RFI Automation: From 45 Minutes to 5

The average RFI costs a thousand dollars. Not in materials. Not in equipment. In labor. A project manager sits down, pulls up the spec, cross-references two or three drawings, writes a question that the architect will actually answer, formats it, attaches the right documents, routes it to the right people, and submits it through Procore. That process takes 15 to 45 minutes depending on complexity. On a mid-size commercial project, multiply that by 200 to 400 RFIs.

Those numbers come from a Navigant study that analyzed 1,300 construction projects and over a million RFIs. The same study found that 22% of all RFIs are never answered. Not rejected. Not closed. Just ignored. The primary reason: lack of a formalized process. The RFI was vague, incomplete, or routed wrong, so it sat in a queue until someone forgot about it.

That's $31 billion in annual rework attributed to miscommunication. $4.2 billion from poor document control alone. And the people who know exactly what's wrong in the field are the ones least equipped to write the RFI that fixes it.

The actual problem isn't writing

A superintendent standing at grid line C-4 on the second floor can tell you exactly what's wrong. The HVAC duct conflicts with the structural beam. There's six inches of clearance where the spec calls for eighteen. The mechanical contractor can't reroute without hitting the return air path. He knows all of this. He's looking at it.

But writing that into a formal RFI with the right spec section number, the correct drawing references, a clear single question, and proper routing? That's a different skill set. So the super describes the problem to the PM in a text message or a Teams chat. The PM translates it into an RFI sometime that afternoon. Or the next morning. Or never, because twelve other things landed on their desk.

The information exists. It's sitting in a chat thread. The bottleneck is translation: turning conversational field language into a document that an architect will take seriously and respond to quickly.

What we built

RFI.msg is an AI tool that takes plain-language input and produces a spec-grade RFI. You describe the issue the way you'd describe it to a coworker. Drop a photo if you have one. The AI figures out what's missing, asks one targeted follow-up question, and drafts an RFI that reads like your best PM wrote it in five minutes.

Not a template. Not a form with dropdowns. A conversational interface that understands construction vocabulary and knows what a complete RFI looks like.

The pipeline works in five stages:

Intake. The AI parses your message and extracts structured fields: location, issue description, trades affected, references mentioned. If you upload a photo, it describes what it sees and pulls context from the image.
Completeness check.Every RFI needs a specific question, a location, spec references, drawing references, and a description of the existing condition. The AI scores what you've provided against what a good RFI requires. It only asks about critical gaps. If you gave it enough, it skips straight to drafting.
Draft.The AI writes the RFI using a writing skill that enforces active voice, construction vocabulary, specification-grade precision, and one question per RFI. No “please advise.” No “it has come to our attention.” No passive hedging.
Rewrite. A second pass runs the draft through a slop filter. Sixteen blacklisted phrases trigger an automatic rewrite. Any sentence over 40 words gets split. Any paragraph over four sentences gets broken up. The result reads like a human wrote it, because the rules were derived from how the best PMs actually write.
Submission. The finished RFI goes directly into Procore with correct workflow routing, assignees, and attachments. It posts as you, not as a bot. The architect sees your name on it.

What the slop filter catches

AI-generated text has a smell. Anyone who's read enough of it can spot it in two sentences. “It is important to note that a discrepancy has been observed between the structural and mechanical drawings.” Nobody on a jobsite talks like that. Nobody should write like that either.

We built a blacklist of phrases that trigger an automatic rewrite:

“It has come to our attention” becomes a direct statement of the issue
“Please advise” becomes the actual question you need answered
“This discrepancy” becomes the specific conflict, named
“The aforementioned” becomes the actual noun
“In order to” becomes “to”
“It should be noted that” gets cut entirely

The naturalness test: read the RFI out loud. If it sounds like something you'd hear in a deposition, rewrite it. If it sounds like something a PM would say standing on the slab pointing at the problem, keep it.

Before and after

Here's what the AI prevents, and what it produces instead.

Before (typical AI output without our rules):

It has come to our attention that there appears to be a discrepancy between the structural drawings and the mechanical drawings at the second floor. The beam appears to conflict with the HVAC ductwork. Please advise on how to proceed with this matter.

After (RFI.msg output):

At grid line C-4, second floor, the W12x26 steel beam (S-201) conflicts with the 24”x18” HVAC supply duct (M-301). Field measurement shows 6” of clearance where spec section 23 05 00 requires 18” minimum. The duct cannot be rerouted without impacting the return air path to the east wing.

Question: Should the mechanical contractor lower the duct routing below the beam, or should structural provide a revised beam detail to restore the required 18” clearance? Mechanical reroute affects 3 additional hangers and 2 days of schedule.

The second version has a location. Spec reference. Drawing references. A measurable condition versus a measurable requirement. An impact statement. A suggested resolution with schedule implications. One clear, answerable question. That RFI gets a response. The first one sits in a queue.

Why completeness matters more than speed

We could have built a tool that just writes faster. But speed without quality creates a different problem: more RFIs that don't get answered. The Navigant data is clear. The 22% that never get a response aren't slow. They're incomplete. They're vague. They bundle three questions into one submittal. They reference “the drawings” without saying which sheet.

Every element in our completeness scoring has a weight. Specific question and location are required. Spec and drawing references are high priority. Impact statements are medium. Suggested resolutions are low priority but the Navigant study found they significantly reduce response time when included. The AI enforces single-question discipline. If you bundle two issues, it splits them.

If the draft scores below threshold on completeness and quality, the tool refuses to submit. It tells you what's wrong in plain language. You can override it, but we track the override rate. A high override rate means the tool needs calibration or the user needs training. Either way, the data tells you something useful.

What this doesn't do

RFI.msg is a writing tool, not an engineer. It will not tell you how thick the concrete should be. It will not recommend structural modifications. It will not cite building codes as design advice. When someone asks an engineering design question, it says so: “That's an engineering design question. I'm not the right tool for that.”

It also won't suggest moving installed steel or re-pouring concrete. If a beam is already erected, the suggested resolution needs to work around it, not through it. We call this the constructability gate. The AI checks whether its suggestion is physically possible given what's already built.

The math

A project manager earning $85 an hour spends 30 minutes on a standard RFI. That's $42.50 per RFI in direct labor. On a 300-RFI project, that's $12,750 in PM time just writing and formatting RFIs. Cut that to 5 minutes each and you save $10,625. On one project.

But the real savings aren't in writing time. They're in response time. A complete, well-referenced RFI gets answered faster. A faster answer means fewer delays. Fewer delays mean less rework. The FMI/CURT study puts the industry-wide cost of rework from miscommunication at $31 billion per year. You don't need to fix the whole industry. You just need your RFIs to be the ones that get answered.

Try it

The web demo is live at contractor-ai.com/rfi-demo. No account required. Describe a construction issue in plain language and see what comes back. If you've written a few hundred RFIs in your career, you'll recognize the difference immediately.

Procore integration is coming next. The full version submits directly into your project, as you, with correct routing and attachments. MS Teams integration after that, so your field teams can draft RFIs from the chat platform they already use.

We're building this because we've written thousands of these things by hand and know exactly how much time gets wasted on formatting when the real value is in the information. The field team already knows what's wrong. The tool just needs to write it down properly.