STATEMETHOD

Guide

AI for construction document control: where it helps and where it does not.

Document control decides which version of a drawing, specification or report is current, who received it and what was approved. Whichever platform a project uses, each document control step raises the same three points: what a model can read and propose, which check must stay a fixed rule in the register, and which decision stays with the document controller.

A model is good at reading content that arrives in inconsistent forms, such as title blocks, cover emails, transmittals and comment sheets. It should not be the record. The register stays the source of truth, software enforces its rules, and a named person accepts every change to it.

Hugo Sequier runs State Method, which sells the assessments, pilots and builds these guides discuss. Apply the same tests to it as to any other provider.

Author
Hugo Sequier
Published
31 October 2026
Reading time
6 min
Updated
3 November 2026

01 / Split

Three jobs, three owners.

Writing this split down before any tool is chosen is the most useful step. It shows where a model adds speed and where it would add risk.

The model reads and proposes

Classifies an incoming file, reads the document number, title, revision, date and discipline from the title block or cover sheet, matches an email attachment to a register entry, summarises what changed between revisions and finds references to a document that has been superseded. Every proposal names the page or field it was read from.

Software enforces the rules

Checks the naming convention, the revision sequence, the status codes, duplicates, required fields and the distribution matrix. These are fixed rules with one correct answer, so they belong in code or in the register’s own validation, not in a model’s judgment.

A named person approves

The document controller accepts an entry into the register, issues a transmittal and marks a revision as superseded. Reviewers keep the status they give a submittal or drawing. Nothing changes the controlled record without that approval.

02 / Step map

Where AI helps in each document control step.

The usual document control steps on a construction project, what a model can prepare in each, and the check or approval that stays outside it. The examples linked below use synthetic project data.

Intake and registration

A model reads incoming files and their cover emails and proposes the register fields. Software checks the proposal against the naming convention and the existing register, flags duplicates and missing attachments, and the controller accepts or corrects it.

Example: RFI and submittal intake →

Naming and metadata

Where filenames are inconsistent, a model can propose the correct document number and metadata from the content itself. Software validates the format. No file is renamed or moved without a register record of the change.

Revisions and supersession

Software checks the revision sequence against the register. A model then finds which open RFIs, submittals, work packages or notices still reference the superseded revision, so the controller can issue one coordination notice instead of searching by hand.

Example: drawing revision coordination →

Distribution and transmittals

Who receives which document stays a rule table, the distribution matrix. A model can draft the transmittal cover text and the summary of changes; the controller sends it.

Review and comments

A model can extract comments from marked-up sets and comment sheets, map each one to a sheet and revision, and surface conflicts between packages with the source cited. The reviewer still issues the review record.

Example: construction document and drawing review →

Search and retrieval

A model can answer questions across the project documents, provided it reads only current revisions and every answer names the document number and revision it came from. An answer without that citation cannot be checked.

Handover and closeout

A model can match O&M manuals, certificates, warranties and as-built drawings to the handover obligations list and show every gap. The handover manager approves the pack.

Example: handover document preparation →

03 / Limits

Where AI should not decide.

These are the points where a fluent model output causes the most damage, because the error looks like a valid record.

  • Which revision is current. That is a rule applied to the register, not a judgment about the content.
  • The status of a submittal or drawing, such as approved, approved as noted or rejected. That belongs to the named reviewer.
  • Who receives a document. The distribution matrix decides, and changing it is a controlled action.
  • Renaming, moving or deleting files without a record in the register.
  • Filling a missing title block field by guessing. A missing field is flagged and returned, not completed.
  • Anything at all when the register itself is unreliable. Fix the register first; AI on a broken register produces confident wrong answers faster.

A model that reads a superseded drawing answers just as fluently as one that reads the current revision. Requiring every output to name the document number and revision it read is the simplest control against that failure.

04 / Before starting

Settle these before buying or building anything.

  • There is one register per project, and everyone treats it as the source of truth.
  • The naming convention and status codes are written down, including the exceptions people actually use.
  • It is known where documents arrive: the project platform, email, a shared drive or several of these.
  • Each project has a named document controller who can say what a correct register entry is.
  • The weekly volume of incoming documents and the time each one takes today are recorded.
  • The cost of a wrong registration or a missed supersession is understood, for example rework built from an old revision.

If documents already arrive and live in one platform, start with that platform’s own features. A custom workflow is worth considering when documents arrive from several parties and channels, the register sits in a separate system, and supersession has to be traced to open items across them.

05 / Trial

Test it on past documents before trusting it.

The same trial works for a platform feature, a specialist tool or a custom workflow, so the options can be compared on equal terms.

  • Pick 30 or more past incoming documents of mixed types, including misnamed files, missing attachments and at least one superseded revision.
  • Write the correct register entry for each before anyone sees an output.
  • Score the proposals field by field against those entries, not by overall impression.
  • Check that the superseded revision is flagged and that every open reference to it is found.
  • Check that each proposal cites the page or field it was read from, and that a reviewer can open it in one step.
  • Record time per document for the current manual process and for the trial.

Before uploading project files to any tool, check its data handling with the vendor questions in the AI proof of concept guide.

FAQ

Before the next decision.

Will AI replace document controllers?

Not in a controlled process. It can move the document controller’s time from typing metadata and searching for references to checking proposals and resolving exceptions. Someone still has to own the register, accept entries, issue transmittals and answer for what is current.

How should AI handle drawing revisions?

Software should check the revision sequence against the register, and a model should only help find what the change affects: open RFIs, submittals, packages or notices that still cite the superseded revision. Every output should name the document number and revision it read, and the controller issues the coordination notice.

Assessment / one workflow

Test one document control workflow.

Describe where documents arrive, the register they go into, who controls it and how long each document takes today. A person reviews every assessment. Do not send confidential documents through the public form.