Start with the records that already exist
Most conversations about AI in construction start with the thing that does not exist yet: a clean, structured, project-wide data model, ideally attached to a 3D model, ideally updating in real time. That is a five-year program and it is not where the money is. The money is in the records a capital program already produces under contract, by the thousand, in formats nobody chose: cost-loaded CPM schedules, submittal logs, RFIs, daily reports, pay applications, inspection reports, and the closeout package. Those records exist because a contract clause requires them. They are the only construction data with guaranteed coverage.
That framing matters because it decides what is buildable this year. The American Society of Civil Engineers put the overall condition of U.S. infrastructure at a C in its 2025 Report Card and set the ten-year investment gap at roughly $3.7 trillion. The Infrastructure Investment and Jobs Act (Public Law 117-58) pushed about $550 billion in new federal spending into that gap. Every dollar of it runs through owners and contractors who are staffing at capacity, which means the constraint on program delivery is rarely analysis. It is review hours, coordination hours, and the time between a field question and an answer.
Our engineers have spent twenty years building production systems for federal agencies, and the pattern in capital programs is consistent with what we see elsewhere: the useful first project is almost always a document or schedule problem, not a sensing problem. Three record classes carry most of the near-term value, and they are worth taking one at a time.
Data readiness by construction record class
Editorial weighting from public sources and practitioner reading. Illustrative, not a measured statistic.
Schedule data: small files, long memory
A cost-loaded critical-path schedule for a $100 million project typically holds 3,000 to 12,000 activities. Exported monthly across a 30-month job, the entire history is maybe 30 snapshots and well under 100 megabytes of XER or XML. By the standards of machine learning this is nothing. Its value is that it is longitudinal: the same activity, re-forecast thirty times, with the logic, the constraints, and the actual dates recorded each time.
That structure supports a specific and testable question: which activities in the current update are likely to slip in the next one. It is a supervised problem with labels you already own, because every past update tells you what the prior update got wrong. The features that carry signal are unglamorous. Remaining duration relative to original duration. Total float trend across the last three updates. Number of predecessors not yet complete. Whether the activity has a hard constraint. Whether it sits on the longest path. Whether the responsible subcontractor has slipped elsewhere on the same job.
The federal apparatus already gives you a scoring vocabulary here. The Defense Contract Management Agency's 14-point schedule assessment checks for missing logic, negative float, excessive lags, hard constraints, and high-duration activities, and produces a Baseline Execution Index and a Critical Path Length Index. Earned value reporting under ANSI/EIA-748 is required on DoD contracts at the thresholds in DFARS 234.201. A model that predicts slip is far easier to sell when its outputs land in the same language the program office already uses for schedule health.
Submittal review: the highest-volume text problem on the job
Submittals are where the paper lives. On a mid-size vertical project it is normal to see 1,500 to 3,000 submittal packages, organized by CSI MasterFormat specification section, running 25,000 to 60,000 pages of shop drawings, product data, samples, and certifications. Specification Section 01 33 00 usually gives the design team a fixed review period, commonly 14 calendar days. Every day of that clock is a day the fabricator is not cutting steel.
Three things a model does well here. First, routing: read the incoming package and predict the correct specification section and reviewer, which sounds trivial and is not, because submitters mislabel constantly. Second, completeness checking: the spec section states what the package must contain, and a retrieval system can compare the two and flag what is missing before a human opens it. Third, prior-art matching: find the equivalent package that was approved on a past project with the same spec section and pull up the review comments that were written then.
The value is measurable without any modeling sophistication. Track first-pass rejection rate, median days to first response, and the count of packages returned solely for a missing item. A Navigant Construction Forum study of more than 1,300 projects put the average cost of processing a single RFI at roughly $1,080. Reducing avoidable submittal churn moves numbers of that shape, and those numbers are already in the log.
As-built records: the hardest class, and the most valuable
Record drawings, closeout documentation, and O&M manuals are the worst-behaved data in the program and the most valuable over a fifty-year asset life. They arrive as scanned red-lines, as PDFs of PDFs, as an IFC export nobody validated, and occasionally as a COBie spreadsheet with half the columns blank. ISO 19650 defines how information should be managed and handed over across a project. Compliance with it in the field is uneven at best.
Do not promise an owner a perfect digital asset record. Promise something narrower and checkable: for a named set of asset types, extract manufacturer, model, serial, install date, warranty term, and location, with a page-and-region citation back to the source document for every field. That is a deterministic extraction problem with a generation layer on top, and it is auditable. Anything that cannot point at the page that supports it does not go in the register. We wrote about that separation in document extraction that survives an audit.
The same discipline applies to infrastructure asset inventories. The National Bridge Inspection Standards at 23 CFR Part 650 Subpart C set inspection intervals and data submission for the more than 620,000 structures in the National Bridge Inventory, roughly seven percent of which carry a poor condition rating in recent FHWA reporting. The inspection narratives behind those ratings are free text and have been for decades. Reading them consistently is a language problem with a very clear consumer.
Volumes, latency, and what has to be fast
Ranges below are planning figures we use when sizing an engagement on a project of roughly $100 million. Scale them linearly and you will not be far off.
| Record class | Typical volume | Latency the work needs |
|---|---|---|
| CPM schedule updates | 3,000–12,000 activities, ~30 monthly snapshots, <100 MB total | Overnight batch. Nobody re-forecasts in real time. |
| Submittals and shop drawings | 1,500–3,000 packages, 25,000–60,000 pages | Minutes at intake, hours for full analysis |
| RFIs | 800–2,500 items, 5–40 pages each | Seconds for duplicate detection at submission |
| Record drawings and closeout | 2,000–8,000 sheets plus O&M volumes | One-time backfill, measured in weeks |
| Reality capture | 15–30 GB imagery per drone flight; 50–500 GB point cloud | Days. Registration and QC dominate. |
| Equipment telematics (ISO 15143-3) | 1 record per machine per 15 min; well under 1 GB/year per 100 machines | 15-minute polling is plenty |
Notice what is missing from that table: any requirement for sub-second inference. Construction and infrastructure work runs on review cycles measured in days and weeks. A system that answers in four hours against a fourteen-day baseline has already won. Designing for real-time streaming when the business process is weekly adds cost and failure modes for no return. The one place latency genuinely bites is duplicate detection at RFI submission, because the value of telling someone their question was already answered collapses the moment they hit send and move on.
What it costs to run
Text work at these volumes is cheap, and buyers routinely overestimate it by an order of magnitude. Take a 40,000-page submittal corpus. Commodity document OCR at list pricing runs in the low single dollars per thousand pages, so the full corpus is on the order of $60 to $200 to convert once. At roughly 600 tokens per page, 40,000 pages is about 24 million input tokens; one full pass through a mid-tier model at a $3-per-million input rate is roughly $70. Even with three passes, embeddings, and re-runs during development, the inference bill for the entire historical backfill is measured in hundreds of dollars, not tens of thousands.
Storage and hosting are similarly small. A single project's complete document corpus fits in tens of gigabytes. A vector index over 40,000 pages chunked at 500 tokens is roughly 250,000 vectors, which sits comfortably on a single modest instance. Reality capture is the exception: point clouds and imagery are where the storage bill actually appears, and where egress charges will surprise a program that did not plan for them.
The real cost is engineering time and the owner's review time. Budget accordingly: for a first engagement, expect the split to run roughly 70 percent data plumbing and evaluation, 25 percent interface and workflow, and 5 percent anything a vendor would call AI. Any proposal with the inverse ratio is describing a demo.
Data quality decides the outcome, not model choice
Four different taxonomies describe the same scope of work on the same project, and they do not reconcile. The estimator works in cost codes. The scheduler works in a work breakdown structure. The design team works in MasterFormat specification sections. The state DOT works in bid pay items inside AASHTOWare Project. A quantity that moves between them loses fidelity at every hop, and no model recovers what the join threw away.
Specification numbering alone breaks more pipelines than any modeling decision. MasterFormat's 2004 revision moved from a five-digit, sixteen-division structure to a six-digit, fifty-division one. Electrical moved from Division 16 to Division 26. Owners with a long portfolio hold both conventions, often inside the same submittal log, sometimes inside the same project. Schedule data has its own version of this: activity IDs get renumbered between updates, which silently destroys the longitudinal join that makes slip prediction possible in the first place.
- Identity drift – the same subcontractor appears as four entities across the log, the payroll, and the schedule
- Numbering conventions – MasterFormat 1995 and 2004 coexisting in one portfolio
- Renumbered activities – schedule IDs that change between updates and break the history
- Scanned-only records – red-lined as-builts with no text layer and no consistent sheet index
- Date ambiguity – submitted, received, logged, and reviewed dates used interchangeably
- Unit and quantity mismatch – pay item units that do not match the takeoff units
Every one of those is fixable, and fixing them is most of the work. It is also why we insist on measuring against a sealed holdout from the owner's own records rather than a public benchmark. A model that scores well on someone else's clean corpus tells you nothing about a portfolio where a quarter of the record drawings are photographs of paper. This is the step where projects die, and we said so plainly in why AI projects fail at the data access step.
The rules that shape the architecture
Federal and federally funded construction carries constraints that change system design, not just paperwork. They are worth naming before anyone draws a diagram.
Labor records are PII. Davis-Bacon Act prevailing wage requirements (40 U.S.C. 3141 et seq.) and the Copeland Act produce weekly certified payrolls on Form WH-347. The Department of Labor revised that form to require only the last four digits of a Social Security number and to omit home addresses, but the underlying records held by the contractor are full personnel data. Payroll analytics belongs in a segregated store with its own access control, never in the same index as the submittal corpus.
Facility drawings are frequently CUI. Physical security details, utility routing, and structural information on federal facilities are routinely designated Controlled Unclassified Information under 32 CFR Part 2002 and DoD Instruction 5200.48. That decides hosting: a system touching those documents needs an environment with the right authorization boundary, and the boundary has to be written down before ingestion, not after.
Hardware provenance is regulated. FAR 52.204-25 prohibits certain covered telecommunications and video surveillance equipment. The American Security Drone Act of 2023, enacted in the FY2024 National Defense Authorization Act (Public Law 118-31), restricts federal procurement and operation of covered unmanned aircraft and the use of federal funds for them. If a scope includes drone-based progress capture, the airframe question is a compliance question before it is a technical one.
Materials data carries a domestic content test. Build America, Buy America requirements from IIJA Sections 70901 through 70927, implemented at 2 CFR Part 184, apply to iron, steel, manufactured products, and construction materials on federally funded infrastructure projects. Material approval submittals now carry a compliance dimension, which is an extraction target with real dollars attached.
Safety data has a hard line around it
Construction accounts for roughly a thousand fatal work injuries a year in the Bureau of Labor Statistics Census of Fatal Occupational Injuries, more than any other industry. OSHA's construction standards at 29 CFR Part 1926, the recordkeeping rules at 29 CFR Part 1904, and USACE's EM 385-1-1 on federal projects all govern the same subject. There is genuine value in reading safety observations, near-miss reports, and daily field notes at scale to find leading indicators before a lagging one arrives.
The line is this: a model may inform a human, and it may not be the control. OSHA's construction rules assign specific determinations to a competent person, defined at 29 CFR 1926.32(f) as someone capable of identifying hazards and authorized to correct them. That is a person, with authority, on the site. The same principle applies to structural and geotechnical judgment, which carries a licensed professional engineer's seal under state law. Our team includes licensed professional engineers, and their role in an engagement is exactly this: the system organizes evidence, and a licensed human signs.
Scoping a first engagement that can fail cheaply
The goal of a first engagement is not a working product. It is a defensible answer to whether the second engagement is worth funding, delivered before anyone has spent enough to need it to succeed. Six to eight weeks, fixed price, one record class, one closed project, one number agreed in writing during week one.
First engagement, structured to fail cheaply
Two details make the difference between a pilot that informs a decision and one that quietly becomes permanent. The metric is set before any data is seen, so nobody gets to redefine success after the fact. And the holdout stays with the owner. If the first engagement returns a no, the owner has spent a small fixed amount and gained a documented map of their own records, which has value whatever they build next. That is what failing cheaply means.
Pick the record class by how quickly you can label it. Submittal routing is the usual first choice because the historical log is the label set and it is already correct. Schedule slip prediction is a close second for the same reason. Anything requiring new labeling from busy field staff belongs in engagement two, after the owner has a reason to spend that time.
Common questions on scope
Do we need a BIM model first?
No. A federated BIM model is useful and is not a prerequisite for any of the work described here. Schedule risk, submittal review, and record extraction all run on documents and logs. If a model exists and an IFC export validates, geometry becomes a useful join key for locating assets. If it does not, the document work still stands on its own.
Will this replace reviewers or schedulers?
It should not, and pitching it that way tends to end the project. Review capacity is the constraint on most programs, and the point is to spend the same reviewer hours on the packages that need judgment instead of on completeness checks a machine can do. The scheduler still owns the forecast. The system tells them which twenty activities to look at first.
What if the historical records are genuinely bad?
Then the inventory in week two says so with numbers, and the owner decides whether remediation is worth funding before more money goes into modeling. A documented statement that 30 percent of a portfolio has no text layer is a real deliverable. It is much cheaper to learn that in week two than in month nine.
Can this run inside our environment?
Yes, and for CUI-designated facility documentation it generally has to. We build for the authorization boundary the data requires, including deployments where the data never leaves the customer's environment and the model runs where the records live.
Bottom line
Construction and infrastructure data is textual, contractual, and messy, and that is good news for anyone willing to do the plumbing. The volumes are small enough that compute is a rounding error. The review cycles are slow enough that latency is a non-issue. The labels already exist inside logs that contract clauses forced somebody to keep. What decides the outcome is whether the records reconcile, and whether the people scoping the work were honest about the regulatory boundary before they drew the architecture. Start with one record class, measure against a holdout you keep, and make the first engagement small enough that a no is affordable.
Frequently asked questions
Cost-loaded CPM schedule histories, submittal and RFI logs with their attachments, specification sections, and daily reports. These exist on every project because contract clauses require them, so coverage is guaranteed. Record drawings and sensor feeds are more valuable long term and far more variable in quality.
Less than most buyers expect. A 40,000-page corpus costs on the order of $60 to $200 to OCR at commodity list pricing, and roughly $70 for a single full language-model pass at a $3-per-million input token rate. Engineering time, not inference, is the cost driver.
Almost never. Review cycles run in days and weeks, schedule updates are monthly, and telematics polling at 15 minutes is more than adequate. Duplicate detection at RFI submission is the one place where a fast answer changes behavior.
CUI handling under 32 CFR Part 2002 for federal facility documentation, PII controls for Davis-Bacon certified payrolls under 40 U.S.C. 3141 and the Copeland Act, covered-equipment prohibitions under FAR 52.204-25, drone restrictions under the American Security Drone Act of 2023, and Build America, Buy America content rules at 2 CFR Part 184.
Six to eight weeks, fixed price, one record class, one closed project, and a success metric agreed in writing before any data is seen. The evaluation holdout stays with the owner. If the answer is no, the owner keeps a documented inventory of their own records and has spent a small fixed amount to get it.