How do AI agents pre-qualify construction RFQs?

A pre-qualification agent reads each incoming RFQ or bid package the day it arrives, extracts scope, schedule, bonding and insurance requirements, geography and owner, and scores it against the contractor's own history: past wins and losses by owner and scope, margin achieved, crew and bonding capacity, subcontractor availability. It returns a go/no-go recommendation with reasons and the three questions an estimator should answer before committing a week to the bid.

Can an on-premise model draft construction proposals?

Yes. A proposal agent assembles qualification statements, safety records, insurance and bonding responses, key-personnel bios and the narrative from the contractor's own past submissions, matched to the RFP's requirements and formatted to its structure. Estimators review and finish rather than start from a blank page. Adapters fine-tuned on past proposals keep the voice and standards consistent.

What is deal underwriting with AI in construction?

An underwriting agent builds a deal memo from the company's records: comparable projects and their actual versus estimated costs, schedule risk from similar scopes, subcontractor performance, owner payment history and market unit costs. It flags where the estimate departs from history and attaches the evidence. The memo is a draft for a human decision, with sources, not an automated approval.

Why does owning the AI stack change a contractor's valuation?

Because buyers and lenders price repeatability and transferability. A contractor whose bidding, estimating and underwriting run on documented workflows, proprietary data and software it owns presents as a software-enabled business with lower key-person risk. The data moat — years of priced bids and outcomes — cannot be replicated by a competitor or rented from a vendor.

Is this a real contractor case study?

It is a representative deployment blueprint — anonymized and generalized from the construction and industrial patterns we design. No company is named and no win-rate or margin figures are invented; every quantitative claim is a structural property of the architecture.

Deployment Blueprint: A General Contractor Runs Bidding and Underwriting on Agents It Owns

This blueprint is a representative reference architecture — anonymized and generalized from the construction and industrial deployment patterns we design. No company is named, and no outcome figures are invented.

The situation

A general contractor — commercial and light-industrial work, a few hundred million in annual volume, a bid pipeline of dozens of RFQs and RFPs a month — holds its competitive edge in what it knows: which owners pay, which scopes it wins and at what margin, what a square foot of a given assembly costs in each market this quarter, which subcontractors show up. That knowledge lives in two senior estimators, a chief estimator close to retirement, and a shared drive of PDFs: drawings, specs, RFIs, submittals, change orders, daily logs, closeout packages, and a decade of bids with outcomes.

Every RFQ package is read by hand. Every proposal is rebuilt from the last one. Every underwriting memo for a development deal starts from a blank page and a phone call. The owners want to grow and eventually sell, and they have been told by an advisor that the company is priced like a project business because its knowledge walks out at five o'clock.

The constraint

The obvious fix is a construction AI subscription. The owners rejected it for a reason that has nothing to do with capability: the data. Bid history, unit costs, margins and subcontractor scorecards are the company's proprietary asset. Pushed through a vendor's API they become the vendor's training data and, over time, every competitor's benchmark. Some of the company's public-sector work also carries confidentiality clauses that make third-party processing of project documents a contract question.

The requirement, then: frontier-class assistance for estimators and the development team, on infrastructure the company owns, with every document, cost and outcome staying inside the building.

The architecture

Model tier. GLM-5.3-Flash at 4-bit on a two-GPU node in the company's server room — about 200 GB of accelerator memory, MIT-licensed, natively multimodal so it reads drawings, scanned forms and marked-up PDFs directly. A 1M-token context window carries an entire bid package or a project's RFI log in one pass. FP8 on a four-GPU node is the growth path when concurrency demands it.

The second brain. A governed knowledge base over the company's records: drawings and specifications, RFIs and submittals, change orders with their justifications, daily logs, estimates and bid submissions, bid outcomes and post-mortems, subcontractor performance, closeout packages. Retrieval is scoped by project and by role, so a project manager sees their jobs and the chief estimator sees everything. Documents carry provenance; every answer cites the record it came from.

Specialist agents as skills. Each workflow is a written procedure the model runs, kept in the company's own skills library where anyone can read and improve it:

Agent	Trigger	Reads	Produces
Pre-qualification checker	New RFQ/RFP in the bid inbox	Bid package; win/loss history; bonding and crew capacity; sub availability	Go/no-go with reasons; three questions for the estimator
Proposal drafter	Go decision	Past submissions; safety and insurance records; personnel bios; RFP structure	Draft proposal matched to the RFP, sources attached
Underwriter	Development deal or negotiated work	Comparable projects (estimate vs actual); schedule history; sub performance; owner payment history; market unit costs	Deal memo with risk flags and evidence
RFI / change-order assistant	New RFI or change event	Contract, specs, drawings, correspondence	Draft response with clauses cited; pricing template pre-filled

The agent that runs the agents. A supervising agent owns outcomes and the calendar. It watches the bid inbox and dispatches the pre-qualification checker; it runs nightly and weekly crons (new RFQs scored before the 7 a.m. estimating meeting; a weekly subcontractor-availability call list; a monthly cost-code drift report comparing estimates to actuals); it checks specialist output against the source of truth and escalates to a person only where judgment is needed. It runs on a self-hosted agent framework — see our comparison of OpenClaw 2.0 and Hermes Agent — with every action logged.

Fine-tunes and LoRAs. After the first quarter, adapters trained on the company's proposals, estimates and correspondence — inside the environment, on its own hardware — teach the model the house voice, cost codes and standards. Drafts land closer to final; estimators trust them sooner. The adapters are the company's property.

Sovereignty properties. No data egress. SSO and role-based access by project. Full audit logging of prompts, retrievals and agent actions. Air-gapped operation available for the public-sector lines. Open weights are static files: no telemetry, no license-check callbacks, no vendor in the loop.

The rollout

Phase	Weeks	What happens
Sovereignty assessment	1–2	Data classes, bid pipeline, systems inventory, contract obligations. Written architecture and cost model.
Second brain + first agents	3–8	Archive loaded; retrieval live; pre-qualification and proposal skills running against real RFQs with estimators reviewing every output.
Underwriting, crons, adapters	Quarter 2	Deal memos in use; scheduled runs; LoRAs trained on the house corpus. Managed operations, evaluations and model upgrades from here.

What it is worth

The estimating team spends its week on bids it can win and starts proposals from a draft instead of a blank page. The development team underwrites from the company's own history rather than memory. The chief estimator's judgment is captured in skills and adapters before retirement rather than lost with it.

And the company looks different to an acquirer or a lender. Its bidding, estimating and underwriting run on documented workflows, proprietary data and software it owns — a software-enabled business with lower key-person risk, whose data moat (years of priced bids and outcomes, sub performance, unit costs by market) cannot be bought from anyone else. That is the argument for owning the stack rather than renting it: the same agents that win the next bid are the evidence, at exit or refinancing, that the company runs on systems.

The construction industry page has the pattern in summary; the fine-tuning solution covers the adapter work; and the on-premise deployment cost guide has the hardware tiers and the break-even against metered APIs.

Deployment blueprints are representative reference architectures — anonymized and generalized from the deployment patterns we design. They are not client testimonials.