AI Agents for Construction RFIs: Build or Buy

AI agents for construction RFIs can find the spec section, pull the drawing reference, and draft the response. They cannot sign it.

For a small general contractor, my answer is to buy first and build only the narrow piece your platform does not cover.

Here is where the time goes, what an agent can do with it, what review stays human, and a pilot small enough to test in a few weeks.

Where RFI time goes

An RFI is a written question about the contract documents. On a typical job there are hundreds. A Navigant Construction Forum study (April 2013) looked at about 1.1 million RFIs across 1,362 projects. The average project had 796 RFIs. The median reply took 9.7 days. The authors estimated the review and response cost at $1,080 per RFI. That is a 2013 figure, so treat it as a ratio, not a price.

The same study cites research finding that 13.2% of RFIs were “not justifiable.” The answer was already in the documents.

A small GC sits on both sides of the RFI:

• Incoming. A subcontractor asks a question. Someone on your team has to check whether the answer is already in the specs or drawings before it goes to the design team.
• Outgoing. Your team writes an RFI to the architect or engineer. A vague one bounces back or earns a slow, vague answer.

Most of the time on each RFI is not the decision. It is looking things up: finding the spec section, finding the sheet, checking the addenda, and writing a question that cites all three. That is the part an agent can take.

One caution on the number that matters most. The delay is often waiting for the design team, not your team. An agent on your side will not fix their queue. It can make your RFIs better, and fewer of them.

Spec and drawing lookup

This is the strongest use case. Given a question like “What is the fire rating for the corridor doors on level 3?”, a good agent will:

1. Search the specification sections and find the door hardware and hollow metal sections.
2. Search the drawings for the door schedule and the wall types.
3. Check addenda and bulletins for changes.
4. Return the answer with a citation: section number, sheet number, revision.
5. Flag conflicts. If the spec says 90 minutes and the schedule says 60, that is the RFI.

Text is the easy half. Specs are long, structured documents, and retrieval over them usually works well when each answer points back to a section.

Drawings are the hard half. Two pieces of evidence:

• A 2025 study put large language models through 689 construction management certification questions. GPT-4o and Claude 3.7 scored 82% and 83% overall, above the typical 70% pass mark. On questions that referred to a figure, accuracy fell to about 40%. The authors call for sustained human oversight.
• A July 2026 benchmark, DrawingVQA, tested multimodal models on 33 real “issued for construction” drawings with 92 expert-written questions. It reports a substantial gap between model and expert performance, especially on questions that need deeper reasoning.

My read: agents are useful for finding and citing. They are not yet reliable at counting, measuring, or tracing lines on a sheet. So design the agent to point at the sheet and the detail, then let a person look. Do not let it report a quantity from a plan as fact.

Drafting responses

The agent can draft in three places.

1. First-pass answer to an incoming RFI. If the documents already answer it, the agent drafts a reply with citations. Your project engineer checks the citations and sends it. This is where the “not justifiable” share pays off, because those are the questions you can close without involving the design team.

2. A well-formed outgoing RFI. If the documents conflict or are silent, the agent drafts the question: what is unclear, which references conflict, what the contractor proposes, and what the schedule impact is. A good RFI proposes a solution. Navigant’s recommendations include submitting RFIs as soon as the need is recognized and at least 10 days before the answer is needed.

3. Log hygiene. The agent tags each RFI by spec division, links related RFIs, and flags ones that are going past due.

Every draft should carry its evidence. A response without a section and sheet reference is a guess. Make “no citation, no draft” a hard rule.

Review that stays human

This is the part you cannot skip, and the contract explains why.

Under the AIA A201-2017 general conditions, the architect’s answer to a written request is a contract act. Section 4.2.11 says the architect will interpret and decide matters concerning the contract documents on written request, and respond in writing within agreed time limits or with reasonable promptness. An AI draft is not that response.

The same document limits the architect’s submittal review. Section 4.2.7 says the architect reviews submittals “only for the limited purpose of checking for conformance with information given and the design concept expressed in the Contract Documents.” An AIA Community Hub article from September 2025 puts the consequence plainly: the contractor remains responsible for compliance with the contract documents regardless of the architect’s interaction with the submittals. So an agent that checks your submittals before they go out helps you carry your own responsibility. It does not move it.

That gives you a clean rule for review:

Step	Agent does	Human does
Incoming sub RFI	Searches documents, drafts reply with citations	Project engineer verifies citations, sends
Outgoing RFI	Drafts question with references and proposal	PM reviews and sends to design team
Design team response	Optional: drafts for their review	Architect or engineer decides and signs
Submittal pre-check	Builds a spec-by-spec compliance list	Reviewer judges equivalents and substitutions
Anything affecting cost, schedule, or safety	Flags it	PM or superintendent decides

Three more controls I’d insist on.

Treat incoming files as untrusted. Subs send PDFs and emails. OWASP’s guidance on prompt injection says indirect injection happens when a model reads hostile content in files or web pages. The mitigations include separating external content from instructions, least-privilege access, and human approval for high-risk actions. An RFI agent should be able to read and draft. It should not be able to send, approve, or change the log on its own.

Log everything. Keep the question, the documents retrieved, the draft, the edits, and who sent it. If a dispute comes up, you want that record.

Measure the misses. Track every draft a person had to correct and why. That list is your accuracy report.

Build versus buy for a small team

A small GC has a few people doing document control alongside other jobs. That settles most of the build-or-buy question.

Buy when: the RFI log already lives in a construction platform and the vendor offers document search or drafting inside it. You get integration with your documents, permissions, and logs without owning any infrastructure. Check the vendor on three things: does every answer cite section and sheet, can you export the log and drafts, and what happens to your documents. My vendor question list is built for that conversation.

A caution on vendor claims. Some vendors publish accuracy figures like 95% or higher. Few say how they measured it, on what documents, or against what answer key. One vendor page for an RFI agent that I read (May 2026) says project managers “review, edit, and approve every response,” which is the right posture. It published no accuracy numbers, timelines, or pricing. Ask for a test on your own documents, not a demo on theirs.

Build when:

• Your documents live somewhere the platforms do not reach, such as a mix of shared drives and email.
• You have a specific workflow the products do not do, such as your own subcontractor prequalification rules or a spec cross-check tied to your standard details.
• You need the data to stay in your own environment. See private AI agents in your own cloud for patterns.

Don’t build a general RFI platform. You will lose to products with a whole team behind them. Build a thin agent for the lookup-and-draft step, and keep the log where it already is.

For the general framework, read build vs buy for AI agents. The short version for a small team: buy the platform, build the gap, and keep the build small.

Data and access needs

What data do you need? Less than you think, but it must be clean.

Documents

• Current specifications as searchable PDF or text, with section numbers intact
• Current drawing set with sheet index, plus the sheet and revision tracking
• Addenda, bulletins, and ASIs, tied to the sheets and sections they change
• Approved submittals and the submittal log

History

• Past RFIs with their answers, if your team can export them. They show how your design team phrases decisions and are a good test set.

Access

• Read-only access to the document repository
• Read access to the RFI log, with write access for drafts only
• No send permission on email in the pilot

Two data problems come up most often. The first is versions: the agent answers from a superseded sheet. Fix this with a rule that only the current issued set is searchable, and that every citation shows the revision. The second is scanned documents. If your specs are scans, text extraction quality sets the ceiling on everything else. My post on document extraction with LLMs vs OCR covers how to test that before you build on it.

Pilot scope

Here is a safe pilot. It is a plan, not a result from a past project.

Scope: one active project, incoming sub RFIs only, specs only. Drawings are read-only references: the agent names the sheet and detail but does not interpret them.

Setup:

1. Load the current specification and addenda for one project.
2. Take 30 to 50 closed RFIs from that project or a similar one. These are your test set, and you know the right answers.
3. Run the agent on them. Score each draft: correct, partly correct with the right citation, wrong, or no answer.
4. Check that citations point to the right section.

Live phase: the agent drafts on new incoming RFIs. A project engineer reviews every draft. Nothing goes out without a person.

Measure:

• Share of drafts accepted with no or light edits
• Share of drafts with a wrong or missing citation (target zero wrong, tolerate some missing)
• Minutes saved per RFI, from the engineer’s own timing
• RFIs closed without going to the design team

Stop rules: stop if a wrong answer reaches a subcontractor, or if reviewers spend more time fixing drafts than writing them. Expand only after the test set and the live phase both hold up. My one-week AI pilot method shows how to keep scope this tight.

Bottom Line

Agents are good at lookup and drafting. They are not good enough at drawings to work alone, and the contract keeps the decision with people.

Buy first. Build only the thin gap. Require citations on every draft and a human on every send.

My recommendation: test on 30 to 50 closed RFIs from your own documents before you spend on anything else.

FAQ

Can AI answer RFIs? It can draft answers when the answer is in the specs or drawings, and it can cite the section and sheet. It cannot give the binding response. Under AIA A201-2017, the architect interprets and decides on written request. Use AI to draft and to pre-screen, with a person approving.

Is it worth building an agent for RFIs? For most small GCs, not at first. Check what your current platform offers, test it on your own documents, and build only for gaps it leaves. Building a full RFI system is rarely worth it. A thin lookup-and-draft agent can be.

What review step stays? A person verifies every citation and sends every response. The design professional’s answer to an RFI stays with the architect or engineer. For submittals, the contractor remains responsible for compliance, so someone on your team still reviews each one.

What data do I need? Current specs with section numbers, the current drawing set with a sheet index, addenda tied to what they change, the submittal log, and past RFIs with answers. Version control matters most. The agent must only search the current issued set.

What is a safe pilot? Start with one project, incoming subcontractor RFIs, and specs only. Test on 30 to 50 closed RFIs first. Then run live with a project engineer reviewing every draft. Set stop rules in advance, such as any wrong answer reaching a sub.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect