CATALINA COMMERCIAL
← Catalina Commercial
Perspective 03Efficiency

Where AI takes hours out of the job, and where it does not.

Construction runs on documents, and most of the hours go into reading them. We have moved that reading onto language models. A person still makes every call.

Macro detail of a blackened steel reveal meeting rift-sawn white oak.

The constraint on a lean team is never ambition, it is hours. Everything described here runs on off-the-shelf models with document extraction, and none of it makes a decision, signs anything or talks to a subcontractor.

Bid leveling

A framing package comes back from six subcontractors in six formats, and the work is finding what each one left out. The model reads all six against the scope sheet and returns a normalized comparison with a line for every inclusion, exclusion and clarification, flagging where one bidder carried something the others did not. A person confirms each flag against the drawings before it goes into a scope letter.

Submittal and RFI triage

The log is where a project quietly loses time. Every incoming submittal is read against the specification section it claims to satisfy and comes back tagged as conforming, deviating or incomplete, with the deviation quoted. Duplicate RFIs get matched to the answer already issued, so the superintendent opens a short list each morning rather than the whole log.

Pay application review

Each application is checked line by line against the schedule of values, the approved change orders and the stored quantities from the last inspection walk. Anything billed ahead of what is installed is surfaced as its own line before a check is cut, which is the gap that otherwise shows up much later.

Plan check and correction letters

A municipal correction letter arrives as thirty numbered comments across five disciplines. Each one is routed to the responsible consultant with the prior response attached where the same comment came up on an earlier cycle, so nobody is rebuilding context every round.

Lease and document abstraction

Executed leases, warranties and closeout documents get abstracted into the fields somebody will actually need later, including term, options, exclusions and who honors what. The value is in retrieval a year from now, not in the abstraction itself.

A model that reads a correction letter and routes thirty comments is doing clerical work quickly. That is the whole claim. It does not design the building, price the job or choose a subcontractor.

The guardrails

Extraction error is the real risk, so every output carries the page and paragraph it came from, and a person checks anything that moves money before it moves. We do not put a model between a subcontractor and an answer. We do not let it write scope language that goes into a contract without counsel reading it. Project documents stay inside our own tenancy and are not used to train anything.

Why an owner should care

None of this makes the building better on its own. What it does is give the people running your job their evenings back for the work that actually needs judgment, which is walking the site, calling the sub whose manpower slipped, and reading the drawing again.

ReadsBids, submittals, RFIs, pay applications, correction letters, leasesDocuments that arrive in inconsistent formats and have to be compared against a standard.
ReturnsA normalized comparison or a short exception listWith the source passage attached, so a person can check it in seconds rather than rereading the original.
Never doesAward a package, approve a payment, answer a subcontractor, or write contract languageThose are judgment calls with money and liability attached, and they stay with people.