Draft the RFP, Then Dissect Vendor Responses into a Claims Table

Workflow area: Solution Evaluation · Tier: Core · Hands-on time: ~40 minutes

Vendor responses are written by proposal teams whose job is to make every requirement sound met. "Our platform supports real-time integration" might mean a streaming API, a nightly CSV drop, or a roadmap slide. Analysts who score responses by reading them absorb the framing; analysts who decompose them into discrete claims — each one tagged by evidence status and cited by page — can defend their evaluation matrix in front of the losing vendor's lawyers. Long-context models make the decomposition affordable; the tagging judgment stays yours.

When to use — and when not to

• Use it when drafting an RFP from a baselined requirements setthe AI converts requirements into weighted criteria and unambiguous questions faster than a template ever did.
• Use it on every substantive vendor responsethree responses at 80–120 pages each is exactly the reading load where human attention fails and claims tables pay off.
• Use it to prepare vendor clarification roundsthe gap and vendor-claim rows are your question list, pre-cited.
• Skip it as a substitute for reference checks and demos-against-scriptsthe claims table tells you what to verify, not that you have.
• Skip weighting decisions delegated to the modelcriteria weights encode business priorities; the AI can propose a starting split, but the steering group owns the numbers.
• Skip it if procurement rules restrict toolingsome public-sector processes constrain what may touch bid documents; confirm with procurement before any upload.

The recipe

1. 1.

   Draft criteria from requirements, not from vibes. Give the AI your requirements pack: "Convert these into RFP evaluation criteria grouped by theme. For each: criterion, what evidence a response must provide, and a proposed weight summing to 100. Flag requirements too vague to evaluate — I will rewrite them first." The flagged list is a bonus quality gate on your own requirements. Take the draft weights to the steering group for adjustment and sign-off before the RFP ships.

2. 2.

   Issue questions that force falsifiable answers. Have the AI rewrite each criterion as a question demanding specifics: versions, dates, named referenceable customers, SLA numbers, "in production today or on roadmap?" A question that cannot be answered evasively is the cheapest verification tool you will ever deploy.

3. 3.

   Decompose each response into a claims table. One vendor at a time, attach the full response:

   Copy

   You are supporting a vendor evaluation. Attached: vendor
   response to our RFP. Our criteria: [PASTE CRITERIA LIST].
   
   Extract every material claim relevant to any criterion into
   a table: CLAIM-ID | criterion | exact claim (verbatim quote)
   | page/section | evidence offered in the response | proposed
   tag: VERIFIED (independent evidence provided: named customer,
   audited cert, contractual SLA) / VENDOR-CLAIM (asserted, no
   independent evidence) / GAP (criterion unaddressed or answer
   evades the question) / CONTRADICTION (conflicts with another
   part of the response — cite both pages).
   
   Rules: quote verbatim, never paraphrase inside quotes. One
   claim per row. If a claim is conditional ("can support",
   "roadmap", "via partner"), preserve the hedge in the quote
   and note it. List criteria with NO claims at the end.
   

4. 4.

   Audit the tags — this is the analyst's job, not the model's. Sample every VERIFIED row and check the evidence actually independent (a vendor's own whitepaper is not verification). Read every CONTRADICTION at both cited pages. Typical output that earns its keep:

   Copy

   C-14 | Integration | "native, real-time bi-directional sync
   with SAP ECC and S/4HANA" | p.34 | none — no version, no
   customer named | VENDOR-CLAIM (note: p.87 architecture
   diagram shows nightly batch for ECC) | → CONTRADICTION
   C-22 | Uptime | "99.95% availability" | p.41 | marketing
   page cited; SLA appendix (p.102) commits only 99.5% |
   CONTRADICTION — contractual number is 99.5%
   C-31 | Data residency | criterion unaddressed | — | — | GAP
   

5. 5.

   Score from the table, not the prose. Fill the weighted matrix using only VERIFIED evidence for full marks; VENDOR-CLAIM caps a score until the clarification round or demo-script converts it. Run the same prompt over every vendor so the tables are comparable — asymmetric extraction is asymmetric evaluation.

6. 6.

   Turn residue into the clarification letter. Every GAP, CONTRADICTION, and load-bearing VENDOR-CLAIM becomes a numbered written question citing the vendor's own page numbers. Written answers become contractual representations; that alone changes vendor behavior.

Make it yours

• Claude long-context for whole-response passes: 100+ page PDFs in one attachment keeps cross-references (like the p.34 vs p.87 catch) inside a single context window — the contradiction sweep is where long context earns its cost.
• Spreadsheet as the scoring system of record: export the claims table to the evaluation workbook; tags drive conditional formatting so the steering group sees the VENDOR-CLAIM density per vendor at a glance.
• A Project per procurement: load RFP, criteria, and all responses into one Claude Project; late-stage questions like "which vendors named a referenceable pharma customer?" get cited answers across the corpus.
• Lightweight mode for an RFI or three-pager: collapse to one prompt — "list every claim with a page cite and tag it" — the tagging discipline scales down; skipping it does not.

Quality check

• Criteria trace to requirementsevery criterion cites the requirement(s) it evaluates, and weights carry steering-group sign-off.
• Quotes are verbatim and citedspot-check ten rows per vendor against the PDF page; a paraphrase in quotation marks poisons the audit trail.
• VERIFIED means independently verifiedsampled rows hold up; vendor self-citations were demoted to VENDOR-CLAIM.
• Every criterion appears for every vendorwith a claim row or an explicit GAP; silent omissions are how evaluations get challenged.
• Contradictions were read in situboth cited pages, in context, before the tag stood.
• Scores derive from tagsa reviewer can walk any matrix cell back to specific claim rows and their evidence.