One project in two hundred: what the capital projects that deliver do differently

Of more than 16,000 big projects in the largest database of its kind, nearly 48 percent finished on budget. Eight and a half percent finished on budget and on time. Half of one percent finished on budget, on time and with the benefits they were built to deliver. One project in two hundred kept all three of its promises.

Those figures come from Bent Flyvbjerg and Dan Gardner's How Big Things Get Done (2023), built on a database Flyvbjerg has assembled with colleagues over decades: more than 16,000 projects, of more than 20 types, in 136 countries, from dams and rail lines to hospitals, IT systems and Olympic Games. They are not a survey of opinion. They are what happened, measured against what was promised when the money was committed.

This post reads that evidence closely, through the eyes of project controls. It asks three questions. Why do big projects miss? What do the few that deliver do differently? And where, in the daily work of schedules, costs, risks and contracts, can project controls change the odds? We gathered the evidence into an eleven-page report, The 0.5%, with every figure cited to its source. It is free at the end of this post.

Read the numbers carefully

A figure as stark as one in two hundred deserves a careful reading before anyone builds an argument on it.

What “on budget” means. In Flyvbjerg's work, cost overrun is measured as actual cost divided by estimated cost, in real terms, against the estimate at the final investment decision. That is the moment the sponsor commits the money. It is not the tender price, a revised budget or the figure that survived three change orders. A project “on budget” finished at or under what it was approved for.

Which way the bias runs. The authors flag their own selection bias, and it runs in the industry's favour: in their words, projects in the wider population “may perform worse than projects in the sample.” Projects that never report their outcomes rarely do so because they went well.

Whether it is new. It is not. In 2002, a study of 258 transport projects worth US$90 billion found that “costs are underestimated in almost 9 out of 10 projects”, that actual costs were on average 28 percent higher than estimated, and that “cost underestimation has not decreased over the past 70 years.” For industrial megaprojects, Edward Merrow of Independent Project Analysis counted a project as failed if its real cost grew 25 percent or more, its schedule slipped 25 percent or more, or it had severe operating problems for two years or more after start-up. By those tests, about two-thirds failed, twice the rate of smaller projects.

Whether the industry is improving. Slowly at best. The McKinsey Global Institute found that construction labour productivity grew about 1 percent a year over two decades, against 2.8 percent for the world economy and 3.6 percent for manufacturing.

Flyvbjerg calls the pattern the iron law of megaprojects: “Over budget, over time, over and over again.” He adds a sentence that is easy to skim past and should not be: “Best practice is an outlier, average practice a disaster.” The rest of this post is about the outliers.

Why big projects miss

The research offers two explanations, and they are less opposed than they sound.

Optimism. People planning a project imagine it going roughly as intended. They picture the sequence, the crews and the deliveries, and they underweight the many small ways reality departs from a plan. Kahneman and Tversky called this the planning fallacy, and it afflicts experienced professionals as much as novices, because it comes from how the plan is built, not from how much the planner knows.

Strategic misrepresentation. The 2002 study went further. Its authors concluded that the underestimation they found “cannot be explained by error and is best explained by strategic misrepresentation, that is, lying.” Projects compete for approval, and a lower estimate wins more often. That conclusion has been challenged: Love and Ahiaga-Dagbui argued in 2018 that framing the question as error or lie is a false choice. For a controls team, the argument matters less than what follows from it. Both explanations have the same cure, because a reference class of real outcomes does not care whether an estimate was optimistic or convenient. It simply shows where similar projects ended up.

There is a third observation in the evidence that controls professionals should find uncomfortable and useful in equal measure. Merrow reviewed 207 failed industrial megaprojects and wrote that “project management was actually culpable in less than 10 of the cases.” The failures were set up earlier: in the front-end work that was skipped, the targets that were imposed, the cost-reduction exercise at sanction. Most of what decides a project's outcome is fixed before delivery starts, which is exactly when most controls teams are least involved.

That is the thread through the rest of this post. The projects that deliver are not run by luckier people. They make different decisions earlier, with better evidence, and they keep making decisions quickly once work begins.

Where overruns live

The averages above blend very different kinds of project. The book's appendix separates them, and the spread is the first lesson.

At the top sit nuclear storage, at a mean overrun of 238 percent, the Olympic Games at 157 percent and nuclear power at 120 percent. Buildings, which most readers of this blog work on in one form or another, average 62 percent. Rail sits at 39 percent, oil and gas at 34, hospitals at 29, roads at 16. At the bottom sit pipelines at 14 percent, wind power at 13, energy transmission at 8 and solar power at 1.

The bottom of the table has something in common, and in a 2026 paper Flyvbjerg, Budzier and Christodoulou name it. The types with low overrun, they write, “are modular”, while those with high overrun “are generally bespoke.” A solar farm is the same panel installed thousands of times. A nuclear plant is close to one of a kind, every time.

For project controls, the practical meaning is blunt. Every estimate starts from a base rate whether anyone writes it down or not. If you are estimating a building, the evidence says the average similar project overran by 62 percent, and 39 percent of them overran by half or more. An estimate that does not know those numbers has not escaped them. It has simply not argued with them.

The average hides the tail

The right-hand column in that chart matters more than the bars. It is the share of projects in the tail, those that overran by 50 percent or more, and the book gives a third number beside it: how far the projects in the tail overran, on average.

Read the IT row slowly. The mean overrun is 73 percent, which sounds bad but survivable. But 18 percent of IT projects overran by 50 percent or more, and those averaged 447 percent. The mean is a blend of a large group that did tolerably and a smaller group that did catastrophically, and it describes neither. The Olympic Games are the opposite shape: 76 percent of Games are in the tail, averaging 200 percent, which means that for the Games the tail is the normal case. Solar has 2 percent of projects in the tail, averaging 50 percent. The tail of the Olympic Games is 38 times fatter than solar's.

Statisticians call these fat tails, and they break a habit most controls teams rely on. Contingency set against the average is no protection against the tail, because the projects that need contingency most need far more than the average. A P50 date, or even a P80, says nothing about the shape of the worst ten percent of outcomes, and the worst ten percent is where careers, claims and companies are lost. We wrote about this at length in A P80 is a date. The tail is the story; the base rates show why it matters at the scale of whole industries.

The controls consequence is a change of question. Not “what is our most likely outcome?” but “what lives in our tail, how often does it happen, and what would remove it?” That second question can only be answered by running the uncertainty, not by reading it off a single date.

Four questions are worth asking of any risk analysis that lands on a director's desk. Which risks appear together in the worst outcomes, rather than one at a time? Which of the inputs were guessed rather than taken from history? Is the spread of dates one hump or two, a world where the permit arrives and a world where it does not? And when was the analysis last run, against which revision of the schedule? An analysis that cannot answer the last question is describing a project that no longer exists.

Habit one: think slow, act fast

Flyvbjerg's most famous example of a project that delivered is old. The owners of the Empire State Building fixed a budget of $50 million and an opening date of 1 May 1931. That gave them “eighteen months to go from first sketch to last rivet.” A company publication of 1931, which the book quotes, put the secret in one sentence: “Even before it was begun, Empire State was finished entirely—on paper.”

Digging began on 17 March 1930. At its peak, in the architect's words, “we erected fourteen and a half floors in ten working days.” The building opened “exactly as scheduled, on May 1, 1931.” It cost $41 million: “That's 17 percent under budget.” Flyvbjerg's name for the pattern is the title of this section. “I call the pattern followed by the Empire State Building and other successful projects ‘Think slow, act fast.’”

The logic is about exposure. Planning is cheap and safe: a mistake on paper costs an eraser. Delivery is expensive and exposed: every month on site is a month in which a supplier can fail, a price can move, a permit can be challenged or a storm can arrive. Flyvbjerg's 2014 overview gives one measure of that exposure from transport projects: on average, a one-year delay or extension of the implementation phase “correlates with an increase in percentage cost overrun of 4.64 percent.” Using a large rail project as an illustration, he put the cost of a one-year delay at 1.2 billion dollars, or 3.3 million dollars a day.

The best plans are also tested, not merely written. The book holds up Pixar, where a film is storyboarded, “photographed and strung together into a video, voices are dubbed in, the video is shown to an audience and the director gets feedback. Then the director goes back, dreams up new ideas, and repeats the whole process, over and over again.” Planning, in this sense, is rehearsal.

Industrial projects tell the same story in different words. Merrow's research found that “what is done (and not done) on the front-end has more bearing on megaproject outcomes than anything else,” and that good front-end loading is “expensive—3 to 5 percent of eventual total cost.” He also found the clearest warning sign in the data: “Nearly 20 percent of the megaprojects in our database attempted a ‘cost-reduction exercise’ just before or just after authorization … Every last project failed.”

For project controls, think slow, act fast is a statement about where the controls effort belongs. Most controls teams are staffed for delivery: they report, update and reconcile once the project is running. The evidence says the cheapest place to change the outcome is earlier, in a schedule that is used to rehearse the project rather than to describe it.

Habit two: take the outside view

Every plan is built from the inside: this design, this team, this sequence, these quantities. The inside view feels precise and is systematically optimistic. The outside view asks a different question: how did projects like this one actually turn out?

The method that makes the outside view practical is reference class forecasting, which builds on Daniel Kahneman and Amos Tversky's work on the planning fallacy. Flyvbjerg describes its first use in practice: in the summer of 2004, the UK Department for Transport and HM Treasury decided to use it in the appraisal of large transport projects. The American Planning Association endorsed it in April 2005. It has since been made mandatory by government in the UK, Denmark and Switzerland, and in 2012 Hong Kong's Development Bureau commissioned a study to test it there.

The UK Treasury publishes the result as a table every appraiser can use. Its supplementary Green Book guidance gives the uplift to apply for optimism bias, by kind of project. The guidance is frank about why: “There is a demonstrated, systematic, tendency for project appraisers to be overly optimistic.”

The Department for Transport's version adds the idea of choosing your risk. For a road project, the required cost uplift is 15 percent if you are willing to accept a 50 percent chance of overrun, and 45 percent if you will only accept 10 percent. For rail, the same two thresholds give 40 and 68 percent. The uplift is not a guess dressed as a number. It is the gap between how projects like yours were estimated and how they turned out, read at the confidence you choose.

For a controls team, the outside view has a practical prerequisite that most organisations lack: a reference class of their own. That means the actual durations and costs of their past work, not the baselines they bid. A library of past baselines is a library of past optimism. A library of as-builts is the organisation's own outside view, and it is worth more than any published table because it was built by your crews, your suppliers and your methods.

Habit three: build with Lego

The bottom of the overrun table is the strongest argument in the book. Solar power overruns 1 percent on average, energy transmission 8, wind 13, pipelines 14. Nuclear power overruns 120 percent and the Olympic Games 157. The difference is not the people or the countries. It is the shape of the work.

A modular project repeats one small unit many times. Every repetition is a lesson: the hundredth turbine base is faster and better than the first because the crew, the sequence and the supply chain have learned. A failure in one unit is small and visible, and it is fixed in the next. A bespoke project does everything once. It learns nothing from itself, because nothing repeats, and its failures are as large as the project.

Few capital projects can become solar farms, but most contain more repetition than their schedules admit: floors, bays, tanks, spools, piles, fit-out zones, turbines, substations. The Lego habit, for project controls, is to find that repetition and measure it. If the controls record knows that the seventh repetition of a unit took less time than the first, and why, the eighth can be planned from evidence rather than from the estimate. If it does not, the project is paying to learn and throwing the lesson away.

The book names other habits too, among them hiring a masterbuilder and getting the team right. Its account of Heathrow's Terminal 5 is a reminder that the people side is not soft. The project used contracts that did not give people and companies incentives to blame each other when something went wrong, and it built psychological safety. In the authors' words, “the fact that the terminal was built by a real team is the single biggest reason why the project beat the odds and finished on budget and on time.”

The decision gap

The three habits have something in common that is easy to miss: each one depends on facts arriving in time. Thinking slow needs real data to test the plan against. The outside view needs the records of past projects. Lego needs every repetition measured while the next one can still benefit. This is where project controls actually sits, and the evidence on how well facts travel is not encouraging.

Put the request-for-information figure next to the Empire State Building. A project that waits a median of nine to ten days for an answer to a question, hundreds of times over, cannot act fast however well it thought slowly. Every one of those waits is a decision made later than it could have been, or made without the fact it needed.

The disputes data points the same way. Arcadis's 2026 report ranks the top causes of US disputes as failing to understand or comply with contract obligations, errors or omissions in contract documents, and delays and time extensions. HKA's database, which covers only projects that ended in a claim or dispute, finds design-related failures on 40.5 percent of them and scope change on 33.4 percent. These are records problems as much as engineering problems: what was agreed, what changed, when it was known, and who was told.

We call the distance between a fact happening and a decision being made with it the decision gap. On a monthly reporting cycle it is measured in weeks. The three habits of the 0.5% are habits of people who close that gap: they test plans while testing is cheap, they decide from evidence about similar work, and they learn from each repetition while the lesson can still be used.

What AI changes, and what it does not

AI arrives in this picture with large promises and, so far, modest use. In RICS's 2025 survey of more than 2,200 construction professionals, about 45 percent reported no AI implementation in their organisations and 34 percent were in early pilots. Just under 12 percent used AI regularly in specific processes, 1.5 percent across multiple processes, and fewer than 1 percent across the whole organisation. Asked where AI could help most, respondents put progress monitoring and project scheduling first, at 36 percent each, then resource optimisation and contract review at 30 percent and risk management at 29. The biggest barriers were a lack of skilled people, at 46 percent, integration with existing systems, at 37, and data quality, at 30.

The studies of AI drafting schedules are a useful corrective to the promises. In a 2023 study in Buildings, six ChatGPT schedules for a single partition wall with a door all left out the door frame, and none seemed to allow for drying time. In 2025, the best fine-tuned model in a study presented at ISARC reached 83.69 percent accuracy on sequencing, which still puts about one activity pair in six in the wrong order. A 2025 benchmark of scheduling constraints found strong models near their ceiling on simple networks, while feasibility collapsed once downtime, time windows and constraints of the “not in the same place at once” kind interacted. The pattern across these studies is consistent: AI misses what is too obvious to write down, and what is physical.

Read against the three habits, the useful role for AI is clear, and it is not the role in the headlines.

• Thinking slow becomes cheaper. When a machine can try many sequences and recovery options against the whole schedule and show the spread of outcomes, rehearsal stops being a luxury of the best-funded projects.
• The outside view becomes reachable. A machine that can read years of a firm's own as-built records and find the comparable work, with sources, turns the reference class from an aspiration into a habit.
• Every repetition gets measured. Progress that arrives from site can be matched to the unit it belongs to, so the eighth floor is planned from the first seven.
• The decision gap narrows. Letters, requests and reports can be read as they arrive and linked to the work they affect, so the fact reaches the person who decides in hours rather than at month end.

What AI does not change is the base rate itself, or who answers for the decision. A model can draft a schedule in a minute, but the overrun table does not care how quickly the plan was typed. The projects that deliver are distinguished by judgement exercised early, on evidence, by people who know what they are signing. AI can put better evidence in front of those people sooner. It cannot sign for them.

What to do on Monday

1. Know your base rate. Before the next estimate is signed, find the mean overrun and the tail for your type of project, from the published tables or your own history, and write down why yours should do better.
2. Report the tail, not just the date. Alongside the P50 or P80, show what lives in the worst ten percent of outcomes and which risks put it there.
3. Rehearse before you commit. Treat the pre-sanction schedule as a model to test. Any late cut to cost or time gets run against the whole schedule before it is accepted.
4. Build a library from as-builts. Take your last three completed projects and record the actual durations of the work you repeat most, with its context: size, method, location, crew.
5. Find the repetition and measure it. List the units your project repeats and track each one, so the next is planned from the last.
6. Measure your decision gap. Pick ten decisions from last month and record how long each fact took to reach the person who decided. That number is the one to shrink.

Where Verian fits

Verian is a project controls service run on Verian's own AI platform. Our team keeps a client's schedule, cost, risk and contract record current every day, from the documents and systems the project already produces, so the facts reach the people who decide while they still matter. Recovery options are tried as scenarios against the whole schedule before anything changes; the risk analysis runs a full critical path calculation in every iteration and reports the spread of dates, not only one of them; first schedules are drafted with a source on every line. The client's people log in, see everything live, and make the decisions.

We wrote The 0.5% because the evidence deserves to be read in one place, by the people who run capital projects. It is eleven pages, built only from published sources, with every figure cited. Leave your work email below and we will send it to you.