Buyers are starting to evaluate tenders with AI: what that changes about your returnables
Your last submission answered a schedule with "refer to Section 4.2 of our technical proposal". A human evaluator would have turned the page. Would a machine?
Buyers are increasingly using AI assistance to read and compare submissions, usually to extract each bidder's answers against the evaluation criteria before a person scores them. That does not change what wins a tender. It changes how harshly a submission is punished for putting the answer somewhere other than where the question was asked.
The fix is the oldest advice in tendering, enforced more strictly than before: answer in the box, restate the question, and never make the evaluator go looking. Checking a submission for answers that are not where their question is, is work Elora Grid does; what those answers say stays yours.
What does it mean for a tender to be evaluated with AI?
In practice it means an assisted first pass. The buyer's tool extracts each bidder's response against each criterion, checks whether required documents are present, and lays the bidders side by side so an evaluation panel can score from a structured summary rather than from a stack of PDFs.
The panel still scores. What has changed is the raw material they score from: not your document, but what a machine managed to pull out of it.
What does machine reading handle badly?
Anything that requires a reader to go somewhere else. A cross-reference in place of an answer. A response given in the covering letter rather than in the schedule that asked for it. An attachment that answers three questions at once, with no signpost to which part answers which.
Then there is the mechanical layer: scanned pages with no text underneath, tables split across page breaks, answers inside images, and locked or flattened forms. A human evaluator squints and copes. An extraction step returns an empty field.
How do you make a submission survive it?
Put the answer where the question is, and make it complete on its own. Restate the question in your response so the text stands alone if it is lifted out. If an attachment carries the detail, summarise the answer in the field and cite the attachment as evidence rather than as a substitute.
Keep the text selectable. Export to PDF rather than scanning, keep tables inside one page where you can, and check that the numbers in your schedules are text rather than pictures of text.
Does this change what wins?
No, and that is the reassuring part. Conformance, clarity and answering the question asked have always won tenders. Machine-assisted evaluation punishes the same faults, just more consistently and without the benefit of the doubt a tired human sometimes gives you.
Which makes this a change with no downside: a submission that survives extraction is also easier for a person to score. Elora Grid checks the pack the way an extraction step would, listing every field where the answer is a pointer, is missing, or sits somewhere other than where the question was asked. Rewriting the answers stays with your bid team.
What breaks extraction, and the fix
| What you did | What the panel receives | The fix |
|---|---|---|
| Answered with a cross-reference | A pointer where an answer should be | Answer in the field; cite the attachment as evidence, not as the answer |
| Answered in the covering letter | An empty schedule field | Put every response in the schedule that asked for it |
| Scanned or flattened pages | No extractable text at all | Export to PDF from the source file; keep text selectable |
| Tables split across pages | Half a table, or rows attributed to the wrong heading | Keep a table on one page, and repeat the header row |
| One attachment answering several questions | The same block returned against every question, or none | Signpost by question number inside the attachment |
- 01Read every returnable field on its own. Cover the rest of the submission and ask whether that field alone answers the question.
- 02Replace every cross-reference used as an answer. Summarise in the field, then point to the attachment as supporting evidence.
- 03Restate the question inside the answer. The response should stand alone when it is lifted out of your document.
- 04Check the text is text. Search the final PDF for a phrase from each schedule; if it is not found, the page is an image.
- 05Confirm every required document is attached and named as instructed. Presence checks are the easiest thing to automate, and the easiest to fail.
Common questions
Are tenders actually scored by AI now?
The scoring decision sits with the evaluation panel. What is increasingly assisted is the pass before it: extracting each bidder's answers against the criteria and checking that required documents are present. Assume your submission will be read that way, because the practices that survive it are good practice regardless.
Should you write differently for an AI reader?
Not differently, just more strictly. Answer in the field, restate the question, keep each response complete on its own, and stop using cross-references as answers. Every one of those also helps a human evaluator scoring forty submissions in a week.
Is a cross-reference ever acceptable?
As supporting evidence, yes, but never as the answer itself. Give the substantive response in the field the question was asked in, then point to the drawing, certificate or schedule that proves it. An evaluator working through forty submissions will not go looking for it.
Can Elora Grid check a submission for this?
Yes. Hand it your draft submission and the issued returnables, and it returns every field where the answer is missing, is a pointer to another document, or sits somewhere other than where the question was asked, each cited to its source page. The rewriting stays with your bid team.