How AI software drafts and verifies a tender response

Why a tender response goes through four stages

A request (a tender, a security questionnaire, a due diligence questionnaire, an auditor's request) arrives as a single block: dozens or hundreds of questions, a deadline, a signature at the end. The temptation is to hand it in one piece to a tool that reads, writes, and reviews. That is precisely what not to do.

The reason is measured. A language model that reviews its own output does not correct its reasoning errors without outside input, and its performance sometimes degrades after this self-correction (Huang et al., ICLR 2024). An AI that reviews itself confirms what it wrote. What actually reduces made-up claims is a verification mechanism separate from the draft: control questions posed and answered independently of the original drafting (Dhuliawala et al., 2023). Verification must come from somewhere other than drafting.

Hence the four stages, and their split across three actors who are never merged.

Stage 1, understand: what does AI software do with a tender question before answering it?

A misunderstood question produces a wrong answer even when the source is good. It is the most common mistake and the least visible, because the answer looks correct.

The agent that analyzes the question therefore classifies each one, before any drafting, into five types of intent: factual, exploratory, comparative, trick, contractual commitment. It also spots risky wording: absolutes ("do you guarantee total availability?"), double negatives, implicit commitments ("confirm that you will comply with any regulatory change"). The risk level detected determines the treatment applied next: a factual question calls for a citation; a trick question calls for careful rephrasing and referral to the validator; a contractual commitment calls for legal review before submission.

This first stage is the one drafting tools almost always skip. They read the question as a writing prompt. The agent reads it as a risk to be assessed.

Stage 2, draft: how does AI software draft an answer from your documents alone?

Before any request comes in, a librarian agent has read the company's documentation (policies, procedures, certificates, already-validated answers, audit reports) and structured it into a usable reference base. It is this agent that turns a stock of scattered files into a base the other agents can produce evidence from.

Drafting is then handed to two agents working together: your role-specific agent and your sector-specific agent. They draft from this reference base, and from it alone. Every answer cites its document, page, and version. An answer that cannot cite its source is not drafted: the gap is flagged, not filled with a plausible-sounding sentence.

Specialization matters as much as the source. A social-responsibility question asked in the energy sector does not call for the same answer as a quality question in food and beverage: the frameworks cited, the reader's expectations, and the vocabulary all differ. A generic AI drafts both the same way. Specialized agents draft each in its own world.

On the Optivalue.ai platform, which publishes this site, this second stage mobilizes 85 specialized agents: 72 role-specific agents, 12 sector-specific agents, and 1 librarian agent. A banking compliance question, for example, mobilizes the legal agent and the finance sector agent together.

Stage 3, verify: how does a second AI team check an answer it did not draft?

A second AI team, independent of the first, takes up every answer. It scores it from 0 to 100, checks that every cited source actually says what it is made to say, flags discrepancies between what is claimed and what is proven, and recommends the action to close each gap: which document is missing, which expert can supply it, before submission.

This independence is not an architectural nicety. It is the condition for verification to actually verify. A team reviewing its own work looks to confirm it; a team reviewing someone else's looks for the error. The mechanism has 5 layers, including 7 anti-hallucination checks, and an answer that fails one of them never reaches the validator.

When evidence is missing, the second team does not fill the gap: it declares it. And it does not stop at abstaining. Where a tool settles for saying "I don't know," it goes further: the gap is named, the right expert identified, the action to close it proposed. Stage 3 thus produces two things: a scored answer, and a list of what remains to be done for the file to hold up.

Stage 4, validate: how does a named expert validate a tender answer?

The last stage is human. Validators are assigned by domain: legal validates contractual commitments, security validates controls, finance validates terms. Every answer carries an approval status, and validation is a named, traced signature, defensible in an audit or a dispute. The whole process can be replayed: for every submitted sentence, you can trace back who drafted it, who verified it, what score it got, and who signed it.

A rubber-stamp signature would prove nothing. That is why a vigilance test accompanies this stage: deliberately wrong answers are inserted into the validation flow, and if one of them is approved, an alert is raised. The frequency of insertions, the alert recipients, and the trigger threshold are configurable per organization. It is the control that measures human attention instead of assuming it.

The concession that clarifies everything

For a draft, an internal note, a text no one will sign, a single pass is enough, and generic AI does it fast and well. The four stages serve no purpose for what commits no one.

They exist for the moment the document leaves the company and someone answers for it. A draft commits no one. Your response does. Separating the stages has a cost per question; not separating them has a cost per file: a false claim signed by an executive.

Three questions that draw the line

To find out whether a system actually applies the four stages, or just claims to:

1. Who understood the question before it was drafted, and where is that written down? If the classification of intent and risk appears nowhere, the understanding stage never happened.
2. Is the team that verifies the same one that drafted? If it is the same model reviewing its own output, the third stage is decorative. Verification must come from a mechanism outside drafting.
3. Who signs, by name, and can the process be replayed? If the answer is "the team" or "the tool," no one signs. If you cannot trace a submitted sentence back to its drafter, its verifier, and its validator, the file is not defensible.

Frequently asked questions

Do the four stages slow down the response? They add time to the question, and take it away from the file. The time lost is not in drafting; it is in the back-and-forth afterward, when an unverified claim is discovered by the customer or the auditor.

Why two AI teams instead of one more powerful AI? Because power does not replace independence. A more capable model that reviews itself remains a model reviewing itself. Separating roles is a property of structure, not of size.

What happens if the two teams disagree? Disagreement is exactly what stage 3 should produce when it applies. The answer gets a low score, the discrepancy is named, and the human validator decides with both readings in front of them. An answer with no possible disagreement was not verified.

Does the validator have to review every answer? They review what the score and the discrepancies point them to, and they sign the whole. The vigilance test guarantees that their review stays real: an approved deliberately-wrong answer triggers an alert.

What happens to a question the reference base cannot answer? It is not drafted. It is flagged as a gap, with the expert to consult and the action to take before submission. Stage 3 produces this list, and stage 4 decides what to do with it.

Written by the compliance and presales team at Optivalue.ai. Last reviewed: 5 September 2026.

Markdown version