Government Contracting

Contractors should prepare to have their next proposal rejected by a chatbot.  If you’re competing for an award under Notice ID H9240026RE001, a US Special Operations Command IDIQ, the government disclosed that it “may employ AI as a tool to assist in the analysis and review of offeror proposals,” for tasks like “summarizing proposal content, identifying compliance with request for task order proposal / solicitation requirements, and highlighting areas of potential strength, weakness, or risk.” 

The government qualifies its intent to use AI, stating that, “all final evaluation judgments and source selection decisions will be made exclusively by duly appointed Government personnel.  The Government retains sole responsibility for all inherently governmental functions and will not delegate decision-making authority to any AI tool.” (emphasis added). 

The government’s willingness to disclose its use of Generative AI (GenAI) is a step in the right direction.  But how will unsuccessful offerors know whether they were rejected by humans or a chatbot?  Anyone following the proliferation of AI in the legal context can hopefully agree that: (1) GenAI makes mistakes due to plausible errors (hallucinations) and sycophancy (telling users what they want to hear); and (2) GenAI cannot substitute human judgment.  My colleague David Timm has written extensively on common issues that arise when contractors and bid protesters attempt to use AI without proper safeguards, citing countless GAO protest filings filled with hallucinated legal and factual citations, which consume unnecessary amounts of public and private resources.  He has also identified concerns that GenAI procurement decisions might not be adequately documented under the FAR’s independent judgment requirement. 

It seems unwise, then, for the government to use GenAI tools for critical procurement functions such as “highlighting areas of potential strength, weakness, or risk.”  In best-value tradeoff procurements, these determinations are essentially the entire basis for award.  It is not difficult to imagine the potential for generative AI tools to hallucinate or otherwise mischaracterize proposal details that are material to an agency’s award decision. 

Disappointed offerors may have little recourse if they are excluded from competition based on GenAI output errors—at least under the procurement identified above—because the AI tool is apparently “an output-only system.  Any generated report data is not saved.”  (emphasis added).  How, then, can an unsuccessful offeror test the sufficiency of an AI tool that takes a first pass at proposal evaluation if the tool’s generated report data is not preserved for judicial review?  What are the “appropriate safeguards and authorizations in place” vaguely referenced in the solicitation?  Earlier this year, a protester claimed the agency “improperly relied upon artificial intelligence (AI), rather than agency evaluators, to conduct the evaluation. . .” but abandoned the argument, seemingly because it lacked information to substantiate the claims.  Salient CRGT, Inc., B-423640.2, B-423640.4 (Jan 05, 2026).  We may see more claims like this one in the coming months.

Even if the government reserves the ability to make “all final evaluation judgments and source selection decisions [ ] exclusively by duly appointed Government personnel,” how can contractors ensure that underlying GenAI tools providing the basis for final judgments and source selection decisions are being employed even handedly, without hallucinations or bias?  One way contractors can consider addressing this risk is by including pointed questions in debriefings regarding the extent to which GenAI was used to evaluate proposals.