What a submittal review actually is
The submittal review is where the architect compares what was specified against what the contractor intends to build.
The supplier sends a binder: product data sheets, performance certifications, shop drawings, warranty terms. The architect reads it line by line and confirms each item meets the spec section it claims to satisfy.
Then a stamp: approved, approved as noted, revise and resubmit, or rejected. Then the supplier orders. Then the material shows up on site.
The process is forty years old and unchanged in shape since CAD replaced ink-on-mylar.
The structural problem
The architect writes a specification in CSI MasterFormat. Section 09 22 16 covers non-structural metal framing.
The architect specifies "20-gauge galvanized steel studs, G60 coating, conforming to ASTM C645, with section properties meeting the loads in the structural drawings". That is a precise, machine-readable claim.
Three pages, three claims, two hundred sections
The supplier responds with a 14-page data sheet in which the gauge and coating sit on page 3, the section properties on page 7, the warranty on page 9, and the install instructions on page 11.
The architect's job is to find three facts on three pages, then repeat that across 200 spec sections per project, across 50 projects in flight, against suppliers who all format differently.
The work is done by senior staff who could be designing buildings. The error rate is real, because humans get tired around page 11 of binder 17 of project 4.
What document extraction actually does
Modern extraction (LLM-grade, not the OCR-and-regex generation) reads a data sheet the way a human does.
It identifies the claims the supplier is making, the standards the product conforms to, the section properties, the warranty terms and exclusions. It outputs a structured record in seconds, and its accuracy on page 11 matches its accuracy on page 1.
Structure makes the comparison mechanical
Once the supplier's claims and the architect's spec are both structured, the comparison becomes arithmetic: ASTM C645 appears on both sides, G60 appears on both sides, and the section properties match within tolerance.
The submittal passes. Or one claim does not match, the system flags exactly which one, and a 30-minute review becomes a 30-second one.
What it does not do
It does not let the architect skip the review. The architect remains legally and professionally responsible for the approval.
The system surfaces the comparison and a licensed human still makes the call. Attention is what moves, from "find the relevant page" to "judge whether this substitution is acceptable", and the second question is the one the architect trained for.
Garbage in still applies
A scan of a faxed photocopy of a 1998 spec sheet remains a scan of a faxed photocopy. Extraction quality is bounded by input quality.
Suppliers who invest in clean, structured documentation end up at the top of the submittal pile, because their products are cheaper to evaluate. That market pressure compounds.
The downstream effect on procurement
The structured record does not stay in the review. It feeds procurement.
A supplier's product data, once extracted and verified against one architect's spec, is available to the next architect specifying something similar, and the platform accumulates a graph of products against specs against approvals.
Suppliers approved on similar projects rank higher. Substitutions accepted before propagate faster. Friction that lived in every individual review compounds into a knowledge base.
The shift that actually matters
The change AI is making to AEC practice in 2026 lands here, in the document grinding, rather than in design generation or in any story about replacing the architect.
Models do the reading that a licensed professional should not have been spending their afternoon on, and the professional gets back the design work they trained for.
The submittal binder is one of those grinding tasks. We are taking it apart, one spec section at a time.