General Contracting
For every new bid, estimators reviewed extensive project documents, identified which work should be subcontracted, and manually created a tailored RFP for each trade. We built an agentic AI product that turns a client bid package into reviewable, source-backed RFPs, while keeping the estimator in control.
Users
Construction estimators
Before
~3 hrs per RFP
Target
~30 min per RFP
When a general contractor received a new bid opportunity, estimators had to read more than 100 pages of scopes, specifications, drawings, and contractual requirements.
From those documents, they had to determine which work would be done internally, which services required subcontractors, what requirements applied to each trade, what information belonged in each vendor's RFP, and which source documents needed to accompany it. The estimator then repeated the process for every subcontractor, often spending 2 to 4 hours on each RFP.
The same bid package had to be interpreted again and again, once for every subcontractor.
The estimator uploads the client's bid documents. The product analyzes the scope, proposes a work breakdown, and generates a tailored RFP for every confirmed subcontracted service. Each RFP follows the contractor's existing ten-section template and connects generated content back to the original source documents.
01
The client's scope of work, specifications, and supporting project documents.
02
The system recommends self-perform, subcontract, or partial for each service. The estimator corrects and adds context.
03
For each subcontracted service, the system extracts requirements and drafts an RFP from the standard template.
04
The estimator reviews the draft beside the original documents and validates the evidence.
05
The final RFP exports as a formatted .docx, ready for the vendor distribution package.
01
Missing a subcontracted service could mean an RFP never gets created. We prioritized recall, and made it easy to add, remove, or reclassify services before generation.
02
Every generated section links back to its supporting source passage, so estimators can verify content without searching the entire bid package.
03
When information like internally supplied materials isn't in the client documents, the system marks the section Needs Input instead of generating an assumption.
We evaluated more than whether an RFP was generated successfully. The system was designed to measure service recall, grounding, coverage, acceptance rate, rewrite rate, and whether review was actually faster than writing from scratch.
If estimators rewrote the draft from scratch, the product hadn't automated the work. It had only added another review step.