Key Takeaways
- AI is now performing real project management work on capital projects. McKinsey and ALICE Technologies report generative scheduling deployed with more than 35 clients across energy, mining and infrastructure, with schedule accelerations of up to 20 per cent.
- Governance is lagging the technology. McKinsey's 2026 AI Trust Maturity Survey finds average responsible AI maturity at 2.3 on a four-point scale, with only around a third of organisations reaching level three or above in AI governance.
- Accountability is the differentiator. Organisations with a clearly accountable function for AI average 2.6 maturity, against 1.8 for those without one.
- Scepticism is warranted. Gartner forecasts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, citing cost, unclear value and inadequate risk controls.
- AI-generated project data is unverified project data. The governance response is the same discipline capital projects have always needed: named accountability, explicit decision rights, and independent verification before the numbers reach a gate or a board.
In 2019, Gartner predicted that 80 per cent of project management tasks would be eliminated by 2030, as artificial intelligence took over data collection, tracking and reporting.1 Seven years on, that prediction looks directionally right and numerically premature. The routine work of project management is being automated, but the discipline has not shrunk. It has acquired a new team member that produces schedules, forecasts and status reports at a speed no human planner can match, and with a confidence that no human planner should.
Across Energy, Minerals and Resources (EMR) capital projects, this shift has moved well past the pilot stage. Generative scheduling is running on live programmes. Owner organisations are writing AI deployment into contractor frameworks. Project management offices are using large language models to draft the reports that reach steering committees and boards.
AI has already joined the delivery team. The question facing boards and project directors is how governance, accountability and assurance adapt when a growing share of project data, forecasts and recommendations passes through systems that cannot be cross-examined.
What Is AI Actually Doing on Energy, Minerals and Resources Capital Projects?
The clearest evidence comes from scheduling. In April 2026, McKinsey formalised an alliance with ALICE Technologies, reporting that generative scheduling has been deployed with more than 35 clients across capital-intensive industries including energy, mining and infrastructure, achieving schedule accelerations of up to 20 per cent.2 In one data centre case, simplifying schedule logic and re-sequencing work enabled a reduction of roughly 40 per cent against the baseline programme. Both firms are candid that these results came with full operating-model redesign and unusually good data, so the figures describe what is possible under favourable conditions rather than a portable benchmark. For a sector where schedule slippage is the norm, they still mark a step change in what planning technology can do.
Owners are moving too. Saudi Aramco's five-year Project Management Consultancy framework, awarded to eleven contractors including Worley in July 2026, explicitly requires the deployment of AI, digital twins and robotics to improve engineering and project delivery.3 Bechtel has created a senior executive role dedicated to transforming EPC delivery through AI and automation.4 AI capability is becoming a condition of winning delivery work, not a differentiator within it.
At the practitioner level, PMI's research on generative AI adoption finds that heavy users report substantial improvements in scheduling, cost and quality management compared with occasional users.5 These are self-reported perceptions rather than measured delivery outcomes, but the direction is consistent: the tools are being used, and the people using them believe they work.
One distinction matters for EMR readers. Much of the sector's verified AI value to date sits in operations: process optimisation, predictive maintenance, reservoir management. AI in project delivery, where forecasts shape capital commitments rather than throughput, is newer, and its evidence base is thinner. That is precisely where governance needs to be strongest.
Why Is Governance Lagging the Technology?
McKinsey's 2026 AI Trust Maturity Survey, covering around 500 organisations, puts the gap in numbers. Average responsible AI maturity stands at 2.3 on a four-point scale, up from 2.0 a year earlier, but only around a third of organisations reach level three or higher in AI strategy, governance and the governance of agentic AI.6 Capability is being adopted faster than the structures needed to control it.
The same survey identifies the single strongest differentiator: accountability. Organisations with a clearly accountable function for responsible AI average 2.6 maturity, while those without one average 1.8.6 The finding will not surprise anyone who has worked in capital project assurance. Controls without an owner are commentary.
The stakes are also changing shape. As McKinsey notes, the shift to agentic AI means organisations can no longer worry only about systems saying the wrong thing; they must now contend with systems doing the wrong thing.6 An AI that drafts a misleading progress narrative is a reporting problem. An AI agent that re-baselines a schedule or reallocates resources on flawed logic is a delivery problem.
Meanwhile, the people expected to supervise these systems are under-prepared. PMI's global chapter survey found that more than 40 per cent of project professionals had received no AI training at all.7 A delivery team cannot challenge what it does not understand.
Can AI-Generated Project Data Be Trusted Without Verification?
Not on its own, and the evidence for caution is substantial. Gartner forecasts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, driven by escalating costs, unclear business value and inadequate risk controls, and estimates that of the thousands of vendors claiming agentic capability, only around 130 are genuine, a practice it calls "agent washing".8
Error rates deserve equal attention. Stanford research on purpose-built AI legal research tools, systems specifically engineered to ground their answers in verified source material, found they still produced incorrect or improperly supported answers between 17 and 33 per cent of the time.9 Legal research is not project reporting, but the mechanism translates directly: these systems produce fluent, confident output whether or not the underlying content is right. A hallucinated case citation and a hallucinated productivity assumption fail the same way. The difference is that one embarrasses a lawyer and the other misprices a capital commitment.
Capital project governance already has a name for this problem: the gap between reported status and observed reality. AI does not create that gap, but it can industrialise it, generating polished reporting at a volume and speed that outpaces any organisation's capacity to check it. The organisations getting this right have understood as much. McKinsey's State of AI research finds that high-performing adopters are nearly three times more likely to have defined human-in-the-loop verification processes than their peers, at 65 per cent versus 23 per cent.10 Verification discipline, not tooling, separates the leaders. It is the same principle that underpins independent project reviews: information that shapes a capital decision should be verified by someone with no stake in what it says.
Who Is Accountable When the Model Is Wrong?
Accountability does not transfer to software. An AI system holds no professional registration, carries no insurance and cannot be sanctioned. When an AI-generated forecast turns out to be wrong, accountability remains exactly where it always was: with the named project director who relied on it, the engineer who signed the deliverable, and the board that approved the investment. Professional ethics guidance in engineering is already explicit that a practitioner whose work proves deficient cannot point to the AI system as the responsible party.11
The governance architecture for this already exists; it needs extending rather than reinventing. ISO/IEC 42001 provides a certifiable management system standard for AI that integrates with the management systems EMR organisations already run, a subject we examined in AI Is Entering Your Capital Programme - And Governance Will Decide What Happens Next. Regulation is following, though not smoothly: the EU has just deferred the AI Act's high-risk system obligations from August 2026 to December 2027, largely because the standards and tools needed to comply were not ready.12 When even the regulator's timetable slips, waiting for regulation is not a governance strategy.
For capital projects, the practical translation runs through the stage-gate. Any AI output that informs a gate decision, whether a schedule, cost forecast, or risk assessment, should be assured before the gate, with decision rights made explicit: the AI recommends, a named individual decides, and the basis for that decision is traceable.
What Should Boards and Project Directors Do Now?
Board capability is the weak link. Deloitte's global boardroom research finds 66 per cent of board members reporting limited or no knowledge or experience of AI,13 and PwC finds only 18 per cent of directors strongly agreeing their board has functional knowledge of generative AI.14 Boards are discussing AI more than they understand it, which is an uncomfortable position from which to oversee its use in capital allocation.
Four moves close most of the gap:
- Inventory AI across the portfolio. Establish where AI already touches delivery data, forecasts and gate submissions; in most organisations it is more widespread than governance records suggest.
- Assign named accountability. Every AI-assisted output that informs a capital decision needs a specific human owner.
- Mandate independent verification. Treat AI-generated schedules, forecasts and status reporting the way disciplined owners treat contractor data: verified before they reach a gate or a board paper.
- Build AI competence into the PMO and assurance function. Supervising AI-assisted delivery is an assurance capability, and it belongs with the people who challenge project information rather than solely with the IT department.
These are extensions of the questions boards should already be asking about their capital projects, applied to a new source of information.
AI has joined the delivery team, and on the evidence it is a capable one. But capability was never the governance question. The organisations that benefit will be those that treat AI-generated project information the way disciplined owners have always treated contractor-generated project information: useful, necessary, and never accepted without verification.
That verification discipline is where independent assurance earns its place. PDAS provides independent governance and assurance for EMR capital projects, verifying that reported status reflects observed reality, whoever, or whatever, produced the report.
