Construction AI: capital, compute, and the case for embedded agents

On 4 May, Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced an AI-native enterprise services company designed to embed engineers and Claude inside mid-market firms to redesign workflows and integrate AI into core operations. Anthropic, Blackstone and Hellman & Friedman are each contributing roughly $300m, with Goldman Sachs at $150m, alongside Apollo Global Management, General Atlantic, Leonard Green, GIC and Sequoia Capital.

The target market is explicit: PE-owned mid-market firms in healthcare, manufacturing, financial services, retail and real estate. Many UK contractors, consultants and supply chain firms either are PE-backed today or sell to companies that are. This venture will set the pace and the language those clients use to talk about AI.

The May 2026 reading from PBC Today reports that 28 per cent of UK construction project professionals now describe AI as fully embedded in their workflows - supporting data-driven insights (24 per cent), administrative work (22 per cent), predicting project outcomes (21 per cent), resource allocation (21 per cent) and risk forecasting (20 per cent). 91 per cent of firms expect to increase AI investment this year, with project delivery (41 per cent) and profitability (36 per cent) cited as the biggest expected gains.

Those numbers describe a market where embedded AI use has crossed from experimental into operationally normal. They also describe a 72 per cent majority that has not embedded it yet - the gap is widening, not closing.

A Beazley-commissioned Opinion Matters survey of more than 3,500 business leaders, published 6 May, ranks construction the least prepared industry for cyber threats. Almost a third of construction and property bosses (32 per cent) say they are worried about cyber security. The 2026 Spotlight on Cyber Threats and Tech Advances report explicitly flags that attackers are now using AI to run large-scale automated reconnaissance and phishing - attacks that are faster, more adaptive, and harder to detect or contain.

The relevance to AI adoption is direct. Every agent that touches your project files, your commercial pricing, your supplier comms or your CDE is part of the attack surface. The same connectors that make agents useful - Microsoft 365, SharePoint, Drive, Slack - are the connectors threat actors aim at first.

Construction Safety Week 2026 ran from 4 to 8 May under the theme "All In Together: Recognize. Respond. Respect." Beyond Turner Construction opening up its SafeT Coach AI safety assistant to the wider industry, contractors used the week to push for standardised safety language ("high hazard", "high energy", "STCKY") and visible deployment of AI-powered cameras flagging missing PPE in real time, plus wearable sensors tracking fatigue and posture.

The pattern this year was different from the one twelve months ago: AI safety tools were treated as a routine part of the conversation, not as the headline novelty. That is a useful temperature check on how quickly the industry's expectations have shifted.

At its Code with Claude developer event in San Francisco on 6 May, Anthropic shipped three big upgrades to Claude Managed Agents. Multi-agent orchestration lets a lead agent break down complex tasks and assign them to up to 20 sub-agents across 25 simultaneous threads. Outcomes let you define a success rubric and have the agent work to it - Anthropic reports up to 10 percentage-point gains on harder tasks versus standard prompting. Dreaming runs as a scheduled background process between sessions, reviewing transcripts, finding patterns and curating the agent's memory store.

This is the shape of construction-grade agentic AI. A submittal pipeline, a clash-detection-to-RFI loop, or a programme drift watcher can now realistically be expressed as one lead agent, several specialists, a clear outcome rubric, and learned memory across runs - without you writing the orchestration glue.

Both Jensen Huang (NVIDIA) and Lisa Su (AMD) used CNBC interviews this week to argue that agentic AI is now a primary driver of compute demand - Huang framing it as "fully accretive" to enterprise software revenue rather than a substitute. The construction read-through is unchanged: the AI Growth Zone and hyperscaler pipeline that UK contractors are bidding on is grounded in real, currently scaling demand, not speculative orderbooks.

That has two practical implications. First, any contractor or consultant with credible MEP, civils, grid connection or commissioning capability should expect the data centre pipeline to keep feeding. Second, the same hyperscaler demand is what is funding the agent capabilities your delivery teams are starting to use day-to-day - the supply and demand sides are now coupled.

- CNBC - Jensen Huang on agentic AI as accretive (CNBC Television, "Power Lunch", 6 May 2026)

- CNBC - AMD CEO Lisa Su on AI agent demand (CNBC Television, "Squawk on the Street", 7 May 2026)