In August 2026, a US federal contractor sued over a $450 million Army award, alleging that a weakness assigned to its proposal in evaluation was a classic AI hallucination - made-up references to its proposal that no member of the evaluation board checked before the report went up the line. The AI causation is alleged and contested; the evaluation error itself was already established by the Government Accountability Office. Whatever the courts decide, the case names the mechanism precisely: a model filled a gap with something plausible, and nobody checked it before it entered a formal procurement record.
That mechanism does not care which side of the table it operates on. In the bids we audit, it operates on yours.
Where models invent things in bids
A language model asked to strengthen an answer will strengthen it - with whatever the answer seems to need. In tender responses, the fabrications cluster in four places.
Performance figures. “Reduced processing time by 38%”, “99.7% first-time fix rate”, “delivered 12% under budget” - generated because the sentence pattern of a strong answer contains a number, not because your delivery records do. These are the most dangerous inventions, because they are exactly what evaluators are trained to reward, and exactly what reference checks and contract-delivery reality later expose.
Contract references and case studies. A named client, a plausible project, a tidy outcome - assembled from the texture of your real work but corresponding to no contract you held. In both Ireland and the UK, serious misrepresentation is a discretionary exclusion ground; a reference the buyer cannot verify because it never happened is disqualification-grade material.
Facts about the buyer. Models pattern-fill about the contracting authority too - a fleet size, a service structure, a strategy the buyer never published. A factual error about the buyer’s own organisation is the single most corrosive mistake a bid can carry: the person marking your answer is the person who knows it is wrong, and their confidence in every adjacent claim drops with it.
Certifications and standards. ISO numbers, accreditations and framework memberships stated because bids like yours usually hold them. If it is checked at bid date - and certifications are - the gap is not a style issue.
The verification pass
The countermeasure is unglamorous and non-negotiable: a claims audit of the finished draft, before the rest of the pre-submission check.
Extract every factual claim in the bid - every figure, date, named client, contract value, certification, and statement about the buyer - into a list. For each, record the source document that proves it: the KPI report, the signed contract, the certificate, the buyer’s published strategy. A claim with a source stays, cited precisely. A claim without a source is either replaced with one you can prove or deleted. There is no third category; “probably roughly right” is how invented figures reach evaluators.
Two rules make the pass stick. The person verifying should not be the person who prompted the draft - authors inherit the model’s confidence. And the standard is external: could this claim survive the buyer ringing the referee, checking the certificate register, or reading their own strategy? In a market where AI-assisted scrutiny is rising on every side, assume every checkable claim will eventually be checked.
The takeaway
Models fill gaps with plausibility; procurement punishes plausibility that fails verification. The fix costs an afternoon: list every claim, trace every claim, and submit nothing a named person cannot stand behind with a document in hand.
Every unverifiable claim, flagged before the buyer finds it. BidReview audits your finished bid for proof-point gaps, phantom references and credibility risks against the RFT. Run the free scorecard →
