RFP Response Drafting from a Reuse Library

Tier: Core · Hands-on time: ~30 minutes

RFP response is the services architect's highest-pressure writing: 2026 RFPs carry GenAI sections that must be answered credibly in days, and the raw material — win themes, past solutions, delivery evidence — is scattered across a decade of documents. RAG over the proposal library turns that scatter into drafting speed. The taught pitfall is structural: a reuse library contains your firm's most optimistic sentences, each written for a context that no longer applies, and a model will confidently transplant a 2024 claim ("we have delivered 40+ Kubernetes migrations") into a 2026 bid where it is stale, unverifiable, or contractually dangerous. The claims audit is not a polish step; it is the playbook.

When to use

• •
  An RFP or RFI response is due and substantial sections overlap with past bids — the normal case.
• •
  The GenAI section of a client RFP needs a credible answer assembled from real delivery evidence, fast.
• •
  Win themes must stay consistent across a multi-section response written by several authors under deadline.

Skip AI when the RFP section asks for something your firm has never done — retrieval over a library that lacks the answer produces confident pastiche; write that section from an honest capability conversation instead.

Prerequisites

• The reuse corpus, curated and clearedpast proposals, case studies, reference architectures, CVs/capability statements; scrubbed of other clients' confidential details before indexing (a past client's name and metrics in a new bid is a breach, not a win theme).
• Retrieval toolinga Claude Project loaded with the corpus for small libraries; pb-18-style RAG (LLM API + pgvector) once the library outgrows a context window.
• The new RFP, decomposedrequirements extracted (pb-01/pb-03 discipline applied to the RFP itself: mandatory response format, evaluation criteria, compliance asks).
• A capability ownersomeone who can say, today, what the firm can staff and deliver; the audit step needs this person.

The workflow

1. 1.

   Decompose the RFP into a response matrix first. Section by section: what is asked, evaluation weight if stated, mandatory format, and which library material likely answers it. This matrix drives retrieval and, later, the compliance check — RFPs are lost on unanswered mandatory items more often than on weak prose.

2. 2.

   Retrieve and draft section by section, with provenance tagged.

   Copy

   Draft the "Solution Approach — Conversational AI" section for the
   Meridian Retail RFP response. Requirement (verbatim from RFP §4.2):
   [paste]. Retrieved library material attached.
   
   Rules:
   - Ground every capability claim, metric, and reference in the
     attached material; tag each inline as [SRC: doc, section].
   - Tag every claim that asserts numbers, certifications, named
     clients, or delivery track record as [AUDIT] — these get human
     verification.
   - Where the library does not support an answer the RFP requires,
     write [GAP: what is missing] — do not bridge with plausible
     capability language.
   - Tone: our house proposal voice (examples attached); tailor to this
     client's stated priorities (speed to pilot, US data residency).
   

3. 3.

   Audit every [AUDIT] tag with the capability owner. This is the step that distinguishes reuse from over-claim, and it is human by design. Three questions per claim: is it still true (the 2024 metric may not be), can we evidence it if challenged (procurement teams check references), and does it commit delivery to something this team can staff (the claim becomes a contract exhibit). Rajesh's rule for Meridian: any claim the delivery lead will not initial gets rewritten to what they will — "we have delivered similar" becomes "we have delivered X and Y, references available," or it leaves the document.

4. 4.

   Close the [GAP] list honestly. Each gap is a bid decision, not a writing problem: partner to cover it, scope it out explicitly, invest to build it before award — or acknowledge it and compete on the rest. The one prohibited resolution is prose that implies the capability exists. Gaps handled honestly cost points; gaps papered over cost the engagement when delivery cannot cash the proposal's checks.

5. 5.

   Run the win-theme consistency pass. Assembled responses from multiple authors drift: "AI-first delivery accelerators" in section 2, silence about them in section 6's staffing plan. One pass over the full draft: "list every win theme and differentiator asserted, where each appears, and where the response contradicts or fails to support it" (pb-16's method, pointed at persuasion). Fix toward support or remove the theme — an unsupported differentiator invites the exact evaluator question you cannot answer.

6. 6.

   Compliance-check against the matrix, then freeze provenance. Every mandatory item answered, in the mandated format, within limits — checked mechanically against the step-1 matrix. Strip the [SRC]/[AUDIT] tags from the final, but archive the tagged version with the bid records: when the client challenges a claim in BAFO or the engagement's month two, you want the provenance in one lookup.

Verify before you trust it

• Every metric and named reference re-verified this bid. Track records age, reference clients churn, certifications lapse. The library's date is not the claim's date — the capability owner's initials are.
• No cross-client leakage. Search the draft for every past client name and project identifier in your corpus. One leaked name converts a proposal into an NDA incident with your best reference.
• Delivery signs the solution. The proposed architecture and team shape get thirty minutes with whoever must deliver them. Presales fiction is a services-industry classic that AI now produces faster — the countermeasure is unchanged.
• The GenAI section survives an expert read. 2026 evaluators increasingly include AI-literate reviewers; have your strongest AI architect read that section for claims that are technically hollow ("our proprietary LLM framework") — hollow AI claims now lose bids they used to win.

Restricted-environment variant

The reuse library is your firm's commercial crown jewels, and the RFP itself is the client's confidential document — both sides of this playbook are restricted by default. Index and draft only on the firm's enterprise tenancy (no-training terms confirmed); never a consumer tool for either corpus. Where the RFP prohibits AI processing of its contents (appearing in 2026 procurement terms — check the instructions section), decompose the RFP manually and use retrieval only over your own library, keeping RFP text out of prompts. Multi-client corpora deserve tenant separation (pb-18 discipline): the model that drafts against Client A's bid should not retrieve Client B's pricing.