Cost Forecasting & Overruns

Cost overruns are one of construction's defining failures. Projects, and especially large and complex ones, routinely finish over budget, often substantially, and the pattern is remarkably consistent across countries and decades. The causes repeat: optimistic initial estimates, scope changes, unforeseen ground and site conditions, design development, delays that cost money, poor productivity, price inflation, and the simple accumulation of many small variances that no one added up in time. Understanding that overruns are systemic, not exceptional, is the starting point - it means the question is not whether cost will move from the budget but when you will notice, and whether you notice early enough to do something.

That timing is everything, and it is why early warning is so valuable. The cost of a problem to fix rises steeply the later it is found: early in a job there are many levers - resequencing, value engineering, renegotiation, scope adjustment, pace changes - and late in a job there are almost none, because the money is largely committed or spent. A forecast that says, three months in, 'on current trends this finishes eight per cent over' gives management a real chance to respond; the same truth discovered at handover gives them only a reckoning. So the practical goal of cost forecasting is not perfect prediction of the final number - which is impossible - but useful, early, directional warning: is the cost drifting, in which direction, driven by what, with enough lead time to act.

Traditionally this is done with structured cost control - comparing committed and actual costs against the budget and the value of work done, and projecting a forecast final cost, often through earned value techniques. This works, but it is labour-intensive, backward-looking and only as timely as the reporting cycle, so drift can hide between reports and be recognised late. This is precisely the gap AI addresses: it can watch the cost and progress data more continuously, learn from the patterns of past projects what early drift tends to precede, and surface a likely overrun sooner and with less manual effort. The promise is earlier, cheaper attention - as always, dependent on the data and always an input to a human decision.

AI cost forecasting works by learning the relationship between a project's early, observable signals and its eventual cost outcome, then applying that to a live job. Fed the cost and progress data of many past projects - how spending tracked against progress, which early variances preceded overruns, how particular trades and packages behaved - a model learns the patterns that tend to end badly. Pointed at a running project, it reads the same signals as they emerge and projects where the cost is heading, updating as new data arrives. In effect it turns the raw stream of cost and progress data into a continuously refreshed forecast and an early flag when the trend turns adverse.

Several kinds of signal feed this. Cost-versus-progress relationships: is spend running ahead of the work actually completed, the classic early sign of trouble. Variance patterns: are small overruns clustering in particular packages or accelerating rather than settling. Leading indicators from elsewhere on the project: schedule slippage (which usually costs money), rising change-order volume, productivity dropping, procurement coming in above allowance. Comparison to precedent: does this project's early cost behaviour resemble past projects that overran. A model can weigh many such signals together, across more data and more continuously than a manager reviewing monthly reports, and raise a flag while the deviation is still small enough to act on.

The value is real and specific: earlier warning, less manual effort, and attention directed to the packages most likely to drift. Instead of discovering at month nine that the job is over, management gets a flag at month three that this trend, on this evidence, points to an overrun in this area - a prompt to investigate. That is the right framing: the AI does not control the cost or decide the response; it forecasts and flags, and a human interrogates the flag, judges whether it is real, finds the cause, and decides what to do. Used this way, AI cost forecasting is an intelligence layer over cost control - a way to see the overrun coming while there is still time to change the outcome, provided the data it reads is honest and current, which the next section makes the condition it is.

AI cost forecasting has a dependency even sharper than estimating's, because it reads a live project, and a live project's cost data is often late, incomplete and not entirely honest. A forecast is only as good as the current data feeding it, and construction cost reporting is notoriously lagged and messy: costs are committed but not yet recorded, invoices arrive late, progress is self-reported and sometimes optimistic, variations and claims sit unresolved, and coding is inconsistent across packages. Feed a forecasting model this and it will still produce a confident, precise-looking projection - one that reflects the gaps and biases in the data rather than the real state of the job. The forecast can be wrong in either direction, and a falsely reassuring one is the most dangerous.

Two data problems deserve particular attention. The first is timeliness: a forecast built on data that is weeks out of date is forecasting the past, and drift that started after the last update is invisible to it - so the model can look calm while the job is already turning. The second is honesty of progress: cost-versus-progress signals depend on progress being reported truthfully, and progress on construction sites is frequently overstated (the classic ninety-per-cent-done that stays ninety per cent for months). If progress is inflated, the model thinks the job is doing better than it is and under-forecasts the overrun - exactly when a warning is most needed. Garbage in, garbage out here does not just weaken the forecast; it can invert it.

The consequence is that the forecast is only ever as trustworthy as the cost-and-progress data discipline underneath it - which is, again, unglamorous data-foundation work: capturing costs promptly and consistently, measuring progress honestly, resolving and recording variations, coding uniformly. Where that discipline exists, AI can genuinely see overruns coming earlier than manual control. Where it does not, the forecast is a confident fiction, and reading it as fact is worse than having no forecast at all. So the forecast must always be read as a data-conditioned signal to investigate, not a fact to bank - and every forecast should come with an honest sense of the quality and currency of the data behind it.

The deepest risk of AI cost forecasting is not that it is sometimes wrong - all forecasts are - but that its precise, continuously updated, authoritative-looking output invites false confidence. A number like 'forecast final cost: 4.2 per cent over, confidence high' feels like knowledge, and a busy management team under pressure is strongly tempted to treat it as settled and stop asking hard questions. This is automation bias applied to money: over-trusting a confident machine forecast, standing down the human scrutiny that would have caught what the model missed. A forecast that is precise and wrong, believed because it looks authoritative, can do more damage than an honest 'we are not sure, but spend is running ahead of progress and we should look' - because it replaces vigilance with a false sense of control.

The discipline that prevents this is to treat every forecast as a prompt to investigate, not an answer to accept, and to keep asking three questions: what data is this built on and how current and honest is it; what is actually driving the flagged trend; and what is the uncertainty around this number, not just the point value. A forecast with a wide, honest range and a named cause is far more useful than a narrow, confident, unexplained one. Good practice pairs the AI forecast with human cost control, not in place of it - the model widens the net and gives earlier warning, the people verify, diagnose and decide.

And the accountability line holds exactly as in estimating. The forecast informs decisions - about scope, pace, procurement, contingency draw-down, and what to tell the client and the board - but those decisions, and the commitments that follow, are owned by people: the project manager and cost manager who run the control process, the quantity surveyor who stands behind the numbers, and the contract that governs who bears which cost. An AI can forecast where the cost is heading; it cannot be responsible for the budget, the commercial decisions, or the contractual consequences of an overrun. Defer every binding cost and contractual commitment to the qualified professionals and the governing contract and law, use the forecast as early, data-conditioned warning, and never let a confident number replace human vigilance over the money.