Construction RFI automation: AI made RFIs cheap to write, not cheap to answer

On 14 September 2026 two construction software vendors announced, in separate press releases on the same morning, that their AI now writes into Autodesk Forma’s RFI log. Primepoint creates an RFI or issue in Forma in one click from its drawing-review AI. SmartPM pulls Forma’s RFI and submittal data into its schedule analytics. Procore’s Digital Coworker starter pack has included an RFI agent since July.

Construction RFI automation, as sold this year, automates the part of the RFI that was already cheap: writing it. RFI response time, RFI tracking, the log that ends up in front of an arbitrator — the expensive side — is untouched. The tools just made the queue easier to fill.

TL;DR: AI tools now draft a construction RFI in a click, and none of the September 2026 announcements from Primepoint, SmartPM or Procore quotes a single figure for what happens after the click. The benchmark that does exist — Navigant Construction Forum’s April 2013 study of 1,083,807 RFIs on 1,362 Aconex projects — puts the cost on the response side: 796 RFIs per average project, a median reply of 9.7 days, roughly eight hours and $1,080 per RFI to receive, log, review and respond, about $859,680 per project. Faster creation without response ownership, a contract clock and a dedup check amplifies that queue, and the RFI log is a legal record, so every generated RFI is discoverable. The working version treats the log as the data contract between contractor, coordinator and design team. The agent drafts into it. A named person owns the answer.

What shipped on 14 September, and what didn’t

Read the two releases for what they automate.

Primepoint, out of San Mateo: users “can create RFIs or issues in Forma in one click directly from Primepoint,” it “automatically posts submittal reviews to Forma,” and drafts “write back to their Forma projects only when a person approves them.” The platform “traces every answer back to the document it originated from,” so an RFI can arrive pointing at the sheet it came from. That last part is the good part. An RFI that cites the sheet and detail it’s asking about is the RFI Navigant wanted in 2013.

SmartPM, out of Atlanta: RFI and submittal data from Forma Build flows “directly into SmartPM’s schedule analytics engine,” and its CEO Michael Pink frames the problem as “knowing which RFI, which submittal, or which delivery date to prioritize” becoming “a moving target.” Reading, not writing.

Procore’s RFI agent, per its own support documentation, takes what you typed and suggests “a subject, question, cost impact, and schedule impact where applicable,” and a Find Related Items button asks the AI whether “your question has already been answered.” Open beta. “Trust but verify.”

Neither 14 September release contains a number. Not one. Procore’s July release has one, and it’s about submittals, not RFIs: Haskell’s Director of Innovation, Hamzah Shanbari, is quoted saying the firm “reduced submittal reviews from seven days to just 10 minutes.” Reading a spec against a submittal is a bounded, document-in-document-out task, and I believe that figure. It’s also the wrong side of the RFI to be measuring.

The cost was never in the writing

Navigant Construction Forum published the largest RFI data set I know of in April 2013: 1,362 projects started between 2001 and 2012, all run on Aconex, all with at least 100 RFIs, 79% of them in Australia and New Zealand. It’s old and it’s skewed. It’s also the only benchmark with a million rows behind it, and the structure of the cost hasn’t changed since — an RFI is still a question a contractor writes and a design professional has to answer.

Average of 796 RFIs per project. 9.9 RFIs per $1 million of construction cost across the 826 projects between $5 million and $5 billion; 17.2 per million on the $5–50 million jobs. Average first reply 6.4 days, median reply 9.7 days, approaching 10 on projects of two years or more.

Then the part vendors don’t put in press releases. Navigant asked three US design and construction-management firms what an RFI costs to handle and landed on roughly eight hours: four administrative (receive, categorize, log in and out) at a median $82 per hour, four technical at a median $188 per hour. $1,080 per RFI. Times 796 is 6,368 hours and $859,680 for the average project in the set. Navigant flags every one of those hourly numbers as an estimate for the purpose of the report. Fine. Halve it and it’s still the biggest line item in the process, and none of it is on the creation side.

Step	Who does the work	What the September 2026 tools touch	Navigant, April 2013
Spot the gap or conflict in the documents	Field engineer, sub, PE	Primepoint reads the drawing set	—
Write the RFI	PE / PM	Procore agent suggests subject, question, cost and schedule impact; Primepoint one-click into Forma	—
Check it wasn’t already answered	RFI coordinator	Procore Find Related Items (beta)	13.2% of RFIs “not justifiable”; ~800 of 4,000 in one claim were resubmissions
Log, classify, route by discipline	Coordinator	Sync into Forma; SmartPM reads it	~4 admin hours at $82/hr
Technical review and response	Architect / engineer	Nothing	~4 hours at $188/hr; first reply 6.4 days, median 9.7
No reply, ever	—	—	~19% (Middle East) to 35% (Asia) of RFIs

The “what the tools touch” column is my read of the three vendors’ published descriptions, not their claim. Navigant’s hourly costs are its own estimates from a three-firm survey.

The last row is the one to sit with. Navigant calls the no-reply rate “a leading indicator of fundamental problems with the project, such as an inefficient RFI management response system or the absence of adequate controls and procedures governing the processing of incoming RFIs.” A fifth to a third of questions, never answered. That isn’t a drafting problem. Nobody owned the answer.

The log is evidence, and the agent just made it easier to fill

Here’s the operator’s problem with cheap RFIs. The RFI log isn’t a to-do list. It’s the record both sides bring to the claim.

Navigant walks through one: a laboratory building at a wastewater plant, a delay and cumulative-impact claim resting on “more than 4,000 RFIs.” On analysis, nearly 500 were submittals or shop drawings the contract already required. About 1,700 were routine correspondence — meeting minutes, safety reports, confirmations of phone calls — filed as RFIs. Some 150 were substitution requests for the contractor’s convenience. About 200 were responses to the owner’s own non-conformance notices. And 800 or so “turned out to be the same RFI, previously asked and responded to, but resubmitted with slightly different wording and a different RFI number.” The analysts called those fishing expeditions. Fewer than 700 of the 4,000 were RFIs at all, and of the ones that took more than a week to answer, several already had settled change orders attached.

That’s a 2013 case, done by hand, on a project where writing an RFI cost someone twenty minutes. Now put a drafting agent in front of a field engineer who has a question every time a detail doesn’t close. The agent proposes a subject, a question, a cost impact, a schedule impact. One click and it’s in Forma with a number. The 2012 University of Wisconsin–Madison highway study Navigant cites found 13.2% of RFIs “not justifiable” — means and methods, design changes, or answers already in the documents — which at Navigant’s averages is 105 RFIs, 840 hours and $113,400 per project spent answering questions that shouldn’t have been asked. Lower the cost of asking and that share doesn’t fall.

Every one of those generated RFIs is discoverable. Navigant’s proposed defence against a “too many RFIs” claim is a log where every entry is classified on receipt — answered-in-documents, design coordination, means-and-methods, duplicate — because “such a system of categorizing RFIs provides a documented defense.” The classification is the control. The vendors have automated the step before it.

This is the same shape as contract review in construction: the readable, pre-signature layer got cheap, and the obligation-tracking layer that generates the disputes stayed manual. Agents amplify what the process already does. If the process produced 800 duplicate RFIs by hand, it will produce more with help.

What the working version looks like

Not a better drafting model. The drafting is fine.

The RFI log becomes the data contract between the contractor, the RFI coordinator and the design team, and the agent is allowed to write into it only through that contract. Navigant’s minimum form from 2013 is most of the schema already: spec, drawing or detail reference; disciplines and trades affected; schedule activities it may impact; priority one to five; the information requested and why; the contractor’s proposed response; a yes/no cost-impact flag. Primepoint’s sheet reference and Procore’s suggested impacts are those fields being filled by software, which is the right use of it.

Three fields the 2013 form predates. A responder who is a person, not a firm — “the architect” answers nothing; a named engineer with a queue does. A due date computed from the contract clock. Navigant’s model clause gives the owner ten working days, and the clock starts “on the date stamped received ‘In from Contractor’” — working days, from the stamp, not calendar days from the record’s created-at timestamp. Compute it the second way and every RFI that crosses a weekend is due two days early, the dashboard is red for the wrong reason, and within a month nobody looks at the dashboard. And a classification field, required at creation, with duplicate-of as one of the values, so the dedup happens before the number is issued rather than in a claims analyst’s spreadsheet three years later.

Then the response side gets the automation the creation side got. Route by discipline to the named responder. Alert at day seven of ten, escalate at ten, and surface Navigant’s other metric — the no-reply rate — weekly by responder, because it’s a measure of the design team’s process, not the contractor’s. The 4,000-RFI claim would have been a 700-RFI log with 800 rejected duplicates and a classification on each one, which is a different conversation in front of an arbitrator.

An agent drafting into that log, with a person approving before it’s recorded — Primepoint’s design, and the right one — is a genuine step removed. An agent drafting into a log with no responder, no clock and no dedup is a faster way to build the plaintiff’s exhibit. The difference isn’t the model. It’s whether the process was written down before the agent showed up, and it’s the kind of process and integration work that decides which of the two you get.

The question got free. The answer still costs eight hours, and now there are more of them.