In a clarifications register published on eTenders this year, nine consecutive queries from one bidder each open with the literal string “| Lot 1 | Service Category 1 | Building Surveying Clarifications” - a pipe-delimited header carried over from whatever table or template generated the batch, pasted into the portal with the scaffolding still attached. Each of the nine is built on the same syntactic skeleton with one noun swapped. One bidder, one session, one template - and the generation process visible to every registered party, including the buyer, in a published competition document.
That is drafting residue: the artefacts of machine production left in a submitted document. The clarification window is where it is publicly visible. In the bids we audit before submission, the same class of evidence turns up in the submissions themselves - and a bid, unlike a clarification, is being read for marks.
What residue looks like in a bid
Assistant scaffolding. The model’s conversational frame left in body text: “Yes, here’s a clean draft of this section”, “I’ve pulled this from our other tenders”, “As requested, the response below addresses all three criteria”. These read as absurd in isolation. In a 40-page document assembled from a dozen drafting sessions under deadline pressure, they survive to submission more often than anyone would like to believe.
Unfilled placeholders. [Client name]. [Insert figure]. [Add case study here]. The model structured the answer and flagged what it needed; nobody supplied it; the brackets shipped.
The wrong client. Boilerplate that names another buyer, another county, another framework - sometimes a competitor’s client. Beyond the immediate embarrassment, this one has teeth: content built for a different competition is, by construction, not answering this buyer’s question, and where it asserts something false about this buyer it damages the credibility of the whole submission.
Formatting artefacts. Markdown headers (##), bold markers (**), stray pipes and table fragments - the syntax of the drafting environment surviving the paste into Word or the portal.
Phantom citations. References to clause numbers, appendices or documents that do not exist in this RFT - the model citing the structure of some other tender pack it has internalised.
What residue tells the evaluator
Most rubrics have no line item for residue, and an evaluator will rarely deduct marks for a stray bracket as such. What residue does is worse than a deduction. It tells the evaluator, with certainty, that this passage was not read by its author before it was submitted - and it licenses a question about every other claim in the document: if nobody read this, who verified that?
For what it is worth, the same tell is spreading on the other side of the process. Our study of 8,052 clarification questions on Irish tenders found questions with visible AI artefacts rising from roughly one in a thousand in 2023 to one in twenty-five in 2026. The tools are everywhere. The reading is not.
The purge, in ten minutes
Residue is the most mechanically detectable of all AI bid writing risks, which makes shipping it inexcusable. Before any human read-through, run document searches for: square brackets; “here’s” and “I’ve”; “as an AI” and “as requested”; ## and **; the pipe character; every client, county and framework name in your boilerplate library that is not this buyer; and every clause and appendix number cited, checked against the actual pack. Then do the human read - because the searches catch the artefacts, and only a full pre-submission check catches what the artefacts imply.
The takeaway
Residue never wins or loses marks directly. It answers a different question - did anyone read this before submitting it? - and in a competition decided on confidence, that answer follows the evaluator through every page that comes after.
Residue, phantom citations and wrong-client boilerplate - caught before submission. BidReview audits your finished bid the way an evaluator reads it. You write; we check. Run the free scorecard →
