Why AI provenance matters for Pacific-aid bid managers
In 2025, the US Government Accountability Office dismissed more than twenty AI-assisted bid protests. The common thread across all dismissals: the AI system had cited clauses that did not appear anywhere in the source document. The legal term for this in procurement law is fabrication. The practical consequence is disqualification, potential debarment proceedings, and in some cases, reputational damage that persists across donor-funded work.
This is not a theoretical risk. It is happening now, at scale, in the same competitive landscape where Pacific-aid infrastructure firms bid.
The problem with standard LLM summarisation
When a bid manager asks a general-purpose LLM to summarise a 200-page tender pack, the model produces fluent, plausible text. The evaluation weights sound reasonable. The eligibility clauses sound familiar. The submission mechanics sound standard. And approximately 15–25% of the specific figures, references, and clause numbers are invented — because the model is completing a plausible document, not retrieving verified facts.
This is not a flaw that will be fixed in the next model version. It is a fundamental property of autoregressive text generation. The model does not know what it does not know.
What citation-anchored extraction looks like
TenderTism’s AI Capture Brief takes a different approach. Instead of asking the model to summarise the pack, the system:
- Ingests every document in the pack — preserving paragraph boundaries, page references, and document structure.
- Runs targeted extraction with a strict schema that requires a source reference for every claim.
- Computes a confidence score from retrieval margin, self-consistency, and citation density.
- Renders every claim as a clickable citation — opening the verbatim source paragraph when clicked.
The model cannot emit a claim without a source reference, because the system prompt requires the reference field to be populated from the retrieval result. If the retrieval margin is below threshold, the claim is either withheld or flagged with “low confidence — verify before acting.”
The result is a brief that a capture manager can hand to their legal team without anxiety. Every figure has a page number. Every eligibility clause has a paragraph reference. Every claim can be verified in thirty seconds by clicking through to the pack viewer.
The GAO precedent matters for Pacific-aid work
Pacific-aid infrastructure procurement follows FIDIC and ADB/World Bank procurement guidelines — not US Federal Acquisition Regulation. But the reputational and legal risk of submitting an AI-assisted bid with fabricated citations is identical. Procurement agencies at ADB, World Bank, and DFAT-funded programmes have access to the same legal frameworks for disqualification.
More practically: the firms competing for Pacific-aid infrastructure contracts are small. Twenty firms are competing for the same forty tenders in any given year. Reputational damage from an AI citation scandal propagates quickly through a tight professional network.
What this means for your capture workflow
The practical implication is straightforward: if you are using AI for any part of your capture analysis — bid/no-bid decisions, evaluation-weight extraction, eligibility checking, competitive positioning — you need a tool that can show its work.
“The AI said so” is not a citation. A verbatim paragraph from page 34 of the Request for Proposals, with the source sentence highlighted and the document reference shown, is a citation.
TenderTism is built around this requirement. The Citation component in our UI is not a design choice — it is the load-bearing structural constraint around which every other feature is built.
TenderTism is in private beta. Contact us to join the early-access programme for Pacific-aid infrastructure firms.