98 percent of megaprojects run over budget, behind schedule, or both. That figure comes from McKinsey's 2015 research across more than $1 trillion of capital projects, and the industry has quoted it ever since. Bechtel, one of the largest builders on the planet, cited the same pattern as its reason to act. Then it did something most firms still have not. It used its own project data to get ahead of the overruns.
The common story about AI in construction is a story of failure. Pilots that stall. Tools that demo well and collapse on real data. Trust that drops year over year. That story is largely true, and we have written about why it happens.
This is the other half. Five organizations where AI is already delivering real, measured results in construction. Read them together and the same pattern shows up in every one. None of them started with the AI. They started with the data, and with a clear decision about who still has to verify the answer.
Bechtel and Swinerton: seeing the overrun coming
Bechtel took on this problem years before most. Its engineers presented at the 2019 Spark + AI Summit, and Databricks published a case study, describing how the firm unified its data engineering and data science work on one platform and trained deep-learning models on its own project history, after years of algorithms being capped by old hardware. Databricks' write-up put the potential value in the tens to hundreds of millions of dollars. Treat that as a projection, not an audited result, and note the work is several years old. The durable point is the order of operations: the models became possible only after the data and the teams were unified.
The idea is not frozen in 2019, and it is not only for mega-builders. Swinerton, a self-performing general contractor, more recently ran a data-science competition with Avanade to predict which of its projects were trending over budget or behind schedule, using its own historical data. Per Avanade's account, it is a proof of concept rather than a deployed product, with no accuracy figures published yet. But it answers the question that matters for a mid-size firm: can your own past projects tell you which current ones are heading for trouble? The early answer is yes, once the data is in order.
Colas: unify first, build second
Colas builds and maintains roads across roughly 50 countries through hundreds of subsidiaries. In Databricks' Colas customer story, Chief Data Officer Philippe Toublant described the starting point bluntly: the company was an aggregation of hundreds of companies not sharing data, and historically not a data-driven organization.
So Colas did the unglamorous work first. It consolidated 54 separate HR systems just to count its own employees accurately at the group level. Only then did the use cases come, and they came fast. During COVID-19, a team built a site-monitoring app in two weeks using notebooks and a scheduler. It sent daily automated emails and text messages and reported reopening rates across 15,000 construction sites. The Materials and Equipment group now tracks energy-intensive equipment worldwide, sets usage baselines, and cuts energy consumption against them. Next on the list is an AI tool that reviews road video footage to flag maintenance needs.
"Colas built a working app across 15,000 construction sites in two weeks. The speed came from the data unification that happened first.
The two-week app is the part people quote. The 54 systems are the part that made it possible.
Novade: a small team, not a big one
If Bechtel and Colas sound like a budget mid-size firms cannot match, Novade is the counterexample. Novade makes field-management software contractors use for defects, safety, and schedules. Its original homegrown system, built on open-source tooling, could not keep up with growing data volumes or share data securely with customers.
Novade moved to a governed lakehouse with a disciplined structure, raw to refined to reporting-ready, that serves both business dashboards and model training from the same place. On top of it, a small data team trained models that predict worker-safety incident risk and project schedule-delivery risk, and fed live dashboards to customers without copying data out.
Per Databricks' Novade customer story, the reported results were a 60 percent reduction in total cost of ownership, a threefold efficiency gain, and a 100 percent increase in new-client acquisition. The detail that matters for a mid-size reader is the size of the team that did it. Small. The constraint was never headcount. It was whether the data was in order.
Procore and Crux: the data you already generate, made answerable
Two more examples point at what comes after the foundation.
Procore, the project-management system most contractors already run, rebuilt its analytics on a zero-copy sharing architecture, as its team described in a 2025 Databricks Data and AI Summit session. Large datasets can now be shared and joined with other enterprise data without the brittle copy-and-move pipelines and high egress costs that used to make it painful. That is the plumbing that lets a contractor combine Procore data with scheduling, finance, and field data in one place, which is the precondition for any serious prediction.
Crux, a demonstration app built for a 2025 Databricks community challenge, shows what sits on top. A project manager types a plain-English question, such as which projects are highest risk or what is blocking a specific job, and gets an answer that combines structured metrics with cited project documents. In the app's sample project, it reported a job 18 days behind schedule, with an 8.1 percent forecast variance, a 3.9 million dollar projected overrun, 7 overdue tasks, and 1 active blocker. Those figures come from demo data, not a real client, but the mechanism is the point.
Here is the part worth copying. Crux shows the exact database query it ran and the document citations behind every answer. The manager can verify the number before acting on it. The AI does the analysis and shows its work. The human makes the call. That is the boundary we map for every client: AI recommends, your leaders decide.
The pattern behind all five
Put the five side by side and the sequence is identical every time.
First, the data gets unified and governed. Bechtel's teams and pipelines. Colas's 54 systems. Novade's lakehouse. Procore's shareable analytics layer. The AI is never step one.
Second, one clear use case goes first. Overrun risk. Site reopening. Safety prediction. Not a platform-wide rollout. A single problem that matters, proven on real data.
Third, a human stays in the decision. The best of them, Crux, is built so the person can check the AI's work. The tools inform the call. They do not make it.
This is the difference between what we call Path 1 and Path 2. Path 1 buys the tool first and hopes the organization and the data catch up. Path 2 defines the problem, governs the data and the decision rights, and selects the tool last. Every firm in this post ran Path 2, whether they named it or not.
One more thing, because it matters for how you read these stories. The five Databricks examples share that platform, and Swinerton's proof of concept ran on a different stack entirely. The lesson is not "buy Databricks," or any other platform. The same pattern holds regardless. The platform is the last decision, not the first.
The bottom line
AI is working in construction, in specific places, for specific firms. The proof is named and numbered: Bechtel on overruns, Colas across 15,000 sites, Novade with a small team and a 60 percent cost reduction, Procore and Crux making existing project data answerable. What none of them did was start with the software.
They started with questions about which problem to solve, governance over the data and the decisions, and only then the tool. That order is the whole game. We wrote the first version of this argument in what should humans still own when AI enters the work. These five firms are what the answer looks like in practice.
Our Discovery engagement is built for the step before the tool: the questions, the governance, and the Workflow Map™ that shows exactly where AI recommends and where your leaders decide. If you can name the one project outcome you would most want to see two weeks earlier, you already have the start of a pilot worth running.
Humanity for what matters. AI for everything else.