Superintelligent Agents for Civil Engineers: RFI and Submittal Logs
An AI agent can track RFI and submittal logs, draft responses and flag overdue items, while an engineer signs off every technical reply.

An AI agent for a civil engineering firm can keep RFI and submittal logs current by reading incoming correspondence, recording each item with its status, due date and linked drawings, and drafting responses for engineers. The agent never issues a technical position on its own: every response that touches design, specifications or code needs an engineer's sign-off before it leaves the firm. This article explains how that division of work is set up, what the agent should and should not touch, and where to start.
Key takeaways
- An RFI and submittal log is a record of open questions, decisions and deadlines, and most of its daily work is administrative, which is where an agent helps first.
- The agent should register, route, remind and draft. It should not decide what a connection detail, a pipe slope or a specification clause means.
- Every technical response leaves the firm only after a named engineer approves the exact final text, and that approval is recorded in the log.
- Permissions should be narrow: read the project mailbox and document register, write to the log, create drafts, and never send to a contractor without a release step.
- Success is measured by whether every open item has an owner, a due date and a complete trail, not by how many responses the agent produces.
What an RFI and submittal log has to answer
A request for information (RFI) is a written question from a contractor or subcontractor about the contract documents. A submittal is a product sheet, shop drawing or sample sent for review before installation. Both logs exist to answer the same practical questions: what was asked, who owns the answer, what was decided, and whether that decision changed cost or schedule. Many firms keep these logs in spreadsheets, shared folders and email threads that slowly drift apart. The usual problem is not that no log exists. It is the gap between the log and the email that quietly changed a detail.
For each entry, the record usually needs a sequential number, the sheet or specification section referenced, the question in plain words, the responsible reviewer, the date received, the response due date, the current status, and links to the response and any revised drawing. A useful log also notes whether the answer affected scope, cost or the critical path, because those are the items a project manager or owner asks about first.
Which parts of the log an agent can own
Most of the work around the log is structured and repetitive, which makes it a good first place for an agent. With read access to the project mailbox and document register, an agent can detect new RFI and submittal emails, assign the next number in your existing sequence, extract the sheet and specification references, check whether the cited drawing revision is the current issued set, and calculate a due date from the review period in the contract. It can remind reviewers, flag items approaching their deadline, and prepare the summary for the weekly coordination meeting.
- Register incoming RFIs and submittals in the log using your existing numbering and field names.
- Check that the referenced drawing or specification revision matches the current issued set.
- Calculate response due dates from the review periods the firm has entered into the agent's memory.
- Send reminders to named reviewers and escalate overdue items to the project lead.
- Draft a response for an engineer, with the source clauses and drawings attached.
- Prepare the weekly open items summary and the log export for the project file.
Each of these tasks is a clerical or retrieval task with a clear right answer. That is why they belong to the agent. The judgment behind the answer does not.
The sign-off rule: every technical response needs an engineer
The line between helpful and risky sits in one place. Any reply that states a dimension, accepts or rejects a product, interprets a specification clause or changes a detail is a technical position. In a firm that carries professional responsibility for the design, that position belongs to a licensed engineer who has checked it. Licensing rules differ by jurisdiction, but the operating principle is the same: the firm needs a named person who owns each position. The agent may prepare the draft, but it cannot release it. The NIST AI Risk Management Framework treats human oversight and clear accountability as part of managing AI risk, and that principle fits this workflow closely.
Incoming RFIs are written by outside parties, and an email can contain text that reads like an instruction, such as a request to mark an item accepted. The agent must treat that text as content to log, not as a command. The OWASP guidance on prompt injection describes this class of risk, and the answer is structural: the agent's permissions do not allow it to change a status to approved or release a response on the strength of an email alone.
Every released response should carry a sign-off record with the following items:
- The name of the engineer who reviewed the final text, not only the draft.
- The drawing and specification revisions relied on, with their issue identifiers.
- Any assumption the engineer added, removed or qualified.
- The time of release and the channel used to send it.
- A link to any revised sheet, stamped submittal or markup issued with the response.
The audit trail matters as much as the answer. OpenTelemetry's observability concepts describe the value of recording what a system did, in sequence, so that it can be reviewed later. The same discipline applies to an agent: which document it read, which clause it cited, which draft it produced and who approved it should all be recoverable without reconstructing the history from memory.
A step-by-step plan for the first engagement
A first engagement should be small enough to check every output and specific enough to show whether the agent is useful. The sequence below keeps the scope narrow.
- Choose one active project with an open RFI and submittal log and a project lead willing to review the output every week.
- Document the log fields in use today, the numbering rule, the contract review periods and the discipline reviewers who sign off on each area.
- Connect read access to the project mailbox and document register, and write access only to the log file.
- Allow the agent to create drafts but not to send anything to contractors. Record each release as a separate, named action by an engineer.
- Run the agent on incoming items and compare every log entry against the manual record for the same period.
- Review each error with the project lead, correct the memory and rules, and only then consider a second project.
Permissions should follow the principle of least privilege described in the OWASP authorization guidance. Grant the agent only the access the current task requires, and review that access when the project team changes or a new discipline joins the work.
Illustrative scenario: a substitute anchor on a retaining wall
Illustrative scenario, not a real project: a contractor submits a product data sheet for a substitute anchor on a retaining wall detail. The agent registers the item in the submittal log with the next number in sequence and links the specification section and the current detail sheet. It checks the approved product list in the project memory and finds that the proposed anchor's load rating is not listed in the specification. The agent drafts a response stating that the item is not yet accepted and listing the missing data, marks the status as awaiting engineer review, and sends the draft with the clause and data sheet to the reviewing engineer.
The engineer rewrites two sentences, requests a test report from the manufacturer, and releases the response through the firm's normal channel. The log records who released it, when, and which revision of the detail was cited. The agent never writes that the anchor is accepted. That word belongs to the engineer, and the log is built so that it cannot appear without one.
How Be Human's managed approach handles the log
Be Human starts with one role, usually a log coordinator for one project, and defines the memory that role needs: the log format and numbering, the contract review periods, the reviewer assignments, and the approved products and standard details the firm already uses. Permissions are scoped to reading the mailbox and document register, writing to the log, and creating drafts. Release to outside parties stays with a named engineer. Managed operations then cover monitoring of the log, recurring evaluations that compare drafts with the edits reviewers make, permission reviews when the project team changes, and incident response if an item is misrouted or a draft cites a superseded revision.
Choosing the first task worth handing to an agent
The best first task is high in volume, clear in its rules and easy to audit. Poor first tasks depend on site judgment or on contract terms that the firm itself interprets differently from project to project. A simple guide helps sort the options:
- Good candidate: registering incoming items and checking revision numbers against the issued set.
- Good candidate: sending reminders and preparing the weekly open items summary.
- Good candidate: drafting responses that quote existing clauses for an engineer to review.
- Poor candidate: deciding whether a field condition justifies a change in design.
- Poor candidate: any response on a matter where the contract language is disputed.
Write the candidates down, rank them by how often they occur and how easily a reviewer can check the result, and start with the top one. If you want to find the first RFI or submittal task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent answer RFIs without an engineer reviewing them? No, and it should not be set up that way. The agent can register the item, check the referenced revisions and draft a response, but any technical position needs a named engineer to review the final text and release it. The release and the reviewer's name should both appear in the log.
Does the agent replace the project coordinator? It removes much of the repetitive tracking, such as numbering, reminders and summaries, so the coordinator can spend time on the items that need judgment and contact with the team. The coordinator still owns the process, the exceptions and the relationship with reviewers. The agent works inside that process rather than replacing it.
What information does the agent need to track submittals correctly? It needs the log fields your firm uses, the numbering rule, the review periods from the contract, the list of reviewers by discipline, the current issued drawing and specification revisions, and the list of approved products or standard details. Without the revision set and the review periods, the agent can record items but cannot reliably calculate deadlines or flag stale references.
How is the agent's access to project email controlled? Access should be limited to the mailbox and folders needed for the project, with read permission for correspondence and write permission only to the log and draft folders. Nothing should be sent to outside parties without a named engineer's release. Access should be reviewed when team members change, and every action should be recorded so the firm can see what the agent read and produced.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


