When the Algorithm Draws the Line

In late 2026, a regional leadership team at Skanska—one of the world's largest construction and infrastructure contractors—convened in a glass‑walled project office overlooking a congested highway corridor in northern Europe. Below them, excavation had begun for a multi‑billion‑dollar urban tunnel and interchange project that would reroute traffic, unlock new development parcels, and upgrade aging utilities beneath a dense metropolitan district.

The project—an EPC contract jointly delivered by Skanska and a consortium of partners—sat at the intersection of several intensifying pressures:

• Complex underground works in variable geology and high groundwater
• Tight political deadlines tied to elections and regional development plans
• Ambitious safety and carbon‑reduction commitments demanded by the city and lenders
• An aging workforce, thin margins, and chronic talent shortages in critical engineering roles

To manage this risk, Skanska's corporate digital team had pushed hard for an integrated "AI‑augmented delivery stack" on the project: automated design checks on the tunnel linings, AI‑assisted clash detection and sequencing based on the BIM model, generative scheduling tools that suggested optimized construction sequences, and computer‑vision‑driven safety monitoring on site. The project had been selected internally as a "lighthouse" for the company's global AI strategy.

But six months into main works, the lighthouse was throwing off mixed signals.

The Promise of an AI‑Augmented Jobsite

The project was, on paper, a perfect candidate for advanced digital delivery. The contract's risk profile hinged on control of:

• Ground conditions and water ingress
• Interface clashes among utilities, tunnel segments, and highway structures
• Safety risks in confined spaces and under live traffic
• Schedule risk from any stoppage in tunneling operations

The digital/AI stack included:

1. Design and Engineering Support

   • AI‑assisted design rule checks on tunnel segments and reinforcement layouts, trained on prior Skanska projects and relevant codes.
   • Automated clash‑detection routines on the 3D BIM model, pushing prioritized "issue lists" to design managers.

2. Planning and Scheduling

   • A generative scheduling engine that proposed alternative tunnel drive sequences, shift patterns, and logistics plans to minimize bottlenecks and idle time.
   • Scenario tools that could simulate the impact of equipment breakdowns, surprise ground conditions, or access constraints.

3. Site Operations and Safety

   • Computer‑vision modules for PPE compliance, exclusion‑zone violations, and unsafe proximities between workers and heavy plant.
   • Automatic daily site logs generated from video, sensor data, and voice notes, feeding into dashboards for senior management.

4. Knowledge Capture

   • A central data platform that captured RFIs, design changes, non‑conformance reports, and incident investigations—intended to become a reusable "playbook" for future underground projects.

Corporate leaders framed the project as proof that Skanska could move beyond fragmented "digital pilots" toward an integrated, human‑centered augmented intelligence model—something that would differentiate the firm with clients, improve margins, and mitigate growing risk on complex urban work.

Early Results—and Early Friction

Within the first months:

• The AI‑assisted clash checks did flag several non‑trivial conflicts between utility diversions and temporary works designs, avoiding potential rework.
• Safety analytics highlighted a pattern of near‑misses during evening shifts in one access shaft, leading to revised traffic management and lighting.

These wins appeared in slide decks sent to group management and even in discussions with prospective clients.

But on the ground, a different narrative was emerging.

Site engineers and foremen began to complain that:

• The issue lists from the AI‑assisted clash detection were long, opaque, and constantly shifting as models updated—creating "alert fatigue."
• Junior engineers were relying heavily on the generative schedules and automated quantity checks, spending less time walking the site, checking drawings, or running manual calculations.
• Some experienced supervisors felt their judgments were being second‑guessed by dashboards and algorithms produced by people who "never crawled through a wet heading in their lives."

In one internal debrief, a veteran tunnel superintendent remarked:

"Twenty years ago, if a young engineer signed off on a sequence, I knew they'd sweated through the logic. Now half of them are just validating whatever the tool spits out. When things go sideways, they don't have the mental map to improvise."