McKinsey’s vision for an agentic AI construction industry: 4 takeaways

Here are four headline takeaways from the report.

AI has the potential to automate 50% of nonphysical work in the architecture and engineering sectors and 39% of nonphysical work in the construction sector, the report proposes.

The firms who will be able to profit from this most are those that control three things: proprietary project data, the workflows where decisions get made, and the ability to charge for outcomes rather than hours.

The idea here is that as AI boosts productivity and tames risk, firms can switch from selling capacity to selling “excellence with fewer surprises, more reliable schedules, lower risk, and better project outcomes”. This will allow them to finally adopt an outcome-based business model, with fixed fees, milestone payments, shared savings, risk-monitoring services, and performance-linked incentives.

The data matter because they can continuously train the AI agents. “A firm that systematically connects estimates to actual results, schedules to real-world progress, design choices to constructability outcomes, and risk decisions to actual claims data can build a self-improving system,” the report observes.

Interestingly, because the data matter so much, commercial friction between technology vendors and construction firms could arise. “As AI capabilities improve, technology vendors are pushing harder to access project data and the rights to learn from it and reuse it to create new products, or to put guardrails around how their products or software can be used to train proprietary models,” the report notes.

“They are also increasingly competing to control workflows, learning loops, and business opportunities built on top of project data. Firms that overlook the details of vendor agreements risk giving away too much control and long-term advantage. When establishing vendor relationships, firms must be careful about whether they can easily move their data, keep it separated from other customers, or leave a platform without losing information so they can retain the benefits of their AI adoption efforts.”

McKinsey places the above scenario some years away. To prepare for that, in the next 18 months, firms should start deploying AI agents on a small handful – three to five – of the trickiest construction workflows that depend most on expert human judgement and experience. Estimating, constructability reviews, and scheduling are examples the report gives.

Additionally, firms should connect these AI led or enabled activities to the whole project and not treat them as isolated optimisation tools. A schedule agent won’t be much use if it’s not connected to procurement, design changes, cost forecasts, field conditions and more.

“The firms that succeed with AI in the near term will be the best at orchestrating agentic work across workflows,” the report says. It adds that firms can start this on projects now; there’s no need to wait for a new one.

The report sets out three levels of AI autonomy – none, some and all – firms should consider in deploying agents on workflows.

The lowest level is human-led but agent-enabled. This is where an engineering manager, say, leads the entire design delivery process with agents generating analyses, reviews and recommendations.

Next level up is agent-led and “human-accountable”. Here, an agent produces a project bid, calculates contingency levels and identifies commercial risks, while the proposals manager approves and takes responsibility for it.

At the top is complete agent autonomy for a given workflow. They continuously collect schedule cost and project data, update forecasts and automatically issue instructions.

This is more long-term when, the report says, “AI will increasingly connect design, planning, logistics, equipment, and site execution into a more automated operating system”.

McKinsey acknowledged that what it called a “quieter challenge” lurks in this future, however appealing it might be on paper.

Turning a construction project into an automated operating system will require embedding the expertise that now resides in the heads of a few experienced people into the agent-led workflows themselves.

The potential problem is that when you automate all the routine, back-office procedures – tasks that junior staff use now to build their own experience and judgment – how will we grow new experts? Or will we even need experts?

McKinsey worries we do. “Firms may need to train junior staff more intentionally through structured reviews, simulations, explicit standards, and exposure to real project failure cases,” the report says.

“Otherwise, they risk creating a generation of workers who learn how to operate AI tools without developing the judgment, pattern recognition, and risk awareness needed to lead complex projects and handle difficult situations in which clients expect a human to own the consequences.”

Keep up to date with DC+: sign up for the midweek newsletter.