Construction AI: the defender models arrive, and the reasoning ceiling shows itself

UK Construction Week London continues today (Wednesday 13 May) from 10:00 to 17:00, with the ConTech & AI Hub running across all three days. Yesterday's first day saw discussion focused on AI-driven decision-making, generative design, automated estimating and the cross-over with the Marketing & Procurement Hub. Today's programming includes the densest concentration of agentic AI and digital twins sessions - and Thursday closes with the Women in Construction sessions and a final pass over the AI-readiness panels.

If you have not been yet, today is the day. The hub is small enough to navigate in a focused two-hour visit if your team has clear questions on procurement-AI, document automation, or digital twin tooling.

Microsoft has partnered with North America's Building Trades Unions (NABTU) to offer free AI literacy courses and industry-recognised credentials to millions of skilled craft professionals across North America. The framing is workforce development for the AI build-out - both for trades who will build the data centres and for the wider construction workforce, where AI is increasingly embedded in project controls.

For UK construction leaders, the takeaway is two-fold. First, expect similar UK-side partnerships to follow - CITB, Constructionarium, the Federation of Master Builders are all natural counterparties. Second, your AI tooling decisions will increasingly be questioned on whether they're supported with structured upskilling, not just licences.

Nous Research's Hermes Agent dropped a desktop app on 9 May, taking the open-source autonomous agent from CLI/server experiment to a long-lived application running locally on the user's machine. The pitch is an always-on agent that can "self-evolve" by editing its own skill set, watching inboxes and queues continuously, and persisting across sessions. Combined with last week's 0.13 "Tenacity" release (multi-agent kanban, /goal, ElevenLabs voice, DeepSeek v4 Pro), Hermes is now a credible self-hosted option for organisations that don't want a cloud-only agent stack.

For UK construction firms with client confidentiality clauses, this matters. An agent that runs locally, persists across sessions and orchestrates multi-step workflows without sending project data to a vendor is a different procurement conversation than a cloud SaaS agent.

Worth a fresh look this week: Claude for Word, launched in April and now available on Pro and Max plans, proposes every edit as a tracked change in Microsoft Word's native review pane. The original text is shown as a deletion, the new text as an insertion, with each change individually acceptable or rejectable. Claude can also read comment threads, understand what they anchor to, edit the anchored passage and reply to the thread explaining what it did.

For construction document workflows - contracts, RFIs, design narratives, JCT correspondence, BSA-related submissions - that is the audit trail that has been missing from most AI-assisted editing. It also pairs cleanly with the CDM/PI accountability conversation: every AI change is signed, dated and reviewable.

OpenAI launched Daybreak yesterday (11 May) as a cybersecurity platform built on GPT-5.5 and the Codex Security agent. Daybreak connects to codebases and infrastructure, simulates attack routes, surfaces vulnerabilities, generates and tests patches inside repositories, and produces audit-ready validation. It is positioned as a direct counter to Anthropic's Claude Mythos (announced April) and Project Glasswing - a partnership that already includes AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks.

The strategic read for construction is sharper than it looks. Last week's Beazley poll ranked construction the least prepared industry for cyber threats. The same week, Palisade Research showed frontier models can self-replicate across vulnerable networks at up to 81 per cent success. Now both major frontier labs are racing to build defender programmes - but those programmes are gated to enterprises in their partner ecosystems. If your stack is Microsoft 365 / Anthropic, you'll want to understand Glasswing. If it's OpenAI / Microsoft Foundry, Daybreak is your route in.

GPT-5.4, Claude Opus 4.6 and Gemini 3.1 all scored exactly 0 per cent on the latest version of the Abstraction and Reasoning Corpus benchmark (ARC-AGI 3), created by François Chollet to test general reasoning that cannot be cracked by memorised training data. Untrained humans score 100 per cent on the same tasks. These same models simultaneously write substantially better code, score higher on professional exams, and handle longer multi-step tasks than their predecessors - so this is a structural limitation, not a model-quality regression.

The construction read-through is straightforward. Frontier AI is genuinely useful for repeatable, pattern-rich, well-documented work - estimating, scheduling, RFIs, document review. It is still beaten by an untrained twelve-year-old on novel abstract reasoning. That is the right line to use in any boardroom conversation about where AI should and should not be making decisions on a project.