We Used AI on a 2,000-Page Submittal. Here Is Where It Helped and Where We Still Checked

We recently received a submittal of approximately 2,000 pages. Inside it were hundreds of product-data records for light fixtures and equipment that had to be separated, named, linked and made searchable.

AI helped us process that information. It did not remove the need to check whether a file had been split at the correct page, named from the correct heading or linked to the correct record.

That example shows where agentic AI can be useful in construction: repetitive information work with clear steps, identifiable exceptions and a responsible person reviewing the result.

Construction has many such tasks. Submittal logs, RFI registers, material trackers, bulletins, ASIs, document naming and closeout records all require continuous attention. Missing one update can leave a model or project team working from the wrong information.

The goal is not to remove the project administrator, BIM coordinator or document controller. It is to give that person an assistant that can handle well-defined repetitive steps, identify exceptions and make the overall process easier to manage.

What makes agentic AI different?

An ordinary AI assistant may summarize a document or draft a response after a person asks. An agentic system is designed to carry out a multi-step process, such as monitoring a folder, identifying a new submittal, extracting its metadata, comparing it with the register, naming the file and routing an exception for review.

Futurum Group research completed in partnership with IFS surveyed 664 enterprise decision-makers across manufacturing, energy and utilities, aerospace and defense, transportation and logistics, construction, and telecommunications. Nearly two-thirds reported that manual, repetitive work consumes more than 40% of employee time. Sixty-six percent said their organizations were likely or very likely to invest in digital workers within 12 months.

These results are not limited to construction, but the underlying capacity problem is familiar on construction projects.

Start with work that has clear inputs and outputs

The safest early use cases are not open-ended technical decisions. They are repetitive workflows where the expected output can be described and checked.

For construction administration and BIM coordination, that can include:

• updating submittal and RFI logs;

• tracking new bulletins, ASIs and drawing revisions;

• comparing model inputs with the current document register;

• monitoring material and procurement records;

• naming and filing documents consistently;

• identifying missing metadata or broken links;

• assembling routine status reports; and

• organizing closeout information.

At DDG, submittal logs, RFI tracking and design-update tracking consume significant time because each new document can affect the model. The administrative task is not separate from coordination. If the register is wrong, the model team may miss a change.

A 2,000-page submittal is a practical example

We recently used AI to help process a submittal of approximately 2,000 pages containing hundreds of product-data records for light fixtures and equipment.

Instead of leaving the information buried in one large PDF, the workflow separated it into individual files, applied useful names and created hyperlinks so the documents were easier to search and manage by version. The benefit was not that AI made an approval decision. It completed a large amount of structured information work that would otherwise require repeated manual steps.

The project team still had to check the result. Files could be split at the wrong page, named from an incorrect heading or linked to the wrong record. AI made the task faster, but the source submittal remained the authority.

Human review must match the consequence of an error

No AI-produced work should be accepted simply because the workflow completed without an error message.

For low-risk, repetitive outputs, our practical starting point is to sample approximately 5% to 10% of the work, similar to reviewing structured work completed by a junior team member. The sample should cover ordinary items and likely failure points. If errors appear, the review expands and the workflow must be corrected.

That sampling rule does not apply equally to every task. Contractual notices, technical interpretations, safety-related information, code decisions, model changes and formal approvals require complete review by the responsible person. The higher the consequence, the stronger the review gate must be.

Define what the agent may see and change

An agentic workflow can act only as responsibly as its permissions allow. Before connecting one to project systems, teams should define:

1. which folders, registers and applications it may access;

2. which fields it may read, create or update;

3. which actions always require approval;

4. how exceptions are escalated;

5. how every action is logged; and

6. how an incorrect change can be reversed.

This is especially important when the same system contains commercial information, contracts, project correspondence and current construction documents.

AI should increase the administrator’s capacity

The best result is not an invisible system making decisions on its own. It is a project administrator or coordinator who can manage more information without losing control of it.

Agentic AI can watch repetitive processes, complete predictable steps and bring exceptions to the person responsible. That leaves more time for coordination, follow-up and judgment. It also creates a clearer audit trail when the workflow records what it found, what it changed and what it sent for review.

Construction does not need automation for its own sake. It needs dependable help with the information work that keeps projects moving.

For teams improving BIM and project-information workflows, DDG can help define practical automation opportunities without removing accountable human review.