Last month I watched a demo of an AI scheduling tool that promised to “predict delays before they happen.” Impressive deck. The presenter showed a 250-day delay averted on a £4.1 billion rail tunnel in London. The AI had flagged risk patterns hidden in a database of 750,000 historical project schedules and £120 million in potential overruns. The audience clapped.
I raised my hand. “How many of those 750,000 schedules are single-family residential?”
Long pause. “Our data set focuses on capital projects.”
Right.
The Machine Learning Data Problem Nobody Mentions
nPlan, the London-based AI scheduling platform, has trained its models on 750,000 historical project schedules representing over $2 trillion in construction spend. The sectors it covers: highways, rail, public buildings, commercial construction, heavy infrastructure, oil and gas, utilities, data centers. Its AI currently manages $500 billion in active projects. Those are serious numbers.
Now look at who builds your house.
A productive custom home builder completes 10 to 20 homes per year. After 15 years in business, that’s 150 to 300 completed project records. A large regional production builder doing 100 homes annually for two decades has 2,000. Even D.R. Horton, the nation’s largest homebuilder, closes roughly 90,000 homes a year. They’d need a decade just to approach the data volume nPlan used from day one.
Machine learning models don’t care about your intentions. They care about sample size, variance, and signal-to-noise ratio. When nPlan identifies that sewer lining work on a London rail project carries a 73% probability of exceeding baseline duration, that prediction rests on thousands of comparable tasks across hundreds of similar projects. When your builder says “framing usually takes three weeks,” that estimate rests on gut feel and a spreadsheet nobody updated since 2019.
Custom Homes Break the Pattern
The data problem runs deeper than volume. Production homebuilders repeat the same floor plans across subdivisions. Their schedules are transferable because the work is standardized. A three-bedroom ranch in Phase II is close enough to the same house in Phase I that historical duration data actually predicts something.
Custom homes aren’t that. A cantilevered deck on a hillside lot in Marin has almost nothing in common with a slab-on-grade colonial in suburban Dallas. The soil is different. The structural requirements are different. The subcontractor pool is different. The permit timeline is different. You can’t train an algorithm on one and expect it to predict the other any better than a coin flip.
This is why nPlan’s listed sectors don’t include residential. Not because they forgot about it. Because the statistical foundation doesn’t exist.
What Your Builder Actually Has
The residential scheduling tools that builders actually use tell a different story than the AI hype cycle suggests.
| Tool | Scheduling Features | AI Prediction |
|---|---|---|
| Buildertrend | Drag-and-drop Gantt charts, task lists, calendars | None |
| CoConstruct | Basic scheduling, task assignments | None |
| Procore | Advanced project controls, resource management | Commercial-focused analytics |
| nPlan | AI schedule risk analysis, forecasting | Yes — for $100M+ capital projects |
Buildertrend and CoConstruct are the tools most residential builders can afford. Useful for organizing the chaos, sure. But they’re glorified digital whiteboards. They tell you what should happen next Tuesday. They can’t tell you the probability that it won’t. I’ve watched a $1.2 million custom build in Westchester slip six weeks because a tile subcontractor took a bigger commercial job and didn’t mention it until the day before rough-in. No scheduling tool on earth flags that.
The Actual Delay Numbers
Meanwhile, the residential scheduling problem remains brutal. Industry data from 2024–2025 shows 98% of construction projects face delays, with the average project duration running 37% longer than originally projected. The 2023 SORCI Report from the Association of Professional Builders found that 35.1% of residential projects ran late, and an average $750,000 build costs the builder approximately $670 per day in delay expenses.
Worse: 48.3% of builders don’t even know their fixed expenses per job per day. Half the industry can’t calculate what delays cost them.
Census Bureau data from 2024 pegs the average single-family home at 9.1 months start to finish: 1.4 months for permitting, 7.6 months for construction. A 37% overrun on that 7.6-month build phase adds 2.8 months. At $670/day, that’s roughly $57,000 in delay costs per house, absorbed entirely by the builder or passed to the buyer as a change order that breeds resentment and Yelp reviews.
What Would Actually Work
I’ve managed residential projects for two decades. The honest answer is that the scheduling problem in custom homes isn’t a prediction problem. It’s a coordination problem. Your framing crew is ready. Your electrician ghosted you. The inspector canceled. The trusses arrived a week late because the lumberyard prioritized a commercial order. No algorithm trained on London rail tunnels can predict that your plumber’s truck broke down on I-95.
What actually moves the needle for a 15-person residential operation:
Weather-integrated scheduling. Connect a weather API to your critical-path schedule. If there’s a 70% chance of rain on your pour day, the system bumps it automatically and notifies the concrete sub. Open-source tools like n8n can wire this up with Claude or GPT and Slack alerts for under $50/month.
Subcontractor reliability scoring. Track which subs hit their dates and which don’t. After 30 projects, you have enough data to rank your electricians by on-time percentage. That’s not machine learning. It’s a spreadsheet. But it works better than any AI tool that has never seen your market.
Rule-based cascade alerts. If the foundation pour slips three days, auto-reschedule framing, notify the truss supplier, and push the rough-in dates. This is if-then logic, not artificial intelligence. Buildertrend can do most of it with task dependencies. Most builders don’t bother setting them up because it takes effort upfront and the payoff isn’t obvious until you’re already behind.
The Counterargument: Just Pool the Data
An industry consortium could theoretically aggregate scheduling data from thousands of residential builders into a shared training set. Pool 10,000 builders at 15 homes/year each, and you’d have 150,000 new project schedules annually. Within five years, you’d approach nPlan’s dataset volume.
The argument has surface appeal. In practice, it hits three walls. First, builders view their scheduling data as competitive intelligence and won’t share it. Second, inconsistent data formats across dozens of scheduling tools would require massive standardization. Third, and this is the hard one: custom home schedules carry so much project-specific variance (lot conditions, local subcontractor markets, jurisdiction-specific permitting timelines) that aggregation may not produce the transferable patterns that machine learning requires. A framing duration in Phoenix means something different than a framing duration in Minneapolis in January.
nPlan’s own case study material acknowledges this implicitly. Their SCS JV partnership on the HS2 rail project identified 140 risk insights across a single £4.1 billion program. That density of insight is possible because mega-infrastructure has repeatable task patterns at scale. Residential doesn’t.
What This Doesn’t Prove
The $57,000 delay cost estimate is a napkin calculation: $670/day (SORCI 2023, based on average $750K builds) multiplied by 85 delay days (37% overrun applied to the Census Bureau’s 7.6-month average build time). It uses a general construction delay rate, not a residential-specific one, because residential-specific overrun data is scarce. The 37% figure comes from Buildern’s 2025 metrics report covering all construction types. Residential-only delays may be shorter or longer. The $670/day figure reflects 2023 costs and hasn’t been adjusted for 2026 inflation, which would push it higher. I couldn’t find a rigorous residential-only overrun study published in the last three years. That gap in the data is, itself, part of the story.
nPlan’s dataset size and sector coverage are from their public website as of March 2026. I have no access to their internal model accuracy metrics for residential-adjacent projects, if any exist.
The builder data volumes (150–300 schedules for a custom builder, 2,000 for a regional production builder) are my estimates based on typical annual completions per NAHB builder-size surveys, not a formal census. The average NAHB member builds fewer than 25 homes per year; the numbers I used are generous.
The conclusion — that AI scheduling prediction doesn’t work yet for custom residential — is a data-availability argument, not a claim that the technology is fundamentally incapable. If someone builds the dataset, the math changes. Nobody has built it yet.