Goal Drift: What's Actually Risky About A.I. Agents
A.I. Excellence in Construction: Research BriefingsShow Notes
Episode summary: The risk profile for A.I. in construction has changed. In a recent study of 847 production agent deployments, hallucinations were only 9.3 percent of observed vulnerabilities. We look at what dominates the rest, and what construction governance needs to do about it. What you'll learn:- Why the biggest agent failures in production aren't hallucinations, with the categories that actually dominate making up nearly two-thirds of the observed risk surface.- What goal drift, tool misuse, and state manipulation mean in plain language, with construction examples that make the new categories concrete.- How a sequence of individually safe tool calls can compose into a harmful outcome, illustrated with a worked example of the kind of chain construction firms are wiring up.- Why governance built for chatbots checks individual outputs, and why agents need governance built for sequences instead. Who this is for:- Construction executives, innovation/digital leaders, operations, and IT/security teams who want a pragmatic way to prioritize A.I. pilots. Full report & free sample: https://www.placersolutions.io/product/agi-in-construction
More from A.I. Excellence in Construction: Research Briefings
Episode summary: Nearly 6,000 executives report almost no productivity gain from A.I. while seven in ten of their firms use it. Four new studies trace the missing hours to checking A.I. output. This episode reads that research through the 2026 A.I. Excellence in Construction report to show where the checking lands inside a contractor.
Episode summary: Three new 2026 studies test whether A.I. makes people lazy learners. This episode walks through what they found and applies it to how construction firms onboard new project engineers, using examples from the 2026 A.I. Excellence in Construction report.
Episode summary: AT&T now routes 45 billion A.I. tokens a day to the cheapest model that meets its quality bar and reports cutting costs by as much as 90 percent, with open models taking a growing share. This episode asks what that shift means for a contractor whose A.I. arrives inside Procore, Copilot, or a startup's product.
Episode summary: Three contractors at the A.I. Excellence in Construction Virtual Summit, held in collaboration with AGC of America, explained how they decide which A.I. platforms to commit to, wait on, bypass, or replace with something built in-house. Their answers differ. Their conclusion about which layer a contractor must own does not.
Episode summary: Vendors, CEOs, IT directors and project managers each want something different from A.I., and construction firms are getting pulled four ways at once. Builders in the Placer Solutions A.I. peer groups describe what happened when a consultant's central hub pitch met 30-plus disconnected systems, and what actually moved things forward.