AI-powered forecasting tools are arriving in infrastructure risk management and they represent a genuine step forward. But the sector’s persistent delivery problem — cost overruns, schedule slippage, benefits never realised — has survived thirty years of increasingly sophisticated forecasting methods. This article argues that the bottleneck was never forecast quality. It was always what organisations do with a forecast once they have one. That is a different problem, and it requires a different kind of investment to solve.

The promise of AI forecasting

There is something genuinely promising happening in construction risk management, and it deserves to be acknowledged clearly before the more uncomfortable argument begins.

AI-driven schedule risk analysis — the kind now being deployed at scale on major infrastructure programmes — represents a credible advance on what preceded it. The ability to generate probabilistic, activity-level forecasts from a dataset of historical schedules, without elicitation workshops, without rolled-up logic, and without the particular theatre of a room full of experienced professionals being asked to put a number on something none of them have actually measured — this ought to be a meaningful improvement on what it replaces. The theoretical case is sound, and the instinct to wave it away as overhyped deserves at least a moment of scrutiny before it’s indulged.

So the question this article asks is not whether the approach is promising. It is. The question is whether improving the forecast is actually the problem the sector most needs to solve.

The Iron Law doesn’t budge

Quantitative Schedule Risk Analysis became the industry standard for analysing delivery risk on major infrastructure programmes some thirty years ago. Over that period it was refined, professionalised, and progressively embedded into governance frameworks across energy, transport, utilities, and beyond. Forecast quality, by the measures the profession uses to assess itself, was assumed to be improving.

And yet. Bent Flyvbjerg, whose database of more than 16,000 projects across 136 countries represents the most comprehensive empirical record of infrastructure delivery ever assembled, has a name for what that record shows: the Iron Law of Megaprojects. Over budget. Over time. Under benefits. Over and over again. The pattern has not improved in ninety years of comparable data. It has not responded to better tools, better methods, or greater institutional sophistication.

If better forecasting was the lever, the sector pulled it for three decades — and the Iron Law didn’t budge. Which should prompt a question the industry has been slow to ask: is the forecast actually the binding constraint?

16,000+ infrastructure projects in Flyvbjerg’s database

136 countries across the dataset

90 years of comparable data outcomes unchanged throughout

The forecast-consumption problem

The risk profession’s answer to persistent delivery failure has, with impressive consistency, been to improve the forecast. Better elicitation techniques. More sophisticated Monte Carlo models. Improved schedule quality standards. And now, AI-derived probabilistic distributions in place of expert estimates. Each iteration has addressed a real limitation of its predecessor. Each has generated genuine enthusiasm. And each has left the delivery statistics largely unmoved. The jury is still out on whether AI-driven forecasting will fare better — the programmes currently using it have not yet run their course, and the sector has a long history of enthusiasm preceding evidence.

What has received almost no equivalent investment — in tools, in methodology, in professional development, or in governance design — is what happens to a forecast after it is produced. The system that receives a risk output and converts it into a decision. The mechanism that translates a P80 date into a concrete set of options with owners and timelines. The accountability structure that makes acting on a credible forecast the path of least resistance rather than the most politically costly move available.

This is the forecast-consumption problem. And it is, in many infrastructure organisations, almost entirely unengineered.

The investment flowing into AI-powered forecasting is real, and the tools it is producing are genuinely better. But better forecasts fed into broken consumption systems will produce better-documented versions of the same outcomes. Before the sector congratulates itself on the quality of its new inputs, it might usefully ask what it has done to improve the quality of what happens next.

Before you read on

Does this sound familiar?

Tick every scenario you have personally witnessed on a programme.

The profile below is indicative, not diagnostic — treat it as a prompt for reflection rather than a precise result. The failure modes will be described in detail later, this is to warm you up!

0 recognised

The risk was on the register, correctly quantified, traffic-lighted amber — and the meeting ended with “we need to keep a close eye on this.” Nobody disputed the forecast. Nobody acted on it either. The risk analyst flagged a six-month exposure. The people who could actually do something about it weren’t in the room. By the time the forecast reached someone with genuine authority, it had become a line in a dashboard. Everyone in the room understood that acting on the risk would make the programme look problematic. So nobody acted. The post-project review noted — with some puzzlement — that the warning signs had been there all along. The tool was better than its predecessor. The delivery outcome was not.

Tick any scenarios you recognise from programmes you’ve worked on.

Your programme’s failure profile

Failure 1 — Decision rightsNo named owner for the decision a forecast implies

Failure 2 — Translation gapForecast never converted into a concrete decision frame

Failure 3 — Accountability asymmetryDeferral was the rational organisational move

When the forecast doesn’t fail

Most senior leaders on major programmes carry a version of the same story. The details differ. The shape is consistent.

The risk is visible. It has been correctly characterised and probabilistically quantified. The report is clear. The meeting happens. The traffic light is amber, tending red. Heads nod. Someone says the words “we need to keep a close eye on this.” And then — nothing changes. Not because anyone disputed the analysis. Not because the forecast was wrong. The risk matures into a problem at roughly the schedule it was always going to mature at, and eighteen months later the post-project review notes, with some puzzlement, that the warning signs were there.

The forecast didn’t fail. Something else did.

What failed — in almost every instance of this shape — was the architecture between the forecast and the decision. Three specific failures, each compounding the last.

The closing window

Watch what happens when action is deferred

Drag the scrubber through 16 weeks of programme time. Each missed decision point narrows the options still available — and compounds the cost of eventual recovery.

Watch the intervention options change as the window closes.

Decision window

Open

Wk 0 Wk 4 Wk 8 Wk 12 Wk 16

Risk identified. P80 shows six-month exposure on a critical contractor interface. All intervention options are open.

Recovery cost 1.0×

Intervention options

The window has closed.

Every intervention option has expired. Recovery is the only path — at approximately 4.8× the cost of acting at week zero. The forecast was accurate throughout. The post-project review will note, with some puzzlement, that the warning signs were there all along.

Three structural failures

1 · Decision rights are not mapped to forecast outputs

Most forecasts are produced at a level where the decision they imply cannot actually be made. The risk team surfaces a six-month exposure on a critical interface. The decisions that exposure calls for — accelerating procurement, renegotiating a contract, restructuring a sequence, having a difficult conversation with the client — sit three levels above the room where the forecast was received. Nobody present has the authority to act. The forecast is summarised, escalated, translated into a format that survives the journey upwards, and by the time it reaches someone with genuine decision authority it has become a line in a dashboard.

The question of who is accountable for the decision a given class of forecast implies — not who receives the report, but who owns the decision the report calls for — is, in many programmes, either unanswered or answered so vaguely it provides no real accountability at all. Fixing this is a governance design exercise. It requires someone to ask the question deliberately and pursue the answer through the inevitable organisational resistance.

2 · The translation gap is not a designed function

A P80 date is not a decision. It is an input to a decision. The work of converting a forecast into a concrete decision frame — here are the realistic options, here are their costs and implications, here is the window within which each option remains available — is neither trivial nor automatic. It requires contextual knowledge of what is actually possible on the programme at that moment, combined with enough analytical rigour to frame the options with the seriousness the forecast deserves.

In many delivery organisations, this work belongs to no one as a defined responsibility. The risk analyst’s job ends at the output. The project director’s job begins at the decision. The translation between them — the step that actually converts information into choice — is informal, inconsistent, and frequently absent. Organisations that want to close this gap need to build this capability explicitly: people who sit at the boundary between forecast production and decision-making, whose job is the translation, not the analysis and not the governance.

3 · The accountability asymmetry is structurally embedded

This is the most uncomfortable failure, because it is structural rather than accidental. In most delivery environments, the short-term cost of acknowledging a risk and acting on it is higher than the short-term cost of not acting. Acting means accepting schedule exposure publicly, potentially triggering contract conversations, making the programme look problematic in a review cycle, or inviting scrutiny the team is not yet ready for. Not acting defers all of that. In an environment where people are assessed on short-term reporting metrics and political positioning — and most infrastructure delivery environments are exactly that — deferral is frequently the rational organisational move.

Until there is explicit accountability for receiving a credible forecast and doing nothing with it, delivery organisations will systematically underuse even excellent forecasts. In practice this means building a named decision log into programme governance — a mechanism that records not just what risks were identified, but what decisions were taken or deferred in response, by whom, and when. Making that log visible at programme board level, and making inaction as accountable as inaccuracy, is the structural change that no forecasting tool, however sophisticated, can substitute for.

Test your diagnosis

Name That Failure

Four short scenarios. Each is a real failure pattern — but which one? Assign each to the structural failure that produced it.

The harder work

None of this is an argument against better forecasting. It is an argument about sequencing — about where the next investment should go, and what problem it should be solving.

The sector is about to spend significant money and political capital embedding AI-powered risk tools into major programmes. The forecasts those tools produce may well be genuinely better — more granular, faster, less contaminated by the cognitive limitations of expert elicitation. This is worth having. The question that isn’t being asked, loudly enough or in the right rooms, is: what are we investing in to improve the quality of the decisions those forecasts will feed?

The forecast-consumption problem has three components — decision rights, translation capability, and accountability for inaction — and none of them are technology problems. They are governance design problems. They require deliberate structural choices that sponsors, clients, and delivery organisations are currently not being asked to make. Better tools will not prompt those choices. Asking better questions might.

The delivery statistics of the last thirty years are not primarily a forecast quality problem. They are a decision quality problem. And the infrastructure sector, for all its sophistication, has not yet seriously built the systems, structures, and governance that turn risk intelligence into risk-informed action.

That is the problem worth solving. The tools to improve forecasting are arriving. The question is whether the industry is prepared to do the harder work that makes those forecasts matter.

Where to start

None of the three failures described above requires a transformation programme to address. Each has a tractable entry point — a question that can be asked this week, in the next programme review, or at the next board.

Decision rights

Take the three most material risks currently on your programme risk register and ask, for each one, who is accountable for the decision that risk implies. Not who receives the report. Not who chairs the risk review. Who owns the decision the risk calls for, and what would it take for them to act?

If that question produces silence or ambiguity, you have found the first thing to fix.

Translation

Look at the last risk report produced on your programme and ask whether it contains a decision frame — a set of options, with costs, implications, and a window of availability.

If it contains a forecast and a traffic light but no decision frame, the translation gap is open. The next report should close it.

Accountability

Check whether your programme governance contains any mechanism that records what was decided — or explicitly deferred — in response to a material risk forecast, by whom, and when.

If that record does not exist, the asymmetry between action and inaction is structurally embedded in your governance. A decision log costs almost nothing to introduce and changes the accountability dynamic immediately.

These are not complicated interventions. They do not require new tools, new software, or a change programme. They require someone with the authority and the appetite to ask uncomfortable questions about how the organisation currently converts risk intelligence into action — and to hold that question open until it has a real answer.

The forecast-consumption problem is solvable. It is just not being solved, because until now the sector has been looking somewhere else.