Why capital projects go off track (and how to prevent it)

The number that walks into the room

On a Thursday morning, the board of a mid-sized industrial company approves a plant expansion. The binder runs fourteen pages and ends in two figures: 18.4 million dollars and sixteen months. Someone asks whether the number has been "properly looked at". The operations manager replies that he built it with the main contractor, who knows the job better than anyone, and that it even "carries a ten percent contingency, just in case". Signatures follow. There are handshakes, and a photo for the board's WhatsApp group.

The company has just made the most important economic decision of its decade on the basis of a number about which nobody in the room knows the essentials: how much engineering sits behind it, what error range it carries, and what it was benchmarked against.

Two and a half years later, the plant opens. It cost 26.8 million —46% more— and arrived nine months late. At the opening barbecue somebody sums it up with the phrase heard at every latitude: "well, everyone knows these things always end up costing more."

That resignation is understandable. It is also avoidable. And the most interesting part is that the available evidence —tens of thousands of projects measured over decades— says the fate of that investment was not decided during construction, amid surprises and contractor claims. It was decided earlier: in three decisions that were made, or skipped, in that boardroom.

The iron law

Bent Flyvbjerg, Oxford professor and probably the person who has measured more projects than anyone in history, built with his team a database of more than 16,000 capital projects across 136 countries: dams, railways, roads, power plants, buildings, IT systems, Olympic Games. For each one he recorded the same thing: the budget approved at the decision to invest and the real final cost, in constant terms. The results, published in How Big Things Get Done (2023), are uncomfortable.

Flyvbjerg calls it the iron law of projects: over budget, over time, under benefits, over and over again. It is not a peculiarity of governments, or of megaprojects, or of Latin America: the law holds in the private sector and the public one, in projects of 500 thousand dollars and of 5 billion.

The canonical example is the Sydney Opera House: budgeted at 7 million Australian dollars, it ended up costing 102 million —an overrun of around 1,400%— and was delivered ten years late. Its architect, Jørn Utzon, resigned midway, left Australia and died without ever seeing finished the masterpiece that posthumously earned him the Pritzker Prize. At the other extreme stands the Empire State Building: built in under two years, opened ahead of schedule, and costing 41 million dollars against a 50 million budget. The difference between the two was not luck. It was method — above all, how much was known about the project before the spending began.

The average is the kind part

There is something worse than the average overrun: the shape of the distribution. Project deviations do not follow a normal bell curve; they have fat tails. Flyvbjerg and Alexander Budzier showed it for IT projects in Harvard Business Review (2011): the average overrun was 27%, an almost reassuring figure. But one in six projects was a "black swan", with an average overrun of 200%. The average describes the weather; the tail describes the storm that sinks a company.

The right question is not "how much does the typical project deviate?" but "what is my probability of landing in the tail — and do I survive it?".

Three causes, one antidote

Why does this happen everywhere, for decades, with competent professionals? The research converges on three mutually reinforcing causes.

The planning fallacy. Daniel Kahneman and Amos Tversky described it in 1979: we estimate for the scenario in which everything goes well, even while knowing our own history of delays. It is a cognitive bias, not a moral defect: it ships with the human brain.

Strategic misrepresentation. Flyvbjerg and Martin Wachs documented something less innocent: early numbers are not built to be right; they are built to get the project approved. The optimistic budget wins the funds; the realistic one loses the internal contest. The result is what Flyvbjerg calls "survival of the unfittest": the projects with the most wrong numbers are the ones that get approved.

Insufficient definition. The most operational cause, and the most correctable: commitments of surgical precision are approved on engineering that barely exists. The number looks exact —18.4 million— but behind it there is a preliminary layout, two quotations and a spreadsheet.

All three causes share the same antidote, and it is not on the construction site: it is in how the decision is made. Edward Merrow, founder of IPA (Independent Project Analysis), which has audited industrial projects for forty years, measured that 65% of industrial megaprojects fail —defining failure as exceeding cost or schedule by more than 25%, or starting up with serious production problems—. And his central finding is liberating: the best predictor of the outcome is not the country, the contractor or the technology. It is the quality of the definition work done before funds are sanctioned, what the industry calls front-end loading (FEL). Put differently: by the time construction starts, the die is largely cast.

Hence the three decisions that follow.

Decision 1 — How much to know before approving

The first decision is explicit in mature organizations and invisible everywhere else: at what level of definition are funds authorized?

The industries that live off projects —oil, mining, process chemicals— solved this decades ago with the phase-and-gate cycle (stage-gate): the investment is not approved once; it is approved in stages, and each gate demands a higher level of definition to release the next tranche of funds.

AACE International (the association of cost engineering since 1956) put numbers to that logic in its Recommended Practice 18R-97, which classifies estimates into five classes according to the percentage of engineering completed. The expected accuracy ranges speak for themselves:

Representative ranges per AACE RP 18R-97; they vary with industry and complexity. The logic is universal.

A Class 5 estimate with a possible +100% error is not a bad estimate: it is an honest estimate for the information available. The classic mistake —the one in the opening scene— is to treat a Class 5 number as if it were a Class 2 commitment, approve it, communicate it to the bank and the board, and later call "overrun" what was really budgeted ignorance. It is nobody's dishonesty: it is the physics of information.

How much does that definition cost to buy? Merrow quantified it: a well-executed front end costs between 3% and 5% of the project's total capital, and it is the cheapest insurance premium in the world of investment: his data shows that projects with poor FEL pile up overruns worth several times that initial saving. The Construction Industry Institute reaches the same conclusion with its own project database: structured early planning systematically improves cost and schedule performance. The practical rule for the board is simple: before authorizing the bulk of the funds, demand a Class 3 estimate or better — and gladly fund the weeks of engineering needed to get there.

Decision 2 — How the number is set

The second decision: is the project approved with a point or with a distribution?

A single number —18.4 million— is not a forecast; it is an unquantified bet. Every serious estimate is a distribution of possible outcomes, and the management question is which point of that distribution is adopted as the budget. The P50, which will be exceeded half the time? The P80, which leaves a 20% probability of exceedance? There is no single answer —it depends on risk appetite and on how lethal the deviation would be—, but there is a rule: if nobody in the room can say which percentile the number is, the number is not ready to be approved.

Two concrete practices follow from that logic.

Calculated contingency, not customary contingency. The "10% just in case" is folklore, not risk management. The right contingency comes from identifying the project's risks, estimating their impacts and running the probabilistic analysis —typically a Monte Carlo simulation over the estimate, as AACE practices recommend—. In early phases, the contingency that results is usually much larger than the ritual 10%; that discomfort is information, not pessimism.

The outside view. Kahneman proposed the conceptual corrective and Flyvbjerg turned it into method with reference class forecasting: instead of building the forecast only from the inside ("our project, our quotations, our schedule"), it is anchored in the real distribution of comparable finished projects. How did the last twenty projects like yours, run by companies like yours, actually end? That data exists and is the best available starting point. The technique stopped being academic long ago: the UK Treasury has required explicit optimism-bias adjustments in public projects since 2003 (the Green Book's optimism bias uplifts), and Hong Kong adopted it for its infrastructure portfolio.

The domestic version of the method is within any company's reach: before approving, put on the table your own last five projects —approved budget against final cost, promised date against real date— and ask why this one would be different. It is an uncomfortable twenty-minute meeting worth millions.

Decision 3 — Who governs the number

The third decision is the least technical and the most decisive: who owns the outcome, and with what system do they govern it? Three components separate governed projects from accompanied ones.

A single owner, with real authority. When the project's outcome is shared among the plant manager, the contractor, the engineering firm and "the committee", the outcome belongs to no one. The right figure is a project director with a mandate from the board, contingency budget under their signature and the obligation to report against the baseline. In large investments, that role is a full-time job, not an extra hat for the manager who already runs an operation.

A frozen baseline and change control. The approved budget is frozen. Every later change —of scope, specification or sequence— goes through a formal mechanism that exposes its impact on cost, schedule and benefit before being approved, signed by whoever is accountable. Without this, the budget is not exceeded all at once: it erodes through small decisions nobody added up. The "surprising" overrun is usually the arithmetic sum of changes that were in fact approved, one by one, with no one watching the total.

Early warning through earned value. A project's cost performance is measured by comparing the value of work performed against what has been spent (the CPI index of the earned value method). And here the evidence is blunt: David Christensen's studies across hundreds of US defense programs showed that the cumulative CPI stabilizes early —by 20% completion it already anticipates the final result within a margin of about 10%— and that from there it tends to get worse, not better. The management translation: projects do not go wrong at the end; they reveal themselves at the beginning, if anyone is looking. "We will recover it in the next phase" is, statistically, wishful thinking.

What unites the three components is one idea: visible governance. An open dashboard, documented decisions, deviations declared at birth. In our experience, the difference between organizations that learn from their projects and those that repeat them is not the management software: it is whether the bad number can be said out loud in time.

The alternative version of the scene

Return to the boardroom at the start. The alternative version is not heroic; it is a little slower and considerably more boring. The board does not approve 18.4 million: it approves 280,000 dollars —1.5% of the capital— for eight weeks of engineering and definition. The number comes back as a range with a first and last name: a P50 of 21.3 million, a P80 of 23.1, with the five main risks quantified and a project director proposed. The board approves the P80, freezes the baseline and schedules reviews at every gate.

The plant opens fourteen months after the main purchase order, at a final cost of 22.4 million: 3% below the approved budget, 22% above the original illusion. The difference between the two versions was not made on the construction site. It was made by three decisions taken when almost nothing had yet been spent — which is exactly the moment when a board holds maximum power over its investment.

Deciding the investment before spending: that, in one sentence, is the craft.