Contractors are open to AI. In the 2026 Workforce Survey from the Associated General Contractors of America and NCCER, 51 percent expect AI and robotics to have a positive impact on construction over the next five years, against 8 percent expecting a negative one. Inside their own companies, they expect the biggest impact on estimating and bidding and on document and contract review.
Source: AGC and NCCER, 2026 Workforce Survey Analysis.
The usual pitch is that a better model gets you there. In our experience running a live product in Europe, it does not. Preconstruction runs on four formats that a general-purpose model handles badly:
- Bid-form PDFs. Often hundreds of pages of scanned forms, exported tables and plan sheets. Sections follow CSI MasterFormat, but the file is a flat PDF.
- AIA contracts. A201-2017 is the common frame, but every project layers supplementary conditions and owner edits on top. A plain PDF-to-text pass loses the section structure.
- Specifications. Organized by MasterFormat division, with each section (03 30 00, 09 29 00, 26 05 19) split into its own Parts 1, 2 and 3. One product can appear in three sections with three acceptance criteria.
- Schedules of values. AIA G702 and G703, usually XLSX. The line items trace back through the contract and the bid form to the solicitation, and each one uses a different schema.
Why does a general-purpose LLM fall short?
A generic model can read any one of these. It cannot keep them aligned. The bid form says 5/8 in. Type X gypsum at corridor partitions. Spec section 09 29 00 adds a 2-hour rating and a named manufacturer. An owner edit adds Davis-Bacon wage requirements. The sub proposal prices a 1-hour assembly.
Someone has to catch that mismatch, cite each source and decide which document governs. Paste all four into one chat and you get a plausible paragraph with no page cites, which an estimator cannot check.
AI-ready, for preconstruction, means three things. A knowledge base of US construction standards. A citation model that ties each finding to a page, clause or line. A project record shared by every agent, so nothing is lost between bid, contract and award.
What does AI-ready look like in US preconstruction?
A standards-indexed knowledge base
An index of the forms and rules the documents point to: AIA A201-2017, ConsensusDocs 200, EJCDC C-700, CSI MasterFormat, FAR 52, state DOT specifications. When a subcontract says pay-if-paid, the system should know that A201 only sets payment timing and that states treat condition-precedent clauses differently. AIA A201-2017 § 9.6.2; Tex. Bus. & Com. Code ch. 56
Citation, not summary
Every number and every risk carries the sheet, section and line. "This sub excludes dewatering. See 31 23 19, 1.4.A, page 47." If the system cannot show the source, the finding does not go into your bid summary.
A project record shared across agents
Clara reads the solicitation and flags DBE goals and LDs. Dora reads the contract against the scope Clara pulled. Nova levels sub bids against the same scope. Theo ties plan quantities to it. If Clara flags a scope gap and Nova finds it in a sub bid, the two findings are already linked.
The record does not stop at award. Patrick, Dario, Lex, Fina, Colin and Pixer carry it into delivery: a delivery note is checked against its purchase order, a punch item keeps the photo it came from, and a pay application ties back to the schedule of values the estimate started.
What should a GC do about it in 2026?
- Catalogue your documents first. How many bid packages a year, in what formats? If you cannot answer in a sentence, no AI will help.
- Pick two workflows where you already lose money. Scope gaps in sub bids. Risk clauses buried in subcontracts. Start with the leak, not the whole lifecycle.
- Insist on cited output, on your own documents. Ask every vendor to run one of your real bid packages. If the result cannot show the clause, sheet and line, it is a demo.
Our bet: for US preconstruction, the answer is specialist agents on one knowledge base and one project record, not a general chat window. See how the BAU AI agents work together, or watch Clara work to check the cites yourself.