Running a competition in 2025, I saw something I had not seen before: more questions than expected, more of them about the procurement rules than the subject matter, and bidders with no visible background in public procurement quoting the directives at debrief. The obvious hypothesis was that they were being helped. The obvious way to test it was to read the clarification record across enough tenders and years to tell a trend from an anecdote.
The clarification window is the one place this can be done. Most of what a bidder writes is private, but under the Irish eTenders process every question a bidder asks and every answer the authority gives is published to all registered parties and remains on the platform after award. It is the only written, public, at-scale record of how the whole bidding market behaves.
So we built one: every tender in a panel of Irish public competitions published January to August of each year from 2023 to 2026 that carried at least two messages on its board, across six frames with software and IT services as the focus. 468 tenders qualified; 8,052 question-and-answer pairs were coded against a written codebook - what each question is about, whether it names the clause and states the bidder’s own reading, whether it seeks to change a requirement rather than understand it, whether an experienced bid professional would already have known the answer, and whether it carries visible artefacts of model drafting.
The question changed more than its number did
The headline is not volume. The median tender attracts about the same number of questions in 2026 as in 2023. What changed is the character of what is asked.
The specific question stepped up. The question that cites the clause and tells the authority how the bidder reads it went from 28% of everything asked to 34% - and in software and IT services, the frame where drafting tools have penetrated furthest, from 20% to 36%. Questions about the rules of the competition - eligibility, evaluation, pricing structure, contract terms - rose from 30% to 37%, with contract-terms questions doubling from 3% to 7%.
The naive question never came. If AI were bridging an expertise gap by bringing tendering novices into the window, the expert-knowable question - whether the deadline is Irish time, whether an ESPD is required when the RFT says so - should have multiplied. It did not: 4.7% of questions in 2023, 3.8% in 2026, flat throughout. The window is not filling with novices. It is filling with questions that read as if a lawyer or a professional bid writer drafted them.
The residue says where the drafting comes from. Questions with visible AI artefacts - stacked lettered sub-questions each proposing its own answer, markdown left in the text, citations to clause numbers that do not exist - went from roughly one in a thousand in 2023 to one in twenty-five in 2026, and 6.4% in the IT frame. The flag is deliberately conservative, so the true incidence is higher. And the flagged questions have a profile: 42% specific against 29% overall, and twice as likely to be negotiating a requirement. The model-drafted question is the expert-looking one.
It happened as a step, not a slope. On almost every measure, 2024 and 2025 sit on the 2023 level; the whole of the change arrives in 2026. That is what a threshold effect looks like - tools that existed in 2024 became good enough and adopted enough that the cost of producing a clause-cited, position-stating question fell to almost nothing. It argues against extrapolating a straight line, and for re-measuring in 2027 on the same panel.
The buyer got firmer, not better
The natural expectation is that sharper questions produce better answers. The opposite happened. Substantive answers drifted from 84% to 80%; deflection to “refer to the RFT” rose from 6% to 10%. And where a bidder asked for a requirement to be changed, the authority conceded in 42% of cases in 2023 and 19% in 2026. The 2023 pattern, in which a reasonable request produced a corrigendum, has largely gone - replaced by hard cut-offs, refusals of late questions, and “the Authority is satisfied that the documents are clear”.
Meanwhile volume moved to the tails. One tender in three now carries twenty or more questions, against one in five in 2023; four 2026 tenders exceeded a hundred. A minority of competitions have become the venue for a documented, pre-award negotiation of the rules - a bidder citing section 3.2.A, invoking proportionality against a turnover floor several multiples of the contract estimate, and proposing the alternative it would like accepted; another tabling a liability cap as a multiple of annual licence fee, before award, in a forum every competitor can read.
What it means
For authorities: the window is no longer only where ambiguities get resolved. It is where contract positions are tested on the record, and the marginal cost of one more precisely worded question has fallen to near zero. Documents drafted on the assumption that only a few specialists will read the contract schedule closely need a different assumption - and deflecting a question about a genuinely unclear term is a debrief argument waiting to happen.
For bidders: the tools work - specific questions get substantive answers, and the concessions that remain go to specific, proportionate asks. But a specific question no longer distinguishes you from the field. The advantage has moved from being able to write the question to knowing which question is worth asking.
One honest caveat: the coding was done by a model reader from a written codebook, checked against a hand-coded pilot that agrees on which questions exist and on the direction of the shift but applies the specificity code more liberally. Treat the figures as a single-coder result until the blind human re-code, which is the next step. The panel, codebook and coded dataset are held by BidReview and available for replication.
The takeaway
The naive question did not rise. The expert-looking question did. That combination is hard to explain by anything other than a fall in the cost of expertise - and it is the signature to watch in every channel where bidders write to buyers.
The same forensic read, applied to your bid before you submit. BidReview audits finished tender responses against the buyer’s own criteria - the evaluation-side counterpart of the analysis above. Run the free scorecard →
