Artificial
intelligence: the invisible infrastructure

The acquisition mechanics reveal strategic intent beyond the headline figure. AECOM wasn't buying software (Consigli's "Autonomous Engineer" platform claims 90% reductions in engineering time), it was buying the development team, the architectural decisions behind the product, and critically, preventing competitors from acquiring the same capability. Internal justification reportedly framed Consigli as an "existential threat" if acquired by rivals. That language from a $16 billion global firm is instructive.

This creates a consolidation dynamic with uncomfortable implications. If Tier 1 firms conclude that proprietary AI capability is competitively essential, acquisition becomes the fastest path to capability. Mid-tier firms with genuine AI implementation and not just licensed tools, but custom-built capability and AI-literate teams then become acquisition targets. Firms without either acquisition capital or proprietary capability face squeeze from both directions. They are generally unable to compete for major contracts requiring AI delivery, and increasingly attractive only for asset value in distressed consolidation.

The investment thesis is straightforward but uncomfortable. AI implementation requires upfront capital, data infrastructure, governance frameworks, specialist talent with returns that materialise over 18-24 months. Meanwhile, AI-enabled competitors are already capturing productivity gains that compound quarterly. The gap between investment required and time to payback creates a cash flow challenge that varies dramatically by firm scale.

The case for AI in construction rests on demonstrable productivity gains that the industry desperately needs. UK construction productivity has lagged 13.5% behind the economy average for two decades (Construction Leadership Council, 2024). With labour shortages intensifying and clients demanding faster, cheaper, safer delivery, AI has moved from advantage to competitive necessity.

An increasingly common scenario: a contractor deploys AI-powered project management software to analyse bid documents and flag risks. The system misses a critical requirement buried in technical specifications. A claim follows, triggering a professional indemnity (PI) policy review — and the insurer discovers undisclosed AI deployment. Cover is disputed. The contractor faces an uninsured liability.

Most businesses assume their existing insurance covers AI-related losses. It often doesn't. Here is what each policy traditonally covers and what fall through the gap

Traditional PI policies were written for human judgment errors, not algorithmic outputs. Whilst some policies remain silent on AI, many now include exclusions for claims related to AI usage in the design process. The liability landscape is fragmented: cyber insurance may cover data breaches caused by AI; general liability might address AI-caused physical injury; professional liability could apply to design errors or none of them might, depending on precise policy language. RIBA has been explicit on this point: architects hold PI insurance and assume liability for all information produced using AI, because AI is not an entity that can be held liable. The same principle applies directly to contractors with design obligations. In short: you deploy it, you own the outcome.

One dimension the market has not yet fully absorbed is the insurance position for AI is liable to change significantly as underwriters gain loss experience. Those who have not yet faced claims may find renewal conversations in 2026–2027 substantially more demanding than those in prior years.

Practical implications: organisations should disclose AI usage during underwriting, review policies for AI-specific exclusions, and consider whether separate AI liability coverage is warranted. Assumptions that existing PI or cyber coverage provides adequate protection may prove dangerously wrong when tested.

Here is where promise meets peril. In finance or law, text is reality. A contract is the deal. But in construction, text is merely a proxy for physical truth. A daily report can be incomplete. A submittal can reference outdated specifications. An RFI response can contain buried qualifiers that fundamentally alter its meaning.

Large language models excel at reading project records and producing confident-sounding summaries. The danger lies in conflating well-written answers with ground truth. When an AI assistant synthesises project documentation and reports that a foundation design is compliant, it may be drawing from an earlier revision that predates critical geotechnical updates.

This risk becomes most acute in work that disappears once covered: foundations, post-tensioning, fireproofing, critical MEP routing. These are precisely the areas where verification matters most and where AI-generated confidence can be most misleading.

There is an assumption embedded in almost every AI governance framework currently circulating in the construction sector: that keeping a human in the loop is sufficient to maintain accountability. That assumption deserves scrutiny.

Consider what "human in the loop" looks like in practice. An AI system drafts a project update, a contract notice, or an RFI response. A human reads it and sends it. The recipient’s AI processes the document, generates a response, which a human on that side sends back. The human has not disappeared, but their role has changed. They are no longer the author or decision-maker. They are the physical interface between AI systems: the relay.

This matters for two reasons specific to construction:

1. The first is liability. Under PI frameworks, under the Building Safety Act’s competence requirements, and under standard of care obligations in most building contracts, people remain the responsible party. The AI that drafted the notice carries no liability. The professional who pressed send carries all of it, regardless of how little genuine judgment they applied.
2. The second is the feedback loop problem. If both parties to an exchange are using AI systems trained on similar construction data and industry conventions, errors and flawed assumptions can propagate and reinforce rather than being caught. The human relay, approving a confident-sounding output before hitting send, is not a safeguard against this.

The question for construction leaders is not whether their people are using AI. It is whether the human oversight they believe exists is genuine, or whether it has become a procedural formality: the click of a button that preserves the appearance of accountability without the substance of it.

Construction generates massive volumes of unstructured data in photographs, site notes, emails, change orders, RFIs, all spread across disconnected systems. AI models are frequently trained on this fragmented and inconsistent data. The construction industry's fragmented nature compounds this: data acquisition difficulties, retention problems across project handoffs, and quality inconsistencies that multiply as information moves through the supply chain. The results can be counterproductive: precision without accuracy.

Consider schedule forecasting. An AI model trained on historical project data from multiple sources might predict completion dates with apparent confidence. But if that data includes projects where delays were coded inconsistently, weather impacts were not documented, or scope changes were buried in email threads rather than formal change orders, the model has essentially learned to be precisely inaccurate.

The unsexy truth is that effective AI requires investing in data standardisation, classification systems, and governance frameworks before deployment. Organisations that skip this foundation discover that AI accelerates their existing data problems rather than solving them. Building this infrastructure is not a technical project, it is a strategic imperative that determines whether AI investment delivers value or accumulates technical debt.

AI has been positioned as the answer to the capacity crisis. But when AI handles complex analytical tasks, expert roles can drift toward passive monitoring. Two problems follow: experts lose the engagement that builds judgment, and when AI fails, organisations discover they have lost the manual expertise needed to diagnose the problem. The longer-term consequence is rarely discussed. If AI absorbs the lower-tier tasks that have historically been the entry point for graduates and apprentices, the pipeline of future expertise narrows precisely at the moment the industry can least afford it. Automation may solve today's capacity problem whilst quietly compounding tomorrow's.

Successful firms view AI as amplifying and augmenting expertise, not replacing it. They develop upskilling programmes that move workers into AI oversight and exception-handling roles. They preserve critical judgment whilst leveraging AI for routine tasks. Treating AI purely as a headcount reduction mechanism is a category error. The organisations that blend AI's speed with experienced builders' judgment will consistently outperform those that conflate automation with augmentation.

None of this argues against AI adoption. Its clear the competitive pressure is real, and the productivity gains are demonstrable. But sustainable deployment requires infrastructure most firms have not yet built.

1. Establish governance before deployment
   Create clear ownership of AI initiatives with structured accountability frameworks. Develop AI inventories tracking what is deployed, where, and for which categories of decision. Assess risks systematically: which applications are high-stakes versus lower-risk? Ensure new tools integrate with existing data governance rather than creating parallel, ungoverned structures.
2. Invest in data infrastructure first
   Define data quality standards covering accuracy, completeness, consistency, timeliness, and accessibility. Implement classification systems with clear stewardship on who owns which data sets, who can access them, and under what conditions. Build continuous monitoring for data drift. Without this foundation, AI models are built on sand.
3. Review contracts and insurance before deployment, not after
   Review all current and prospective contracts for AI-specific clauses covering liability allocation, data usage, intellectual property ownership, and disclosure requirements. Examine insurance policies for AI exclusions and coverage gaps. Disclose AI usage during underwriting. These reviews should happen before deployment and not when an incident forces the conversation.
4. Address sub-contractor and supply chain AI risk
   AI-related liability does not stop at the main contract boundary. If AI is deployed by sub-contractors, particularly those with design responsibility, the contractor may bear liability for outputs it did not directly produce. Contracts should address who may deploy AI, what training and proficiency standards apply, and how liability flows down the supply chain. Where AI represents a significant project component, collateral warranties from key AI-deploying sub-contractors may be appropriate.
5. Maintain human oversight for critical decisions
   Crucially, AI should inform decisions and not make them autonomously particularly for safety-critical functions, financial commitments, or contractual obligations. Ensure experts remain genuinely engaged with underlying work, not merely monitoring AI outputs. This preserves the judgment that makes AI an amplifier rather than a liability.

In 2026, construction firms face a question that will define the next decade: are you building AI capability that compounds into sustainable advantage or accumulating technical debt disguised as innovation?

The organisations that will lead are not necessarily those deploying AI fastest. They are asking harder questions first. Do we have the governance structures in place before deployment? Is our data reliable enough to produce trustworthy outputs? Do our contracts and insurance actually cover AI-related risks? How do we preserve expertise whilst leveraging automation? And can we demonstrate compliance with the Building Safety Act, data protection law, and emerging AI standards if we are ever challenged? These questions do not generate headlines. They do not fit neatly into transformation narratives. But they are the difference between AI that genuinely transforms operations and AI that creates expensive problems that are only visible when it is too late to fix them cheaply.

Your competitors are making this choice right now. Some are building foundations of data governance, contractual clarity, insurance alignment, workforce strategy. Others are deploying tools without infrastructure, accumulating liabilities they do not yet recognise. The gap between these two approaches widens every quarter.

When a client asks about your AI governance framework, when an insurer requests your AI usage disclosure, when a tender requires demonstrable compliance with the Building Safety Act's data requirements, when a claim tests whether your AI-assisted decisions were properly validated will you have built the infrastructure, or will you be explaining why you did not?

Navigating AI adoption in the built environment requires more than technical expertise. It demands strategic foresight, robust governance frameworks, and the ability to identify uncomfortable questions before they become expensive problems. Trinity works with clients across the built environment to build the foundations for AI deployment that is commercially sound, legally defensible, and operationally resilient. Our advisory practice specialises in helping organisations plan and implement for successful AI adoption: