AI Is Making Analysis Cheap. Decisions Are Still Expensive.

One important part of that shift is now measurable. Adoption has followed. Construction is moving with it.

The gains are real, but they are uneven.

Output is becoming cheap. Correctness is not uniform, and the people using AI are not always able to tell the difference. That is the executive problem this article addresses.

The economist Daron Acemoglu makes a related point. Early productivity evidence comes largely from easy to learn tasks, while some of the future effects will come from hard to learn tasks:

That is a fair description of most consequential capital project decisions.

So the constraint moves upstream. The harder questions become:

What are we actually deciding? Which evidence deserves weight? What are we assuming? What could change the answer? What consequence are we prepared to accept? And who has the authority to commit?

In our experience, consequential project decisions are rarely improved simply by adding another report. Public evidence points the same way. In an ongoing review of how long decisions take in Northern Ireland’s public sector approval process, officials involved raised one observation often enough for the Strategic Investment Board to record it: “it appears that the more data decision-makers collect, the less comprehensible the problem becomes,” and the report notes no evidence that such additional analysis has led to better decisions.8 Bent Flyvbjerg’s research shows that persistent cost overruns have remained a major project problem across decades.9 We read that as an important warning: better data and methods alone do not guarantee better decisions.

AI is already helping project teams summarize and review content, identify patterns, generate scenarios, and coordinate multi step work. In a 2024 global survey by the Project Management Institute of 500 project professionals who already use generative AI, across 18 sectors and 12 countries, 53% reported using it to summarize and review content, and 62% of those said it had improved their productivity.10 Deloitte’s 2026 outlook reports that many engineering and construction firms are piloting agentic AI systems “to autonomously manage complex scheduling, coordinate workflows, and mitigate risk,” without quantifying how many.11 A 2022 systematic review of AI in construction, covering literature published before the generative AI wave, identified reducing the time spent on repetitive tasks as the major opportunity.12

Those capabilities are useful. They are not the same as decision quality.

• A faster analysis can still use the wrong baseline.

• A strong summary can still rely on superseded evidence.

• A sophisticated model can still hide a weak assumption.

• A precise forecast can still answer the wrong management question.

• A confident recommendation can still be wrong, or exceed the authority of the system that produced it. The generative AI risk profile published by the National Institute of Standards and Technology names this directly: “confabulation,” the production of “confidently stated but erroneous or false content.”13

In a capital project, a weak decision does not remain on a screen. It becomes a procurement commitment, a contract position, a schedule consequence, a retained risk, or a change event. The scale of that exposure is visible in dispute data.

Capital projects already generate large volumes of information across schedules, estimates, commitments, procurement, design, contracts, RFIs, submittals, risk, BIM, field reporting, and change management.

The industry does not have a data shortage. It has a decision shortage.

The problem is whether that information carries a consistent meaning. One system identifies work through a cost code, another through an activity ID, another through a location, procurement package, or contract line item. Those systems can be technically connected and still describe different versions of the same project.

Three recent studies point the same way:

Two of those three studies are sponsored by technology vendors, which is a reason to read the direction of travel rather than the precision of any single figure.

If scope, quantity, location, responsibility, schedule, cost, procurement, risk, and change are not linked through a governed project structure, AI may process fragmentation faster. It does not make the fragmentation disappear. AI does not automatically improve a weak management process. It can make that process faster, more scalable, and harder to question.

The same principle applies to evidence. A drawing may be current or superseded. A historical case may be comparable or misleading. A benchmark may provide context without establishing a contractual requirement. Information management standards already treat this as a governance issue. ISO 19650-1 uses revision and status codes to identify the permitted use of information containers. The 2018 edition remains the current published edition while a second edition is under development.17

For executives, the practical question is:

What should we trust, for this decision, at this time?

It matters whenever the organization must award or hold, proceed or pause, negotiate or clarify, revise a forecast, change contingency, repackage procurement, escalate an exception, or commit capital.

We would not begin a consequential decision with What does the AI recommend? We would begin with seven questions.

1. 01What are we deciding?
2. 02Which evidence?
3. 03Which context?
4. 04Which assumptions?
5. 05Which scenarios?
6. 06Which consequences?
7. 07Whose authority?

The sequence moves from information to meaning, from meaning to choice, and from choice to accountability. The executive constraint is rarely the absence of another twenty page report. More often it is one unresolved assumption, one commercial exception, one missing approval, one untested scenario, or one threshold that could reverse the recommendation.

AI does not replace them. It can help answer them earlier, more continuously, and with a broader body of evidence.

Q1 to Q5

Q1 to Q5 are the five xPM management questions — what should be the case, what happened, what is happening, what could happen, and what should we do — that organize decision support so that evidence answers a defined management question rather than the other way around.

The current project record, organizational history, standards, professional guidance, and external information can all contribute to a decision. They should not be blended into one undifferentiated pool.

D1 to D5 are five evidence jobs, not a maturity ladder.

D1 to D5

D1 to D5 are the five distinct evidence jobs in the xPM framework — frame the decision, prove it from the live project record, remember comparable organizational history, test it against standards and disciplined methods, and challenge it with outside evidence — so that the availability of information is never mistaken for its authority over the decision.

The xPM Strategic Decision architecture is a decision method. It is a structured way of working that makes the path to a commitment explicit. It is not an autonomous system, it is not a software product, and it does not produce “the answer.”

Four things travel that path, and they are not interchangeable. A fact is what is observed. A forecast is what is expected. A recommendation is what should be considered. A commitment is what the organization decides to do and will be held accountable for. Each can support the next. None of them is the commitment.

Q1 to Q5 organize the management question. D1 to D5 organize the evidence. The next responsibility is to expose what could change the conclusion: which assumptions are material, which thresholds matter, what evidence is missing, which scenario would reverse the preferred position, and when the decision should be revisited.

A small number of well documented techniques covers most of the need:

Once structure, evidence architecture, and decision logic are established, AI earns its place. That is why we put it late in the architecture rather than at the beginning.

Generative AI can summarize controlled information, classify documents and events, compare alternatives, draft management narratives, and surface contradictions for review.

Agentic workflows can coordinate multi step work within defined permissions. As an xPM design pattern, we expect an agentic workflow to interpret intent, plan, retrieve governed information, use approved tools, check results, escalate exceptions, and confirm completion. That pattern is ours, and no vendor endorses it. Published engineering guidance points in a compatible direction. Anthropic advises that agents gain “ground truth” from the environment at each step, that they can pause for human feedback at checkpoints, and that they run with stopping conditions and guardrails, with simpler workflows preferred wherever those suffice.21 For governance, the anchor should be a governance class source such as the NIST AI Risk Management Framework,23 not a vendor engineering post. For consequential work, verification belongs inside the workflow rather than after the answer.

Predictive AI can support early warning, pattern recognition, forecast ranges, and sensitivity to key drivers. The objective is to make uncertainty more visible, not to convert it into false precision.

Smaller or locally deployed models may be appropriate where privacy, latency, cost, control, or domain specific behavior matter more than raw scale.

Human oversight is necessary but not automatically sufficient. NIST identifies automation bias and over reliance as risks of the human and AI configuration itself,13 and the Center for Security and Emerging Technology at Georgetown University cautions that “human-in-the-loop cannot prevent all accidents or errors.”22 Oversight works when the reviewer has the time, context, and authority to disagree.

The boundary can be drawn as a matrix. This is an xPM editorial position rather than an industry standard: as consequence rises and reversibility falls, the autonomy granted to a system should fall with it, until the final commitment sits with a named person.

As analytical capability grows, the bar for a decision ready case should rise with it. This is an xPM editorial position rather than an industry standard, and it is meant to be applied in proportion to consequence. The full bar belongs on decisions that commit capital, change a contractual position, or alter risk allocation. Routine decisions need only the first four criteria. For a consequential capital project decision, we would expect the case to be:

The xPM Decision Bar

The xPM Decision Bar is a nine-criterion test — framed, evidence-backed, context-aware, assumption-visible, scenario-tested, consequence-conscious, challengeable, explainable, and accountable — that a consequential capital-project decision case should clear before commitment, applied in proportion to the consequence of the decision.

The practical test is simple. At your next gate review, ask which of these nine your current decision case would fail today, and fix that one first.

The last criterion needs one clarification. Boards, investment committees, and delegated authority matrices often decide collectively. That does not remove accountability. AACE International’s terminology states that accountability “cannot be delegated but it can be shared.”27 In the xPM approach, shared or collective authority should still have a named decision owner who carries the decision, its conditions, and its follow through. Depending on the governance structure, that may be the sponsor, chair, or delegated executive. A related profession specific example appears in the global RICS standard on responsible AI use, in effect since March 2026, under which the surveyor “remains accountable for every piece of professional advice, regardless of the tools used to produce it.”28

Our experience has led us to a simple sequence. We offer it as professional judgment rather than as a validated finding, and we say below what evidence would test it.

Structure First

Structure First is the opening step of the xPM sequence, which requires the management architecture — scope structure, coding, baselines, evidence roles, decision questions, governance, thresholds, and decision rights — to be defined before analytics or AI are asked to operate inside it.

The closest research support we can point to concerns the front end. A National Academies assessment of federal project management records that data from the Construction Industry Institute show a positive correlation between front end planning and project performance in cost, schedule, change orders, and operational performance.29 That correlation draws on a Construction Industry Institute research population of more than 600 capital projects representing approximately $37 billion in installed cost.30 That supports structuring decisions early. It does not yet test our full sequence. An honest test would compare decision outcomes across portfolios that govern structure and evidence before deploying AI against portfolios that do the reverse, holding consequence and contract type constant. We would publish that result either way.

The cost of producing an answer is falling. The cost of acting on the wrong one is not.