Where to start with AI · First workflow No. 7

Responding to RFPs

An RFP response assistant, rolled out in two stages: it drafts a sourced first pass, then, on its own, completes the authorized standard sections. Commitments stay with people.

A first AI use case doesn’t need to be spectacular. It does need to be narrow enough to test, measure, and improve.

Download the Scoping Sheet (PDF, 2 pages) View the Carousel (PDF)

Low risk, but only within a strict scope

Responses to requests for proposals (RFPs) meet all four tests of a good first workflow: they come up often, they have a clear owner, they call for a clear output (a sourced first draft, question by question), and they are a pain point that sales and the experts they pull in already feel.

The use case is low-risk only within a strict scope:

An approved answer library kept current, with an owner for each answer.
Current internal sources data sheets, certificates, approved references.
No commercial commitment left to AI: pricing, timelines, warranties, signature.
Unusual, sensitive, or ambiguous bids routed to the sales engineer.

The assistant draws on past proposals, product data sheets, and customer references the company already has. The risk does not go away. It becomes more visible, measurable, and manageable once the sources, the scope, and the escalation thresholds are defined.

A three-step approach

Build the answer library

Gather approved answers, data sheets, certificates, and cleared references; give each answer an owner and an expiration date. The tool comes after.

The assistant drafts the first pass, the sales engineer decides

The assistant extracts the requirements from the RFP package, searches the approved answers, product data sheets, and cleared customer references, and drafts a first pass that cites every source. The sales engineer completes, corrects, and prices it. Every correction is logged, section by section.

The assistant completes authorized standard sections

On its own, it completes sections that match an approved answer backed by a current source cleared for use, with no pricing, timeline, or commitment involved. It routes to the sales engineer any new or ambiguous question, any question no source covers, and anything involving pricing, timelines, warranties, or deviations. An authorized signatory always reviews and signs the bid before submission.

Moving from stage 1 to stage 2

The move to stage 2 should not be decided in advance. It depends on what the sales team observes during stage 1: the accuracy of the proposed answers, the corrections needed, and the system’s ability to recognize questions that fall outside its scope.

Accuracy and sourcing of proposed answersCorrections needed, by sectionExceptions correctly routedNo significant incident observed within the tested scopeA current library, each answer with an owner

The company overview and certifications could move to stage 2, while technical answers stay in stage 1.

Reference architecture: how the system stays under control

Requirements extracted from the RFP package, permission-filtered hybrid search over a library of approved answers, deterministic rules for source validity and commitment sections, answers drafted by an LLM instructed to cite every source, and input and output guardrails with routing to the sales engineer. A loop connects the log, sales review, and the evaluation set.

Click the diagram to open it full size.

The controls required before any autonomous answer:

Evaluation. A reference set of about 40 past bids, pseudonymized or minimized, with the final answers annotated by sales engineers. Metrics: answer accuracy and sourcing, unsourced claims, requirement coverage, and the rate at which exceptions are correctly routed. Rerun whenever the model, the library, or the instructions change.

Observability. Every run is logged: requirements extracted, IDs of cited sources, routing decision. Buyer data, pricing, and résumés: log minimized, access restricted, retention defined by purpose and applicable obligations.

Identity and access. SSO (SAML or OIDC). Permissions apply at retrieval: pricing, margins, and résumés visible by role. Only an answer’s owner can change it. Signing authority stays with authorized signatories.

Data. Hosting and data transfers verified for the chosen solution; contractual terms barring use of the data for model training; encryption; buyer confidentiality and RFP rules taken into account; documented Swiss FADP assessment.

Application security. Safeguards against prompt injection, including in the buyer’s documents. Permitted outcomes: draft a sourced first pass, complete an authorized standard section, or route to a person. Submission and signature stay with people.

Corrections go first into improving the library, rules, instructions, and evaluation set. Retraining the model is neither the starting point nor a substitute for that work.

The goal

The goal is not to take bid responses away from the sales team. It is to cut the time spent finding and rewriting the same text, get a complete first draft sooner, and free up sales engineers’ time for the solution, the pricing, and the commitment.

Download the Scoping Sheet (PDF) View the Carousel (PDF)

Illustrative case: fictitious company; all figures are approximate. Sergio Castagna Consulting is a member of the OpenAI Partner Network and a registered partner in the Google Cloud Partner Network. The tool comes after the problem is defined.