Proposal & RFP Intelligence Agent

The Problem

• Proposals and RFP responses written from scratch each time — 8–20 hours per submission depending on complexity
• No structured reference to past proposals — writers rely on memory or dig through shared drives
• Won and lost proposals stored but never analysed — the pattern of what works is invisible
• RFPs contain client-specific requirements that must be cross-referenced against real org capabilities, but this is done manually and inconsistently
• Facts, figures, and claims in proposals (team size, past project values, delivery timelines) not systematically verified before submission
• No feedback loop: when a proposal is lost, the reasons rarely make it back to improve the next one
• Senior time is consumed by formatting, boilerplate, and research — not the high-judgment positioning work that actually wins

What Is Being Built

Agent 1 — The Intake Agent (brief analysis and scoping)

Reads the incoming RFP or brief and produces a structured intelligence pack before any writing begins.

• Receives the RFP document or brief via email, form submission, or file upload
• Extracts and structures: client name, sector, project type, scope, budget signals, deadline, evaluation criteria, and any mandatory requirements or exclusions
• Runs a web search to enrich context: client background, recent news, known priorities, any public statements about the initiative
• Searches the internal proposal archive for past work with the same client, same sector, or same project type — surfaces win/loss history and relevant case studies
• Produces a "Brief Intelligence Pack": a structured summary of what is known, what is unknown, and what the proposal must address to score well
• Flags any red flags: misaligned scope, unrealistic timelines, requirements the org cannot meet
• Human checkpoint: proposal lead reviews and approves the intelligence pack before drafting begins

Agent 2 — The Researcher (live and archival intelligence)

Fills evidence gaps identified in the intelligence pack — pulling from the web and the internal library.

• Receives the approved intelligence pack from Agent 1
• Runs targeted web searches on: client organisation, sector trends, competitor positioning, relevant regulation, comparable projects, pricing benchmarks
• Searches the internal knowledge base for relevant past project summaries, team bios, capability statements, methodology docs, approved statistics
• Cross-references any data points with sources — flags anything that cannot be verified
• Produces a sourced evidence pack: every fact tagged with its source and confidence level (verified / inferred / assumed)
• Human checkpoint: proposal lead approves the evidence pack; assumed data points must be verified or excluded

Agent 3 — The Writer (proposal and RFP drafting)

Assembles the proposal from the approved brief pack, evidence pack, and relevant past proposals — following the org's established voice and structure.

• Retrieves the approved intelligence pack and evidence pack
• Searches the won proposals archive for the closest structural match — uses it as a reference, not a template to copy
• Selects the correct proposal template (standard client proposal vs formal RFP response vs capability statement)
• Drafts each section using only the approved evidence pack — no hallucinated figures, no unverified claims
• Inserts placeholders where information is still missing rather than inventing content
• Mirrors the language and framing patterns from the org's highest-scoring past proposals
• Runs a self-check: are all evaluation criteria addressed? Any placeholder texts remaining? All figures traceable to the evidence pack?
• Human checkpoint: full draft reviewed by proposal lead and at least one subject-matter expert

Agent 4 — The Fact-Checker (verification and risk review)

Independently audits every factual claim in the draft before it leaves the building.

• Receives the full proposal draft
• Extracts every specific claim: named past projects, project values, team credentials, delivery timelines, statistics, quoted figures
• Cross-references each claim against the approved evidence pack, internal knowledge base, and (where appropriate) public sources
• Produces a verification report: each claim marked Verified, Needs review, or Remove
• Flags any claim where the draft wording is stronger than the evidence supports
• Checks internal consistency: do figures in the executive summary match the body? Do team sizes and project values align across sections?
• Human checkpoint: proposal lead resolves all flags before submission — nothing proceeds with unresolved red flags

Agent 5 — The Memory Agent (win/loss learning)

Closes the feedback loop — updating the proposal knowledge base with outcomes and extracting patterns over time.

• When a proposal outcome is known (won / lost / shortlisted / no decision), it is logged against the original record
• Where feedback is available (debrief notes, client comments, evaluator scores), the agent extracts structured learnings
• Enriches the proposal archive: each past proposal is tagged with outcome, sector, project type, value, win/loss reason, and differentiators
• Monthly cadence: produces a "Proposal Intelligence Report" (win rate by sector, win rate by project type, common themes in wins, common objections in losses)
• When Agent 3 drafts a new proposal, this agent surfaces the most relevant winning proposals as reference — prioritising recent wins in the same sector
• Human checkpoint: monthly report reviewed by business development and leadership; insights feed back into the playbook

Agent Flow

RFP / brief received
        |
   Agent 1: Intake
   (brief analysis + archive search)
        |
   [Human checkpoint: approve intelligence pack]
        |
   Agent 2: Researcher
   (web + internal evidence gathering)
        |
   [Human checkpoint: approve evidence pack]
        |
   Agent 3: Writer
   (drafts from approved evidence only)
        |
   [Human checkpoint: proposal lead + SME review]
        |
   Agent 4: Fact-Checker
   (verifies every claim independently)
        |
   [Human checkpoint: resolve all flags before submission]
        |
   Submission
        |
   Outcome logged -> Agent 5: Memory
   (win/loss learning + monthly intelligence report)

Projected Outcomes

• Proposal first draft time: from 8–20 hours to 2–4 hours (agent drafts, human refines)
• Research phase: from 3–5 hours manual to 30–45 minutes (enriched by web search + archive)
• Fact-checking: from ad-hoc or skipped entirely to 100% of claims verified before submission
• Win/loss pattern visibility: from zero to a monthly intelligence report updated after every outcome
• Institutional memory: every proposal, outcome, and learning retained and searchable — survives staff turnover
• RFP response quality: proposals reference real winning patterns rather than starting from a blank page

Key Metrics

Next-Level Improvements

• Scoring simulation: model how the proposal scores against stated evaluation criteria before submission
• Competitor awareness: flag if a known competitor has recently won work with the same client
• Pricing intelligence: surface past project values and market benchmarks to inform fee proposals
• RFP no-bid scoring: agent assesses fit and win probability before the team commits time — recommends bid or pass with reasoning
• Multi-language support: draft proposals in the client's primary language with a translation review checkpoint
• Client relationship history: cross-reference CRM data — flag any prior unresolved issues before the team submits

Human-in-Loop Design Principles

This system is explicitly designed so that humans control every decision that matters.

• No agent submits or sends anything — every output routes to a human before action
• No agent invents facts — all figures and claims must trace to the verified evidence pack or be flagged as placeholders
• No agent overwrites the archive — the Memory agent appends and tags; it never deletes prior records
• Fact-Checker is independent — it does not see the Writer's sources, only the draft, so it cannot be primed to approve what the Writer assumed
• All web search is sourced — every fact pulled from the web is tagged with its origin and surfaced for human review before use

This case study is based on a conceptual architecture developed for a professional services organisation. Human-in-loop checkpoints are non-negotiable design constraints, not optional features — the system is built on the assumption that final judgement on every client-facing output remains with a human.