Somewhere in the last eighteen months, procurement quietly crossed a line. Buyers stopped asking whether suppliers use AI to write bids. They started asking whether the bid survives checking. Two companies have already learned the difference the hard way, one court has blessed AI-assisted bidding outright, and the UK has written the new expectations into official guidance. If you answer tenders for a living, these four documents are worth ten minutes of your time.
Poland: excluded for tax rulings that never existed
In October 2025, the Polish road-maintenance contractor Exdrog was thrown out of a public tender after the buyer noticed something strange in its filing: the bid cited tax rulings that had never been issued. The National Appeals Chamber upheld the exclusion. The document read fluently, the legal references were formatted correctly, and the rulings simply did not exist. A classic large language model hallucination, submitted under a company's name in a regulated procedure, treated by the tribunal as misleading the contracting authority.
The lesson is not that AI wrote the bid. The lesson is that nobody checked the citations before a competitor and a tribunal did. Full report at Notes from Poland.
United States: GAO turns fabricated citations into a sanction
A month earlier, the US Government Accountability Office went further. In Oready, LLC (B-423649, September 25, 2025), GAO dismissed a contractor's protests as a sanction after finding AI-fabricated citations in its filings. Not "please refile". Dismissed, as a penalty. Legal commentators tracking the trend counted roughly thirty AI-misuse filings across 2025 procurement disputes, and GAO's message in Oready was that fabricated authority is now treated as an integrity problem, not a formatting slip. Analysis at Burr & Forman.
Italy: the court that said yes
Before concluding that AI in bids is radioactive, read the Italian counterexample. Italy's Council of State upheld a contract award where the winning bidder had used ChatGPT to help prepare its submission. The losing side argued the AI use itself should void the award. The court disagreed: AI can legitimately enter the competition. What matters is the content of the bid, verified the ordinary way. Bird & Bird's summary calls it what it is: a tool, not a free pass.
Put the three rulings together and the pattern is clean. No tribunal has punished a bidder for using AI. Two have punished bidders for submitting unverified AI output. One has protected a bidder who used AI and stood behind every word.
The UK writes it into the rulebook
The UK has now formalised the buyer's side of this bargain. PPN 017, the procurement policy note that took effect with the Procurement Act regime in February 2025, encourages contracting authorities to ask suppliers whether AI was used in preparing a bid. Disclosure itself is unscored, and the guidance is explicit that using AI is not a reason to mark a bid down. But the same note warns authorities, in as many words, that AI tools may introduce "an increased risk of misleading statements via hallucination", and tells them to undertake due diligence on the accuracy, robustness and credibility of bids. Source: gov.uk, PPN 017.
Read that as a bidder and the operating reality is simple. You can use AI freely. You should expect your bid to be checked more carefully than it was in 2023, precisely because buyers now assume AI was involved. The verification burden has moved onto you, before submission, where it is cheap, instead of after award, where it is fatal.
What careful teams do now
The teams getting this right have converged on the same few habits.
- Every factual claim gets a source. If a sentence asserts a certification, a case study, a statistic or a legal reference, someone can point to the document it came from. If nobody can, the sentence goes.
- Requirements are tracked in a matrix, not in memory. A tender hides dozens of mandatory requirements across hundreds of pages. Winning teams work from an extracted list with a page reference for each one, so review means checking rows, not rereading the pack.
- Coverage is checked before submission. The question "does our response actually answer requirement 34" is answered by inspection, not by confidence.
- AI output is treated as a draft with a chain of custody. Who generated it, from what sources, and who verified it. That is the audit trail PPN 017 quietly expects.
This is also, transparently, how we build Lucius. The platform extracts every requirement from a tender into a compliance matrix with a page reference for each clause, marks which extracted clauses were verified verbatim against the source document, checks whether your draft addresses each requirement, and exports the whole record as a single auditable archive. You can run a free analysis on a live tender and see the page-cited matrix on your own document, or read more about AI bid management done compliance-first.
Frequently asked questions
Is it allowed to use AI to write tender responses?
Yes, in every jurisdiction discussed here. Italy's Council of State has upheld an award won with AI assistance, and UK guidance states that AI use is not itself a reason to penalise a bid. What gets bidders excluded is submitting AI output containing false statements, such as citations to documents that do not exist.
Do I have to disclose AI use in UK public tenders?
PPN 017 encourages contracting authorities to ask, typically through a disclosure question in the tender pack. The disclosure is unscored. Answer it honestly and expect the authority to look harder at the substance of your bid, because the same guidance tells them to verify accuracy and credibility.
What actually gets AI-assisted bids disqualified?
Unverifiable content. The two public casualties to date, a Polish contractor and a US protester, were both punished for fabricated citations that nobody checked before submission. A bid where every claim traces to a real source document carries no special AI risk at all.
Get help with your bid