Skip to content

Reference: RFP / Proposal Response

Shape: supervisor-routing with parallel specialists. Primary ROI lever: response time and win rate. Variant of: the flagship. Scaffolding reference — adapt the flagship src/ to this scenario via /scaffold-from-brief.

When to use this

Customer responds to complex RFPs / RFIs / security questionnaires. Today: a cross-functional scramble (sales, pricing, legal, security, technical) that takes days per response and recycles inconsistent boilerplate.

A chatbot can't help meaningfully here — the work is composing a multi-section document grounded in customer-specific policy, past-proposal content, and current product capability. An agentic solution with specialist workers aggregates correctly-grounded sections in parallel, with HITL on legal and pricing sections.

Chatbot baseline vs agentic lift

Capability Chatbot Agent + HITL
Surface past Q&A ✓ (poorly) ✓ (grounded retrieval over past proposals)
Draft a pricing response ✗ ✓ specialist worker w/ HITL
Draft a legal/contractual response ✗ ✓ specialist worker w/ HITL
Draft a technical/security response ✗ ✓ specialist worker
Assemble multi-section proposal ✗ ✓ aggregator
Write to proposal DMS ✗ ✓ with HITL

ROI lever: response time (days → hours), win rate lift, deal-team hours saved per RFP.

Proposed agent graph

  Supervisor (RFP intake + section routing)
  ├── PricingWorker     — prices against rate card + approved discounts
  ├── LegalWorker       — answers from policy + past contracts, flags deviations
  ├── TechnicalWorker   — answers technical/architecture Qs w/ citations
  ├── SecurityQWorker   — answers security questionnaires from policy corpus
  └── ReferenceWorker   — pulls case studies + references
  Aggregator            — assembles full proposal draft + coverage report

Side-effect tools

  • dms_upload_draft — proposal DMS (SharePoint / custom) — HITL: always before writing final
  • flag_deviation — creates a review task when Legal/Pricing detects deviation from policy
  • request_sme_input — pages a human SME for a gap the agents can't fill

Grounding

  • Past-proposal corpus (AI Search index)
  • Policy corpus: pricing guidelines, legal positions, approved boilerplate (AI Search index)
  • Product docs (AI Search index)
  • Security controls library (AI Search index; source of truth for SecurityQWorker)

HITL policy

  • always: dms_upload_draft (nothing auto-published)
  • always: pricing sections' approval before aggregator includes them
  • always: legal/contractual deviations flagged for legal review before inclusion
  • threshold: technical/security sections with confidence < 0.85 trigger SME review

Success criteria to fill in

  • Time to first complete draft (current → target, in hours)
  • Coverage: % of RFP questions answered on first pass (target ≥ 95%)
  • Consistency: % of responses that match approved boilerplate (target ≥ 90%)
  • Win-rate lift on RFPs scored above quality threshold (leading indicator)

RAI risks specific to this scenario

  1. Agent makes up a technical capability the product doesn't have → mitigated by grounded retrieval + validate.py rejecting unsupported claims.
  2. Agent quotes pricing outside approved rate card → mitigated by PricingWorker HITL + schema-validated rate card input.
  3. Agent answers a legal question that commits the company → mitigated by LegalWorker HITL-always.
  4. Agent leaks sensitive past-proposal content between customers → mitigated by tenancy-scoped index + redteam case.
  5. Agent over-claims security controls not in effect → mitigated by SecurityQWorker sourcing from controls library only.

KPIs to instrument (events)

  • rfp.received
  • rfp.section.drafted (per section type)
  • rfp.section.coverage_met
  • rfp.hitl_pricing_approved / rfp.hitl_pricing_revised
  • rfp.hitl_legal_approved / rfp.hitl_legal_revised
  • rfp.draft_uploaded
  • rfp.time_to_first_draft_ms

How to scaffold from this reference

  1. Paste relevant sections into docs/discovery/solution-brief.md.
  2. /scaffold-from-brief.
  3. /add-worker-agent for each of Pricing, Legal, Technical, SecurityQ, Reference (or rename existing workers).
  4. /add-tool for dms_upload_draft, flag_deviation, request_sme_input.
  5. Build the past-proposal index in infra/modules/ai-search.bicep with proper row-level security.
  6. Write 20+ golden cases = real past RFPs with expert-reviewed answers.