Construction Is Spending Millions on AI. Full Integration Remains Rare.

Illustrative stock photograph by Daniel McCullough on Unsplash. Not a Suffolk project or a photograph of the surveyed firms.

New research from BRG puts numbers on what many owners already sense. Nearly seven in ten North American respondents reported spending at least $250,000 on AI over the previous two years, and 37 percent reported spending $1 million or more. Yet only 15 percent said AI was fully integrated across their projects and operations. [BRG]

The money is real. Full integration remains uncommon. That gap is the story.

That distinction matters. Incomplete integration does not mean no integration, and the survey also reports meaningful benefits. The question is whether firms are connecting those benefits to a consistent operating model.

Where the money goes, and what it buys

Adoption now spans pre-construction (46 percent), construction (49 percent), post-construction (40 percent), and business operations (43 percent). Planning and feasibility leads among individual functions at 58 percent, rising to 80 percent among builders. [BRG]

Early project work is a logical place to start. Decisions made during planning shape risk, scope, cost, and delivery strategy. Better analysis at that stage can matter more than simply producing the same document faster.

Respondents report gains in effectiveness, meaning better quality and outcomes, more often than efficiency across most functions. BRG also reports widespread self-reported double-digit cost savings, although those findings should not be read as independently audited reductions in total project budgets. [BRG]

The workforce story is equally telling: 11 percent reported reducing headcount in certain areas because of AI, while 18 percent added it. More than half changed their reliance on outside vendors, nearly half changed team structures, and almost four in ten reskilled existing employees. In this sample, AI is reorganizing work more often than reducing headcount. [BRG]

The obstacles name the real problem

When BRG asked what could limit AI’s benefits, respondents cited lack of skilled talent (40 percent), limited relevance to the work (38 percent), and data and integration challenges (37 percent each). More than a third identified data or integration as a constraint. [BRG]

My reading is that these findings point beyond a technology-purchasing problem. They point to an operating problem.

AI does not automatically resolve fragmented documentation, inconsistent workflows, or teams running the same process in different ways. Without deliberate process design, it can reproduce those weaknesses at greater speed. A firm that cannot get project teams to document decisions consistently should not expect an AI tool reading those records to produce consistent organizational intelligence.

Technology can help improve the process. It cannot substitute for deciding what the process should accomplish, who owns it, and how its outputs will be checked.

The governance findings sharpen the point. Liability for errors in AI-generated work concerns 43 percent of respondents, and unclear treatment of AI in contracts concerns 40 percent. Yet only 42 percent say their policies address human oversight and accountability, and 34 percent say they address accuracy and reliability. [BRG]

Those figures do not prove that every firm outside those groups lacks an AI policy or produces unreliable work. They do identify an important control gap. If AI-assisted information enters the project record without clear review responsibility, traceability, and retention practices, apparent productivity can become avoidable exposure.

Governance before automation. Define the responsibility before accelerating the work.

A useful example: build the foundation

Suffolk offers a useful comparison. ENR reports that its leadership committed more than $100 million to data, technology, and innovation initiatives beginning more than a decade ago. Its structured construction-data environment now holds approximately 293 terabytes. The company’s Jobsite of the Future initiative embeds AI engineers on jobsites, concentrating on design, scheduling, and process improvement. [ENR]

This is not proof that every contractor needs Suffolk’s budget, nor that its operating model can simply be copied. It illustrates a principle: meaningful AI deployment depends on the information, people, and operating practices surrounding the tool.

The lesson is not to spend more before starting. It is to build a reliable foundation while implementing targeted uses that address defined needs.

The question for owners

The BRG findings raise a familiar risk: buying AI as though a product purchase can resolve an unresolved process. Before approving the next investment, ask whether project information is reliable enough, responsibilities are clear enough, and the intended outcome is defined well enough for the tool to help.

For firms still working toward integration, I recommend the following:

Audit a defined workflow end to end. Pick meeting minutes, RFIs, or change management. Map how teams actually work today, not only how the manual says they should. Identify information gaps, handoffs, and review responsibilities.
Agree on a usable standard with the people doing the work. Establish core templates, naming conventions, approval paths, and process ownership. Preserve justified project-specific flexibility. Test the standard before scaling automation.
Assign responsibility for AI-assisted output. Define a named human reviewer for documents entering the project record, the checks required, and how approval is documented. Address confidentiality, approved systems, and escalation in the same operating standard.
Measure effectiveness alongside efficiency. Establish a baseline before deployment. Track documentation completeness, review quality, rework, and decision reliability alongside cycle time and administrative effort.

The firms that do this work will be better positioned to connect AI spending to defensible operational results. The objective is not a larger collection of tools. It is a more capable organization and a better project outcome for the owner.

Research basis and limitations

Benchmarked: All survey figures above are self-reported findings from BRG’s July 2026 research, published October 9, 2026 and publicized October 14, 2026. The survey included 109 executives involved in their organizations’ AI efforts, including 65 in North America. Unless otherwise identified, the report’s headline percentages concern the North American respondents. These findings are not a census of the construction industry and do not establish that AI caused every reported outcome. Suffolk’s investment and data-volume figures are reported by ENR. Interpretations and recommended actions are the author’s analysis.

References

BRG. AI in Construction: Measurable Results, Unclear Benchmarks. October 2026. See also the publication page and October 14 press release.

Engineering News-Record. Suffolk CEO John Fish on AI and the Jobsite of the Future. Interview by Jeff Yoders.

Terry Devlin, PMP, CCM, is the founder of Devnua, an advisory and implementation business focused on organizational improvement, program leadership, team development, documentation systems, and responsible technology integration. View his professional background.