I watched a project engineer named Sarah spend most of last Tuesday doing something that would have baffled anyone outside construction. She had a 1,200-page specification document open on one monitor and an Excel spreadsheet on the other, manually scanning each section for submittal requirements, copying them into her log, cross-referencing CSI division numbers, and flagging which ones needed shop drawings versus product data versus samples. By 4 PM she had gotten through Division 07. She still had Divisions 08 through 33 ahead of her.
Down the hall, her company had just installed a $45,000 AI-powered scheduling platform that promised 20% schedule compression through predictive analytics and resource optimization. Nobody in the office could explain what it was actually doing that their old Primavera setup did not, but the sales demo had involved a drone, so leadership was excited.
This is the state of AI in construction in 2026. We have taught machines to fly over job sites, drive excavators without operators, detect hard hat violations from security camera footage, and generate 3D point clouds from photographs. What we have not taught them to do is the thing that actually eats Sarah's week.
Where the Research Money Actually Goes
A 2025 study from the University of North Florida swept 24,978 AI-in-construction research papers published between 2004 and 2025. They sorted every paper into application domains. Sustainability and environmental performance led with 8,096 papers. Sensing, monitoring, and safety: 6,343. Digital design technologies: 6,145. Cost, scheduling, and productivity: 4,179.
Administrative and management processes, the category that includes submittals, RFIs, change orders, transmittals, meeting minutes, and compliance documentation: 215 papers. Less than one percent of all AI construction research.
Read that again. The work that consumes a documented 14 hours per week of every construction professional's time receives less attention from AI researchers than any other category. Not by a small margin. By a factor of 30 compared to the next-smallest domain.
The UNF researchers called it a "structural bias." AI research follows sensor data because sensor data is abundant and publishable. Drones produce photogrammetry datasets. Cameras produce safety violation images. IoT devices produce time-series data. All of it maps neatly to machine learning pipelines that produce impressive-sounding accuracy numbers in conference papers.
Submittals, RFIs, and change orders produce Word documents, Excel logs, and PDF specifications. Less photogenic. Harder to publish. Equally devastating to project timelines when mishandled.
The RFI Numbers Are Brutal
The Navigant Construction Forum analyzed 1,362 construction projects using data from ACONEX, examining 1.1 million individual RFIs. The average project generated 796 RFIs. Each RFI costs approximately $1,080 in direct costs when you factor in the time to write it, route it, wait for a response, review the response, and distribute the answer to affected trades.
That is $859,680 in RFI processing costs on an average project. Not the cost of the design changes the RFIs reveal. Just the cost of asking and answering questions about the drawings.
It gets worse. Twenty-five percent of RFIs receive no response at all, per Procore's data. A quarter of the questions your project team carefully documents, routes through proper channels, and waits on simply vanish into someone's inbox. On a residential custom home running 40 to 80 RFIs, that means 10 to 20 questions go unanswered, which means 10 to 20 decisions get made in the field by someone who is guessing.
Guessing does not show up on the schedule. It shows up six months later as a change order.
What $177 Billion Looks Like at Project Scale
Rhumbix documented that field foremen spend five to eight hours per week on paperwork alone, translating to $260,000 to $820,000 in lost productivity annually per firm. Paper-based systems create 15 to 25 percent revenue leakage. That is not a typo. A quarter of potential revenue, gone to bad data flow.
The PlanGrid/FMI research attributed 48 percent of all rework to poor communication and 22 percent to poor project information, which together account for 70 percent of the rework that, according to the American Society of Concrete Contractors, averages 10 percent of original contract value. On a $500,000 custom home, that is $50,000 in rework, $35,000 of which traces back to people not knowing what they needed to know when they needed to know it.
Not a technology problem. A paperwork problem. A filing problem. A "did anyone ever answer RFI number 47" problem.
AI Can Already Do This. Almost Nobody Uses It.
The same UNF study built a prototype tool that extracts submittal requirements from construction specifications. It processes a 1,000 to 1,500 page spec document in six to eight minutes at a cost of less than ten cents per file. It generates transmittal-ready Excel and Word documents. It lets project teams query the extracted content through a conversational interface.
Autodesk shipped AutoSpecs, which does something similar within their Construction Cloud platform. Datagrid's AI agents automate RFI routing, deadline tracking, and documentation assembly. Buildxact's Blu AI generates residential construction estimates in 30 seconds using local market data.
These tools exist. They work. They cost less per month than the scheduling software nobody in Sarah's office can explain. And almost nobody in residential construction is using them, because the entire industry conversation about AI has been about robots and drones for so long that the mundane stuff never gets a seat at the table.
Why the Exciting Stuff Wins Every Time
I have been managing projects for twenty years, and I have watched this pattern repeat across every wave of construction technology. The demo that gets the budget is the one with the video. Drones flying over a roofline. A robot laying bricks. An excavator driving itself across a graded lot. Leadership watches the video, writes the check, and posts about it on LinkedIn.
Nobody shoots a video of a project engineer processing submittals 40 percent faster because an AI extracted the requirements from the spec automatically. Nobody posts about an RFI response time dropping from 11 days to 3 because an AI flagged the overdue items and routed them to the right reviewer. Those outcomes are invisible because they prevent problems instead of solving spectacular ones, and preventing problems has never been a marketable skill in this industry.
That invisibility is also the strongest counterargument to what I am saying here. Procore, Autodesk BIM 360, and PlanGrid have been digitizing construction workflows for years. The tools keep getting better. Adoption keeps being terrible. Only 20 percent of construction firms use mobile tools effectively for collaboration, per the PlanGrid/FMI data. The bottleneck might not be that nobody built the AI. The bottleneck might be that a 55-year-old superintendent does not want to change how he processes RFIs regardless of what software you put in front of him.
That is a real objection. I have met that superintendent. I have been that superintendent. But the cultural resistance argument is also a convenient excuse for not building the tools in the first place, and the 215-out-of-24,978 research gap suggests the AI community has not tried very hard to find out whether better tools would change adoption.
What This Means If You Are Building a Home
If you are hiring a general contractor for a custom home in the $400,000 to $1,000,000 range, ask what software they use to manage submittals, RFIs, and change orders. If the answer involves the phrase "Excel spreadsheet" or "we email those around," understand that 14 hours per week of your contractor's project management capacity is going to administrative overhead that proven tools could cut in half. That is time not spent on your project.
Ask specifically whether RFI response times are tracked and what the average turnaround is. If nobody knows, that is your leading indicator for how many field decisions on your home will be made by someone who is guessing because the architect never answered the question.
If you are a small builder running three to five projects: the submittal extraction tools are real, they are cheap, and they will give your project engineer back two days per month. The RFI tracking and routing tools are real and they will cut your unanswered rate from 25 percent to something you can actually defend in a warranty dispute. Neither of these will make a good LinkedIn video. Both of them will make your projects run closer to schedule.
Limitations
The 796-RFI-per-project average comes from a 2013 Navigant/ACONEX dataset skewed toward large commercial projects in Australia and New Zealand. Residential custom homes generate fewer RFIs, typically 40 to 80, but the per-RFI cost dynamics are similar. The $177.5 billion waste figure from PlanGrid/FMI encompasses all U.S. construction, not residential specifically. The UNF bibliometric analysis covers papers indexed in Scopus through early 2025 and may undercount recent documentation-AI work published in trade journals or preprint servers. Adoption rates for document automation tools in residential construction are not systematically tracked by any industry body, so "almost nobody uses them" reflects trade anecdotes and vendor penetration data, not a rigorous survey. The submittal extraction prototype was tested on ten specification documents, an informative but small sample.