Short answer. Split the questionnaire into atomic questions, retrieve from an approved answer library with version and expiry metadata, draft with citations, and route security, legal, pricing and SLA answers to their owners. Never auto-submit a generated answer.
The pipeline
An RFP response system has six stages: intake, question extraction, retrieval, drafting, approval and submission. Each produces a record the next stage can audit. The model belongs in extraction and drafting; approval belongs to the person who owns the claim.
- Intake: preserve the questionnaire, amendments and deadlines.
- Extraction: one atomic question per row, with section and word or format limits.
- Retrieval: find approved answers and source documents that address the question.
- Draft: write to length and format, with citations and an abstain option.
- Approval: SME, security, legal or pricing sign-off by topic.
- Submission and learning: record what was submitted, then capture win/loss and buyer feedback.
Extract atomic questions
A single cell can contain a question, a sub-question, an attachment request and a word limit. Split them. A question with two asks cannot be answered or scored cleanly.
| Question type | Retrieval source | Approver |
|---|---|---|
| Capability | Product documentation, approved answer | Solutions engineer |
| Security | Security policy, control description | Security owner |
| Legal | Current contract template, counsel guidance | Legal counsel |
| Commercial | Pricing model, deal desk | Deal desk or finance |
| Reference | Approved reference list | Customer marketing |
Give every question a stable ID derived from the section and row, not its position. Amendments move rows. Keep the original text next to the normalized text.
Build the answer library
The answer library is the product. If it is a folder of old proposals, retrieval returns stale claims with confident prose. Store each answer with answer_id, question pattern, answer text, product, version, plan, region, owner, approver, approval date, expiry date, confidentiality, source URL, source hash, supersedes and status.
Expire answers on purpose. A two-year-old security answer may describe a control that changed, a certification that lapsed or a subprocessor no longer used. Filter expired and retired answers before the model sees them. If nothing current exists, the output is “needs owner input,” not a plausible reconstruction.
Draft with citations and abstention
Retrieve first, generate second. Embed the question, search the approved library, and pass the top passages with their IDs. Require the answer, the IDs used, a confidence value and an abstain flag, then validate that every citation points to a retrieved passage.
import json, math
from openai import OpenAI
client = OpenAI(api_key="YOUR_KEY", base_url="https://api.plugsky.com/v1")
def embed(text):
return client.embeddings.create(model="plugsky-embed", input=text).data[0].embedding
def cosine(a, b):
dot = sum(x * y for x, y in zip(a, b))
return dot / (math.sqrt(sum(x*x for x in a)) * math.sqrt(sum(y*y for y in b)) or 1)
def draft_answer(question, approved_answers):
qvec = embed(question)
ranked = sorted(
((cosine(qvec, embed(a["text"])), a) for a in approved_answers),
key=lambda pair: pair[0], reverse=True,
)[:5]
context = [{"id": a["id"], "text": a["text"]} for _, a in ranked]
schema = {"answer": "string", "citations": ["answer_id"], "confidence": "number",
"abstain": "boolean", "missing": "string"}
response = client.chat.completions.create(
model="plugsky-pro",
response_format={"type": "json_object"},
messages=[
{"role": "system", "content": "Answer only from the approved passages. Cite at "
"least one ID. If the passages do not answer, set abstain true and say what is "
"missing. Never claim a certification, contract term or customer name not in them. Return JSON."},
{"role": "user", "content": json.dumps({"schema": schema, "question": question,
"passages": context})},
],
)
draft = json.loads(response.choices[0].message.content)
allowed = {p["id"] for p in context}
assert draft["abstain"] or (draft["citations"] and set(draft["citations"]) <= allowed)
return draft
The assertion blocks a citation to an unretrieved answer. It does not prove grounding; add a claim-level check and a human reviewer for high-risk answers.
SME and security review
Capability answers can go to a solutions engineer; security, legal, pricing and SLA answers need the owner of that claim. Build routing into the question record.
- Security: the owner confirms the control, report date and scope. A SOC 2 report is usually shared under an NDA; the model must not assert a certification or imply coverage the report does not support.
- Legal: counsel reviews contract terms, data processing, liability and governing law. The current template is the source.
- Commercial: deal desk owns pricing. The model may format a price, never invent one.
- References: customer marketing confirms the customer and use case may be named. A logo is not consent.
Use states: draft, needs_sme, changes_requested, approved, submitted. Record the approver, timestamp and version. If the answer changes, it returns to review. No generated answer should be submitted without a named approver.
Win-rate tracking
Win rate is easy to manipulate by choosing the denominator. Define it: won divided by decided submissions (won plus lost), excluding withdrawn, no-decision and disqualified opportunities. Segment by deal size, industry, product and renewal versus new logo.
Do not attribute a win to an answer without a comparison. Track answer reuse, edit rate and cycle time alongside win rate, and compare bids that used the library with bids that did not.
| Metric | Definition | Why it matters |
|---|---|---|
| Answer reuse | Approved answers used without substantive edit ÷ total | Is the library trusted? |
| Edit distance | Changed words ÷ drafted words, by section | Locates weak retrieval |
| Cycle time | Questionnaire received → submission approved | The operational win |
| Win rate | Won ÷ (won + lost), decided bids only | Outcome, not quality by itself |
Evaluate the system
Build a labelled set from past submissions. For each question, record the approved answer, source and whether the draft was usable. Measure groundedness, citation validity, abstention on unanswerable questions and the share of drafts a human could approve with minor edits. Evaluate security answers separately; a fluent wrong answer about encryption is worse than an abstention.
Run suggest-only for one cycle. Compare drafts with what the team submitted and review every abstention. The goal is fewer repeated SME reviews.
This is an engineering guide, not legal or compliance advice. Security claims, certifications, contract terms and data processing statements require owner approval and, where appropriate, counsel review. Verify the current status in the contract and report.
FAQ
How does RAG improve RFP responses?
It retrieves approved answers and source passages before generation, so the draft is grounded in current material and can cite the answer ID. That reduces stale claims and makes review faster, but it does not remove the need for owner approval.
Can the model answer security questionnaires?
It can draft from approved security documentation, but the security owner must confirm the control description, scope and report date. The model must not assert a certification, a coverage period or a contractual term that is not in the source.
How do I measure RFP automation ROI?
Track answer reuse, edit distance, cycle time and review hours before and after. Report win rate with a written denominator and segment it. Do not attribute wins to automation without a comparison group.
What belongs in an answer library?
The approved answer text, product and version, owner, approval date, expiry date, confidentiality, source link and status. Expired or retired answers should be filtered out before retrieval.
Should the system auto-submit responses?
No. A named approver should release each submission, with security, legal, pricing and SLA answers routed to their owners. Auto-submission turns a drafting error into a contractual representation.
Sources: the RAGAS paper on retrieval-augmented generation evaluation and the NIST AI Risk Management Framework Generative AI Profile (NIST AI 600-1). Both checked 20 November 2026. Reviewed 20 November 2026. This page is not legal or compliance advice; security claims and contract terms require owner and counsel review.