How do you know what your AI agent did not see?
In brief.
When a company entrusts an AI agent with analysing its tenders (reading the client’s file, extracting every requirement, preparing the answers), one question outweighs all the others: how do you know what it did not see? Text generation is visible: you read it, you correct it. Failed extraction is invisible: the list produced looks complete, and nothing points to the requirement missed on page 87 of an annex, the one discovered at bid opening, in the form of a disqualification.
The asymmetry between the visible and the invisible
When an agent writes badly, everyone notices and corrects it. When it leaves something out, nobody notices anything: the output is clean, coherent, convincing. Trust sets in precisely because the most serious errors are the ones you cannot see. And tender documents are the worst possible terrain: requirements scattered across the tender rules, the specifications and the annexes, legal wording, cross-references, tables, scanned documents.
The only serious answer: measure coverage
Extraction coverage is measured by comparing, on reference files annotated by humans, the list of requirements actually present with the list the system identified. Without this test bench, any claim of completeness is an act of faith. The question to ask any team, in-house or vendor: “on which set of files do you measure your coverage, and what is your score?”. No answer is an answer. The organisational rule that follows: as long as coverage is not measured, a human has to redo the extraction in parallel, which cancels out the point of the agent.
The Optivalue.ai approach
Optivalue.ai produces a requirements matrix traceable down to the page number and treats coverage as a compliance object: what is covered, what is not, what remains to be examined. Completeness is not a promise there; it is a verifiable deliverable.
Why is an omission more serious than an error?
Because an error is corrected at review; an omission goes unseen until the penalty.
What is extraction coverage?
The proportion of the requirements actually present in the file that the system did identify.