timeline. TRAX's lawsuit is the mirror image of that story. It is a reminder that the evaluation side of a bid can now run on the same kind of model doing the writing, with the same failure mode: a confident, plausible sounding fabrication that nobody in the room happened to check. For a company built around agentic AI that helps contractors respond to RFPs, that is not an abstract concern, it is the whole argument for building in traceability from the start, a citation back to the actual document behind every claim, on both sides of the desk. If a machine is going to help decide who wins a nine figure contract, the industry is going to need a much better answer than a three sentence letter for how it made up its mind.

Sources