A new hospital wing is announced: finished in three years, for a stated price. Four years later it opens, and the bill has grown by a third. Nobody was dishonest in the headline sense, and the engineering was sound. The delay was not a mystery either: a delivery came late, the wrong crew was waiting for a space that was not ready, and the soil under one corner was not what the survey said.
Building a structure is the easy half of civil engineering. Getting it built, on time and on budget, by hundreds of people who do not work for one organisation, is the other half, and statistically the half that goes wrong more often. A graduate of construction management does not see a building rise. They see a network of tasks, a crew chart, a price with a risk range, a contract that decides who pays when the soil surprises, and a site where people could be hurt.
Every project balances three things: scope (what you build), cost and time. Change one and another moves. The picture of a triangle with these corners, and quality in the middle, goes back at least to Martin Barnes, who taught a version of it in 1969. The slogan "good, fast, cheap: pick two" is its folk version. Real projects are subtler than the slogan, because you can trade, but not freely: adding people to a late project does not make it proportionally quicker.
The first tool for taming a project is the work breakdown structure (WBS): cut the whole job into pieces until every piece has one owner, one budget and a way of knowing when it is done. A house extension might split into site preparation, foundations, frame, roof, services and finishes. Each in turn splits further. The tree makes the scope visible, so nothing is missed and nothing is built twice, and it is the skeleton for the schedule, the estimate and the contracts that follow.
Here is the trade-off in miniature. Tick the pieces of scope; then choose how many people to put on site. More workers finish sooner, but each new worker adds less, because people get in each other's way. The time-related cost (wages and daily site overheads) falls at first as the job shortens, then rises again.
Illustrative: each person's productivity is 1 ÷ (1 + crowding × (crew − 1)), wage $300 a day, site overhead $1,500 a day, labour-days and material costs invented for the example (the same logic as Fred Brooks's 1975 observation about adding people to late software projects).
Once the job is in pieces, each piece has a duration and some must wait for others: you cannot raise the frame before the foundations. Draw that as a network and you have what developers at DuPont and Remington Rand worked out in the late 1950s as the critical path method (CPM), around the same time as the US Navy and Booz Allen Hamilton developed the related PERT. The method is a pair of sweeps. The forward pass finds the earliest each task can start and finish, assuming every predecessor finishes as early as possible. The backward pass finds the latest each can finish without making the whole project later. The difference between them is the task's float.
The critical path is the longest chain of tasks from start to finish. Its length is the project's duration, and a day lost on any task along it is a day lost for the project. Tasks off the path have float: they can slip that many days without consequence. Below is a house in nine tasks, including a window order that takes twelve days to make. Pick a task and change its duration. Shorten a critical task and the whole project shortens (until another path becomes critical). Shorten a non-critical task and nothing happens. Delay "Services" by a few days and watch where float runs out.
A=Excavate, B=Foundations, C=Frame, D=Roof, E=Services, F=Cladding and windows (needs D and I), G=Interior finishes (needs E and F), H=Handover, I=Windows made (no predecessor). Finish-to-start links only; durations in days; invented for illustration.
CPM assumes you have as many workers as you need. You never do. Draw the number of workers on site on each day of the schedule as a bar chart (a resource histogram) and you will find peaks and valleys. Peaks cost: the extra crane, the extra bunkhouse, workers with nothing to do the day after. Resource levelling shifts tasks that have float, so that crews overlap less. The finish date does not move, because float is free. The peak does.
Here the same nine tasks have crews attached (for instance, six people on the frame, five on services). Left alone and started as early as possible, the roof and services overlap and the site peaks at 11. Services have five days of float, so slide them back and the peak falls. Change a duration in the network above and both charts redraw: the schedule you edit is the one levelled.
Crew per task: A 3, B 4, C 6, D 6, E 5, F 4, G 6, H 1, I 0 (made off site). The algorithm tries each start between earliest and latest, keeping the arrangement that minimises the sum of squared daily crews. Illustrative.
Real levelling is harder (software uses rules of thumb because the exact problem is famously hard), but the principle is what a graduate keeps: float is a resource you can spend.
An estimate is a prediction made when you know least. The common practice, set out by AACE International, classifies estimates by how much of the design is done: an early concept-stage estimate can easily be out by tens of percent either way, while one made near tender, when drawings are nearly complete, is within roughly ten to fifteen percent. The honest way to state an estimate is as a range, not a single number.
The second trap is that the sum of the most likely costs of each package is not the most likely total. Costs are skewed: a package is more likely to overrun by a lot than to come in far below. The sum of the most-likely values is lower than the average outcome. The solution used on serious projects is a quantitative risk analysis: give each package a low, most likely and high value, simulate thousands of combinations (a Monte Carlo simulation), and read off the cost you can be 50% or 80% sure of. The gap above the base is the contingency.
Seven invented work packages with low / most-likely / high costs (in $ thousands) such as Structure 270 / 300 / 420, triangular distributions, 4,000 trials with a fixed random seed. "Move together" means that in that share of trials all packages hit the same percentile (bad weather, a bad market). Illustrative, not a pricing method.
Two lessons come out of the simulation. Contingency is a calculated number, not a pad. And correlation matters more than people expect: if everything goes wrong together (a strike, a price spike), the risk does not average out, and the contingency you need grows.
A contract answers the question every project eventually asks: who pays when something goes wrong? How the work is organised determines the answer. Several families are in use around the world:
| Method | How it works | Trade-off |
|---|---|---|
| Design-bid-build | Owner hires a designer; a finished design is tendered to contractors, usually on price. | Clear roles, competitive price; slow, and designer and builder do not talk until late. |
| Design-build | One firm is responsible for design and construction. | Faster, fewer design disputes; the owner has less control of detail. |
| Construction manager at risk | A builder advises during design, then guarantees a maximum price. | Early buildability advice; depends on the quality of the guarantee. |
| Alliance or integrated delivery | Owner, designer and builder share risk and reward under one agreement. | Aligns incentives; needs trust and open books. |
| Public-private partnership | A private consortium designs, builds, finances and often operates for decades. | Private money and long-term upkeep; the public pays over decades. |
Standard forms differ by country. FIDIC, the international federation of consulting engineers, publishes widely used international forms (its colour-coded Red, Yellow and Silver books, first issued in 1999 with second editions in 2017). In the UK the NEC suite and JCT forms are common, while US projects often use AIA or ConsensusDocs forms. Public work is tendered under national or regional procurement law, such as EU procurement directives. Price terms vary too: a lump sum fixes the price and puts the overrun on the contractor; cost-plus reimburses actual cost plus a fee and puts it on the owner; a target cost shares the gap. Slide through those three.
Target $10.0 M, fee $0.5 M. Lump sum: owner pays target + fee. Cost plus: owner pays actual + fee. Target cost: owner pays actual + fee + share × (target − actual). Simplified; real contracts add caps, change orders and claims. Illustrative.
No form is best. A lump sum rewards efficiency but invites cutting corners and fighting over changes; cost-plus is flexible and gives the builder no reason to save; sharing the gap, as in NEC-style target cost contracts, tries to keep both sides pulling together.
The most reliable fact about large projects is that they run over. Bent Flyvbjerg, with Mette Skamris Holm and Søren Buhl, studied 258 transport infrastructure projects worth about $90 billion (published in 2002). Costs were underestimated in about 9 of 10. Averages: rail projects ended 44.7% above their estimate, bridges and tunnels 33.8%, roads 20.4%. In Flyvbjerg's 2014 overview of megaprojects (projects of a billion dollars or more), overruns of up to 50% were common and over 50% not uncommon: the Sydney Opera House 1,400%, the UK NHS IT system 400–700%, Boston's Big Dig 220%, Denver International Airport 200%, and the Channel Tunnel 80%, all in real terms. Overruns had stayed high for the 70 years for which comparable data exist, and all countries with data suffered them. In later work with a database of more than 16,000 projects, he reports that only about 0.5% meet cost, schedule and benefit targets together (secondary reports of his 2023 book).
Why? Flyvbjerg's answer is two-part: optimism bias (planners imagine this project going well, a psychological effect Daniel Kahneman and Amos Tversky named the planning fallacy) and strategic misrepresentation (the projects that win approval are those that look cheapest, so there is a selection pressure for under-estimates). The remedy he advocates is the "outside view": ask how projects of this class actually finished, and uplift your estimate accordingly. The idea is not undisputed. Peter Love and Chukwuka Ahiaga-Dagbui argued in 2018 that the study's methods and its "error or lie" framing are flawed. Try the reference-class idea on a budget of your own.
Averages (rail 44.7%, fixed links 33.8%, roads 20.4%) are from Flyvbjerg, Holm and Buhl (2002), 258 projects, 1927 onward, and describe typical outcomes, not any single project's forecast.
Construction is among the most dangerous peacetime occupations. The numbers below are official counts, and they are not directly comparable: the US counts by occupation, the EU by accidents at work, Britain by incidents reportable to the regulator, and workforces differ in size.
| Where | Year | What was counted | Source |
|---|---|---|---|
| United States | 2024 | 5,070 fatal work injuries in all jobs; 1,032 of them to construction and extraction workers, about one in five | BLS Census of Fatal Occupational Injuries |
| European Union | 2023 | 792 fatal accidents at work in construction, 24.0% of all; 6.3 per 100,000 people employed, third highest among sectors after mining and quarrying (10.8) and agriculture | Eurostat |
| Great Britain | 2023/24 | 51 construction workers killed (47 the year before); provisional 2025/26 figures show 25 of 126 worker deaths | Health and Safety Executive |
In the US, falls, slips and trips are the largest single cause: in 2023 they accounted for 421 of the 1,075 deaths in the construction industry, about 39% (BLS). The International Labour Organization lists falls, electrocution, crush injuries and being caught between objects as the leading hazards worldwide. The standard way to think about them is the hierarchy of controls: first eliminate the hazard (design out the work at height), then substitute, then engineer a control (guardrails, scaffolds), then change procedures, and only last rely on personal protective equipment. Personal protective equipment such as a harness is the weakest layer; designing the hazard out is the strongest.
Deaths are a lagging indicator. Good sites watch leading ones: near-miss reports, stop-work authority and whether the newest labourer will tell the foreman something is wrong. That is psychological safety (34/53) with literal consequences. For the engineering side of how designs fail, see the Hyatt Regency case in Engineering Drawing & CAD (34/18).
Building information modelling (BIM) replaces a set of 2D drawings with one shared 3D model whose elements carry data: this beam, this duct, this supplier, this cost. The first payoff is clash detection: finding in the model that the ventilation duct runs through a structural beam, which is cheaper than finding it on site with the crane booked. Later the same model feeds estimates, the construction sequence and the facility manager. ISO 19650, published from 2018, sets an international framework for managing information on a BIM project, building on earlier British standards.
Governments have pushed it: the UK set out a strategy in 2011 aimed at collaborative 3D BIM on its projects by 2016, and Hong Kong made it mandatory for government projects above HK$30 million from 1 January 2018. A 2004 NIST study estimated that poor interoperability cost the US capital facilities industry $15.8 billion a year. Cautions are fair, too: adoption outside rich countries has been slow, a 2019 claim of savings up to 20% was an official estimate and not a measured result, and the industry warns about "BIM-washing", a label on work that does not use it properly. The model is only as good as the information put in.
Break the job into pieces, find the critical path, and spend the float. Level the crews so the peak does not set the bill. Price with a range, and hold a contingency that a simulation, not hope, supports. Choose a contract that puts risk with whoever can best control it. Look at how similar projects actually ended before believing yours will differ. And design out the fall before you hand out the harness.
Try this: the next time you pass a building site, pick one activity you can see and ask what must happen before it can start and what is waiting on it. You have just drawn a bit of a network.
Sources. Project triangle: Martin Barnes, 1968 thesis and 1969 course, as described in the Wikipedia article on the project management triangle. CPM and PERT: late 1950s, Walker and Kelley (DuPont, Remington Rand); Booz Allen Hamilton and the US Navy. Cost estimate classes: AACE International (ranges are described in rough terms here). Overruns: Flyvbjerg, Holm & Buhl, "Underestimating Costs in Public Works Projects: Error or Lie?", Journal of the American Planning Association 68(3), 2002 (258 projects, about US$90 billion; rail 44.7%, fixed links 33.8%, roads 20.4%); Flyvbjerg, "What You Should Know About Megaprojects, and Why: An Overview", Project Management Journal (2014); Flyvbjerg & Gardner, How Big Things Get Done (2023), figures for the 16,000-project database taken from secondary reports; critique: Love & Ahiaga-Dagbui, Transportation Research Part A 113 (2018). Safety: US Bureau of Labor Statistics, Census of Fatal Occupational Injuries (2024 release; 2023 data for the falls figure); Eurostat, accidents at work by economic activity (2023); UK Health and Safety Executive, work-related fatal injuries in Great Britain. BIM: ISO 19650; NIST GCR 04-867 (2004) as cited in the Wikipedia article on BIM; national mandates as described there. FIDIC, NEC, JCT, AIA and ConsensusDocs forms are named from general knowledge. Interactive figures were computed and checked in code; datasets are illustrative.