The Construction Data Problem: Why AI Needs Clean Data to Work

AI is hard to implement in construction because project data is fragmented across tools, paper and PDFs, and much of it is inaccurate, incomplete or inconsistent. Autodesk and FMI estimated that bad data may have cost the industry $1.85 trillion in 2020. The construction data problem is the main reason many AI pilots stall before production.

Our view: the data problem is real, but the 2026 survey evidence says it is a systems problem more than a cleanliness problem. Firms that get AI into production talk about integration first. That changes where a contractor should spend the first dollar. For the wider picture of what is deployed and what is marketed, see AI in Construction: The State of the Industry in 2026.

What counts as bad data in construction?

Autodesk and FMI define bad data as information that is inaccurate, incomplete, inaccessible, inconsistent or untimely, so that it cannot support decisions. That five-part definition is useful because each failure looks different on a jobsite.

• Inaccurate: a quantity carried forward from a prior revision.
• Incomplete: a daily log with no weather or headcount.
• Inaccessible: a submittal approval sitting in one person's inbox.
• Inconsistent: the same cost code used two ways on two projects.
• Untimely: an RFI answer that arrives after the work is in place.

Most of these are ordinary workflow gaps rather than technical failures. They only become "AI problems" when software tries to read the record and finds nothing reliable there.

How much does bad data cost the industry?

The best-known estimate comes from the Autodesk and FMI study, which surveyed more than 3,900 professionals. It put the cost of bad data at $1.85 trillion globally in 2020 and attributed 14% of all rework to decisions made on bad data. Thirty percent of respondents said more than half of their project data is bad. The study was published in 2021, so treat the dollar figure as a 2020 estimate rather than a current measurement.

One caution on the rework figure: published summaries of the study disagree on the dollar total, so the 14% share is the number to rely on.

Why is construction harder for AI than other industries?

Construction has structural features that make data hard to standardize. None of these is new, and none is the fault of any one firm.

1. Every project is a prototype. Each job has a different owner, design team, site and delivery method, so records rarely line up across projects.
2. Temporary organizations. Teams assemble for one project and disband. Data conventions are set by whoever is in the room that month.
3. Documents, not databases. The contractual record is mostly drawings, specifications, proposals and PDFs, which are made for people to read.
4. Many parties, many tools. Owner, architect, GC and trades each keep their own systems and rarely share a schema.
5. Paper is still present. Bluebeam's 2025 survey of more than 1,000 AEC firms across five countries found only 11% fully digital, with 52% still using paper during design and 49% during planning, according to Bluebeam's release.

Takeoff tools show the practical effect. ConstructConnect lists low-resolution scans, handwritten markups, unusual symbols, multi-sheet revisions and incomplete or conflicting plan sets among the conditions where AI takeoff struggles. Those are data-quality problems, not model problems.

Is data quality or integration the bigger obstacle?

Here the survey evidence complicates the popular story. Among firms with AI in production, Bluebeam's July 2026 survey of 750 US technology decision-makers found the top success factor was integration with existing systems at 47%. Cross-team collaboration followed at 29%, and high-quality data at 26%. Ninety-seven percent said tools must fit existing workflows. That survey was published by the vendor, so read it as one market participant's data.

Two readings fit these numbers. One is that firms already in production have cleaned up enough to stop worrying about it. The other is that respondents underrate data quality because its cost is spread across rework and delay rather than a line item. The survey cannot tell us which. What it does show is that firms getting value describe plumbing, meaning how data moves between tools, as the decisive work.

What does the data problem mean for productivity?

McKinsey figures reported by the wire service ANI in July 2026 put construction labor productivity growth at 10% in total (about 0.4% a year) from 2000 to 2022, against 90% (3% a year) in manufacturing. The same coverage said AI and automation could unlock about $228 billion in annual value in US AEC by 2030 and automate 39% of nonphysical work in construction. We did not access the underlying McKinsey report, so check it before quoting these figures in a board deck.

Those are projections. They also assume the inputs exist. A model cannot automate a coordination task if the RFI history, drawing revisions and cost data it would learn from are scattered or unreliable.

Illustrative example: one RFI log, two outcomes

Illustrative example: a contractor wants an AI tool to draft RFI responses from past projects. On Project A, every RFI has a spec section, discipline, status and closing date in a single system. On Project B, half the RFIs live in email threads, with free-text subjects and no closing date. Assume the tool can learn from A and cannot read B at all.

The tool's output will look strong in the demo, because the demo uses Project A. The pilot then runs on Project B and disappoints. Nothing about the model changed; the records did. That gap, between curated demo data and the real project record, is the most common way the construction data problem shows up in practice.

What are firms doing about it?

Evidence of specific data-infrastructure programs at named firms is thin in the sources we reviewed, so we describe patterns rather than case studies. Common moves include consolidating workflows into a single platform, standardizing templates for logs and submittals, and choosing AI tools that sit inside systems the team already uses. Product direction fits this: Procore's Agent Builder, announced as generally available on October 15, 2025, lets firms build agents for RFIs, submittals and daily logs on top of data already inside the platform. The approach sidesteps the integration problem by working where the records live.

Platform control cuts both ways. ENR reported that Trunk Tools was denied Procore API access, a reminder that data access can be restricted by whoever owns the system of record. Contracts for any AI tool should say who can read, export and reuse project data.

What this means for GCs, owners and designers

General contractors and subs. Pick one workflow and fix it before buying anything. RFIs are a good candidate because the fields are few, the cycle time is measurable and the history is reusable. The measurement side is covered in How to Measure the ROI of AI in Construction.

Owners and developers. Owners control what gets handed over at closeout. Specifying structured, consistent turnover data in the contract is the most direct lever an owner has, and it benefits facility operations as well as any future AI use.

Architects and engineers. Naming conventions, sheet indexes and model hygiene are data quality. A drawing set with inconsistent sheet titles will degrade every downstream reader, human or machine.

A short data-readiness checklist

Before a pilot, check the following for the workflow in question:

• Is there one system of record, or are records split across email and spreadsheets?
• Are status values and naming conventions defined and enforced?
• Can you export the data in a structured format?
• Who owns data quality for this workflow?
• What share of records is complete today? Measure it.

Bad inputs also raise error risk. Language models can produce confident but wrong output when fed unreliable text, which is the subject of AI Hallucination in Construction.

What to watch through 2027

• Whether the share of firms running AI in production rises above Bluebeam's 28%, and whether integration difficulty (15% as a barrier) climbs as more firms try to connect tools.
• Whether platform vendors open or restrict data access, following the Procore and Trunk Tools dispute.
• Whether independent studies replace the vendor-sponsored surveys that dominate this topic. Almost all current numbers come from vendors or sponsors.
• Whether owners begin writing structured-data turnover requirements into standard contracts. We found no AIA AI guidance as of 2026, so the contractual language is still unsettled.

For adoption patterns at larger contractors, see Why Large GCs Are Slow to Adopt AI.