Answering RFP questionnaires with agentic RAG.

Responding to RFPs meant subject-matter experts repeatedly answering questions they had already answered in past proposals — the same security questionnaires, the same capability descriptions, over and over. Meanwhile, genuinely new questions bounced between contract-building teams looking for the right owner, adding days of latency to proposal timelines.

The waste was structural: expert time went to repetition, and the routing of novel work ran on institutional memory rather than any system.

We split the problem the way a good proposal desk would: recognize what's been answered before, and know who should answer what hasn't. A curated dataset of previously answered SME queries became the knowledge base for an agentic RAG (retrieval-augmented generation) system.

A Standard Query Agent matches incoming questions against that library and returns the expert-approved answer when the match is strong. Questions that don't match go to a Routing Agent, which determines which contract-building team should own them. And because both decisions matter, dedicated validator agents independently check the Standard Query Agent's answers and the Routing Agent's assignments before anything is returned — the system reviews its own work.

The proof of concept showed that the bulk of repeat RFP questions can be answered instantly from approved content, with novel questions landing directly on the right team instead of wandering the org chart.

The validator-agent pattern — every agent's output checked by another agent — proved its worth here and carried forward into the ticket-triage system that followed.