Flyvbjerg and Gardner’s How Big Things Get Done

 Projects Minor, Major and Mega

 

 

 

These days we live and work in a world of projects, with everything from planning a holiday, a product launch or a political campaign seen as a project. Organisations have project teams and use project-based management systems. There is a project management book of knowledge, known as the PMBoK, taught in the many PM courses now available.

 

Major projects, and public projects in particular, are frequently associated with cost blowouts and schedule slippages. Some projects become notorious for cost overruns, like the Sydney Opera House (1,400 percent), the Scottish parliament building (1,000 percent) or Boston’s Big Dig tunnel (600 percent). In fact, the great majority of large projects like airports, pipelines, tunnels, railways and roads are not delivered on time or within budget. Software projects are rarely delivered on budget. However, about 20 percent of major projects are delivered on time and on or below estimated cost, so although that may be difficult and unusual it is not impossible.

 

The 20 percent success rate comes from a database of major projects built by Bent Flyvbjerg, a Danish researcher now at Oxford University. Starting in the early 1990s he began collecting project cost and time data, initially for transport (roads, rail and bridges) then extended to include water, power, oil and gas, IT, and aerospace projects. That data became the basis for many journal papers on project performance and the 2003 book by Flyvbjerg, Bruzelius and Rothengatter Megaprojects and Risk: An Anatomy of Ambition. 

 

Flyvbjerg and his colleagues coined the phrase ‘Survival of the unfittest’ to describe projects that get approved and built despite their poor economic and financial characteristics and outcomes. Their key characteristics of projects are:  

• They are inherently risky due to planning horizons and complex interfaces between the project and its context, and between different parts of the project;
• Costs and benefits are many years in future, and are large enough to change their economic environment with unintended consequences;
• Stakeholder action creates a dynamic context with the escalation of commitment driven by post hoc justification of earlier decisions;
• Decision making and planning are processes with conflicting interests;
• Often the project’s scope or ambition changes significantly after starting work;
• No allowance is made for unplanned events (known as ‘black swans’) so budget and time contingencies are inadequate;
• Misinformation about costs, benefits, and risks is the norm, and in some cases is strategically misrepresented to get a project started and ensure commitment;
• The result is cost overruns and/or benefit shortfalls with a majority of projects.

 

There is also the 2023 book from Flyvbjerg and Dan Gardner, How Big Things Get Done: The Surprising factors Behind Every Successful Project, From Home Renovations to Space Exploration. Thanks to Gardner’s contribution this is a brisk, readable book, not another academic tome, and it made its way onto the best business books lists of the Economist, Financial Times and McKinseys. Although it covers the factors and issues in the 2003 book there is less data and analysis, and this book has more examples of different types of projects (buildings, films, tunnels, railways etc.) and the people responsible. Each of the nine chapters addresses a specific issue, illustrated by interesting stories about the people and projects featured, and presents a key concept, the ‘universal drivers that make the difference between success and failure’. 

 

The first chapter is ‘Think slow, act fast’, meaning plan thoroughly and as completely as possible before staring work, or ‘think first, then do.’ Once started a project should be delivered as quickly as possible, to reduce the risk of something going wrong. The chapter has data from Flyvbjerg’s database of 16,000 projects: 91.5% go over their time and budget; 99.5% go over cost and time and under-deliver on benefits. The typical project has underestimated costs and overestimated benefits, and the risk of a project going disastrously wrong (not 10%, but 100% or 400% over budget) is surprisingly high. 

 

Chapter 2 is ‘The commitment fallacy’, where projects are approved before alternatives are explored and/or continued after money has been spent (the start digging a hole strategy). Strategic misrepresentation is an organisational and institutional explanation where project promoters produce biased appraisals at the approvals stage (underestimated costs + overestimated benefits = approval) and projects that get funded are ones that look best on paper (i.e. have largest errors) not the best projects. Another explanation is psychological, from Daniel Kahneman’s ‘planning fallacy’ for decisions based on delusional optimism about the time needed to complete tasks. Premature commitment leads to poor outcomes because people assume What You See Is All There Is – the WYSIATI fallacy – focusing exclusively on what is in front of them and not exploring alternatives. 

 

In chapter 3 a kitchen renovation is the example, a project that expanded and grew after starting and blew its budget. Although planned well it did not start by asking ‘why are you doing this?’ The point is to decide what the project is for first, before thinking about how to achieve that goal. The first requirement for a successful project is to select the right one, and whether or not to proceed. Chapter 4 is ‘Pixar planning’, which is spending a lot of time exploring an idea with many iterations to get to proof of concept stage. Often repeated advice is the three words ‘Try. Fail. Again.’ The authors say we are good at learning by tinkering, ‘which is fortunate because we’re terrible at getting things right the first time.’ Chapter 5 argues for the importance of experience and tacit knowledge, and shows how common it is for such a basic insight to be ignored. 

 

Chapter 6 introduces Reference Class Forecasting, a solution to optimism bias and the illusion that a project is unique. This involves three steps: Identification of a relevant reference class of past, similar projects; establishing a probability distribution for the reference class; and comparing the specific project with the reference class distribution. From the comparison reliable forecasts of a project’s budget and schedule can be made. In chapter 7 the idea that ingenuity and creative chaos leads to great outcomes is refuted, it is the occasional success, which is an exception, that makes this such a good story. 

 

The importance of getting the team right is Chapter 8, with British Airport Authority’s 2007 Heathrow Terminal 5 the example project. This was a famously successful megaproject. The delivery of T5 on time and on budget, with a remarkable safety record, was due to the three inter-related factors of risk management, integrated teams, and the alliance contract. BAA held all the risk and an incentive contract meant suppliers and contractors were motivated to find solutions and opportunities. BAA used in-house project management teams where traditional boundaries were broken down and replaced by colocation, so people from different firms worked in integrated teams in BAA offices under BAA management. The focus was on solving problems before they caused delays.

 

Chapter 9 argues modularity delivers projects, faster, cheaper and better because it allows repetition, and repetition allows learning by doing. Rather than building one huge thing the Lego approach is to make modules that can be assembled into buildings, cars, cakes, satellites and subway stations. In the database the most successful projects (i.e. least likely to have cost overruns) are energy projects for solar, wind and thermal generators that are inherently modular. At the other extreme are nuclear power plants and waste storage, hosting Olympic Games, and hydroelectric dams, ‘all classic ‘one huge thing’ projects’. The chapter closes with an appeal to address climate change through building out the energy transition as quickly as possible. 

 

The book ends with eleven heuristics for better project leadership that collect the book’s key points. These are ‘rules of thumb used to simplify complex decisions’ such as: Hire a masterbuilder; Get the team right; Take the outside view (i.e. use a reference class); Build with Lego; Think slow but act fast; Think right to left (i.e. start with your goal, then identify the steps to get there); and Say no and walk away. Although these may seem obvious, the point is how often they are not followed and how many projects go over time and budget, and areled by people with only partial competence with no provision made for black swan events.

 

Flyvbjerg and Gardner argue a significant reason for poor decisions on projects is unwarranted optimism about outcomes, the planning fallacy. Planners underestimate the time, costs, and risks due to size, gestation and time for delivery, and overestimate the benefits of projects. In some cases there is strategic misrepresentation of costs and benefits. After a project has started there are the risks of escalated commitment and lock-in, scope changes, and conflicting interests. None of these risks are unknown or mysterious, which raises the question of why so many projects have such poor outcomes. 

 

The answer is often the quality and competence of project managers. A 2016 infrastructure report from the McKinsey Global Institute, the think tank for the management consultancy, found ‘Cost overruns for large projects average 20 to 45%. We often see cost differences of 50 to 100% in similar projects carried out by different countries, even those in similar income levels. If countries apply the best practices that have already been proven effective elsewhere, they can achieve remarkable results.’ McKinsey argued a key factor was the quality of the project manager, as their research ‘across thousands of projects indicates that top quartile project managers consistently deliver projects ahead of time and below cost, whereas the opposite is true for the bottom quartile’. 

 

That said, how likely is it that project managers will read How Big Things Get Done? Probably not enough, if McKinsey is right about how little best practices are copied. Although the book is about projects, it does not specifically include or refer to the PMBoK toolkit of processes and knowledge areas, that project management qualifications are based on. Also, while the examples used of architects like Jørn Utzon and Frank Gehry, Pixar movies, iPods, the Empire State Building and Heathrow Terminal 5 are interesting and revealing, because they are unusual and exceptional projects many project managers might not accept that the lessons taken from those projects are widely applicable.

 

It may be the real audience for the book is clients and owners rather than project managers. The client ultimately has responsibility for a project, even if they try to unload this onto a project manager. Much of the advice, on project selection, planning, iteration, contingencies and modularity for example, is about the development stage of a project when the client is or should be in control, not the delivery stage after work commences when the project manager is responsible. And the important message the book sends is that the success or failure of the great majority of projects is determined early on, during planning and development.  

 

 

                                                                      *

 

 

Bent Flyvbjerg and Dan Gardner, 2023. How Big Things Get Done: The Surprising factors Behind Every Successful Project, From Home Renovations to Space Exploration. New York, Currency Press. 

 

Flyvbjerg, B., Bruzelius, N. and Rothengatter, W. 2003. Megaprojects and Risk: An Anatomy of Ambition, Cambridge, Cambridge University Press.

 

Review of Ed Merrow's book Industrial Megaprojects

 

It is well known that the future is uncertain, where uncertainty is an unmeasurable or truly unknown outcome, often unique. This can be clearly seen on large infrastructure projects, which often bring into focus the issues around project selection. A remarkable number of these projects are unsuccessful, by exceeding their time and cost estimates, or inefficient because their returns and/or benefits are well below forecasts.

Major infrastructure projects and other megaproject costs and benefits are many years into the future, and any estimates of them will depend on the assumptions and type of model used. They change their economic environment, generate unintended consequences, and always have the possibility of escalation of commitment driven by earlier decisions.

Ed Merrow did the first published study on major projects costing over US$1 billion (known as megaprojects) for the US military think tank RAND Corporation in 1988, on 52 private sector projects – refineries, oil, transport, and nuclear. It looked at time and cost performance and the factors that drive the outcomes on these projects. Most met performance and schedule goals, but only four came in on budget with an average cost growth of 88%. He concluded “The larger the project, the more important is the accuracy of early estimates.” (1988: 80). This remains the key issue.

Merrow set up Independent Project Analysis to provide project research for heavy industry and the process and extraction industries. Depending on the project, between 2,000 and 5,000 data points are collected over the initiation, development and delivery stages. From this database companies can compare their project with other, similar projects, across a wide range of performance indicators. The data gives estimates on approval, design and documentation, and delivery times for the type of project, and allows for factors like location, access and complexity in costs. When Merrow published his book Industrial Megaprojects in 2011 the IPA database had 318 megaprojects, out of about 11,000 projects in total, from industries like oil and gas, petroleum, minerals and metals, chemicals, and power, LNG and pipelines.

In his 2011 book Merrow recommended a process he called front-end loading, and his best examples of project-definition reduced project timelines and cost by roughly 20 percent. He saw projects having three stages, the first evaluates the business case, the second is scope selection and development, and the third is detailed design. His argument was that there needs to be gates between the stages that prevents less viable projects from getting to authorisation. He emphasises the ‘period prior to sanction of the project.’

Using evidence from the 11,000 projects in his database Merrow argued the best form of project delivery is what he called ‘mixed’: hiring engineering design contractors on a reimbursable contract then construction contractors on a separate fixed price contract. His view was this is the most effective form of project organization, basically traditional construction procurement where consultants are appointed to do the design and a competitive tender is held for one or more contractors to execute the works against a complete design.

Merrow also argues the owner’s job is to select the right project and the contractor’s job is to deliver the project as specified, on time and on budget. In his view contractual relationships are more tactics than strategy, and cannot address any fundamental weaknesses in the client’s management of the project, in particular the client ultimately has to own the design. This crucial point became widely recognised by the private sector clients/owners of large engineering projects that Merrow studied, because they understood that significant risk transfer from clients to contractors is structurally impossible on the projects they undertake.

Design and delivery of major projects can be contracted separately to reduce project costs and risks so that, as far as possible, design and documentation is complete or nearly complete before tendering. The ‘nearly complete’ qualifier is important. A simple project can be fully specified just because it is simple. However, there is a limit to how much design can be completed in the initial stages of a major project, because the specification of a major project develop over time as the project details are refined and defined. Therefore, it is unreasonable to expect a major project to be fully specified at tender, and in most cases this would not be possible. On the other hand, it is not unreasonable for tenderers to expect the documentation they receive to be sufficient, because the extent and clarity of the design determines their project time and cost plans.

There are a number of advantages of this strategy of unbundling design and construction, particularly for major projects. Breaking a project into smaller, sequential contracts spreads the cost out over time, and does not incur interest costs if a loan is not used for design work. It makes quality control easier and more effective, by being focused on each stage, which is an important risk management tool. Separating the design stage from tendering and construction will also improve opportunities for consultation with the community and stakeholders. Most importantly, completion of design and documentation before tendering reduces contractor risk and therefore total project cost.

This argument is for design and construction of projects to be contracted separately, because this will reduce project costs and risks. As far as possible, design and documentation should be complete or nearly complete before tendering or starting the works. The key factor is therefore the extent of the specifications, on some projects there may be a limit to how much design should or could be completed upfront. For many major projects these develop over time as the project details are refined and defined. It is unreasonable to expect a complex project to be fully specified at tender, and in most cases this would not be possible. It may also be advantageous to look for innovative ideas or design options, so for these projects an incremental approach would allow contractors and suppliers the opportunity for input during the development of the design. This also has the advantage of reducing uncertainty from poor tender documentation, thus lowering risk and cost for tenderers.

To deliver better results in on-time and on-budget delivery, Merrow argues project developers or sponsors should spend 3 to 5 percent of the cost of the project on early-stage engineering and design. This is because the design process will often raise challenges that can to be resolved before construction starts, saving time and money.

If more realistic, and therefore more accurate, time and cost estimates were given for major infrastructure projects before they are approved, and during the design and development stages, there would be fewer recriminations about project performance and less incentive to find scapegoats on completion, which is typically over budget and schedule. There would be fewer of the common accusations of poor productivity, management failures or poor planning, thus lessening the atmosphere of acrimony that often surrounds major projects in their later stages

Merrow, E.W. 1988. Understanding the Outcomes of Megaprojects, RAND Corporation, Santa Monica.

Merrow. E.W. 2011. Industrial Megaprojects: Concepts, Strategies and Practices for Success, Wiley, Hoboken, NJ. Second edition 2024. 

Can AI Solve the Construction Productivity Problem?

 Will artificial intelligence reduce hours worked or time taken to complete a project?

 

 

There has been little or no growth in construction industry productivity for many decades, and there have been many different reasons given. These include larger and more complex buildings, longer and more global supply chains, increased regulation and compliance requirements, temporary project teams, the cycles of boom and bust in construction work, increasing use of prefabrication and offsite manufacturing, the number of low productivity small firms, low industry capital intensity, and a lack of workforce skills and training. All these and more contribute to a complex problem. 

 

Over the decades various solutions have been proposed, such as total quality management in the 1980s, lean construction in the 1990s, BIM in the 2000s, and digital twins in the 2010s. Often criticised as top-down, technocratic solutions when imposed by governments, with BIM mandates for example, the reality is the great majority of firms in the industry, particularly small and medium sized ones, have not adopted them unless required by clients. Clients, in turn, have not been prepared to increase costs or risk innovation on their projects to benefit the industry, a classic catch 22 situation.

 

Now there is a new solution to the productivity problem. Since the launch of Chat-GPT in November 2023 there have been dozens of artificial intelligence (AI) models released, and widely used software from majors like Microsoft, Oracle, Trimble and Autodesk now has AI incorporated into their systems. Will this time be different, or will AI follow BIM and the other solutions proposed in the past, and be adopted piecemeal when required without changing industry productivity and performance? 

 

In the previous post around 100 companies with construction related AI systems were included, divided into six areas: preconstruction and estimating; project and document management; site monitoring and safety; equipment management and maintenance; design and planning; and materials. This post discusses the potential of AI to solve the construction productivity problem. It looks at how AI might reduce the number of hours worked or decrease the time taken to deliver projects, considering its potential effects in equipment management and maintenance, preconstruction and estimating, project and document management; and site monitoring and safety. Issues in worker and workplace monitoring are discussed, because there is also the possibility AI will make the industry a worse place to work, which has been the experience of call centres, warehouses and Uber drivers, industries where AI is already intensively used. 

 

Can AI Improve Construction Productivity?

 

Labour productivity is the ratio of an industry’s inputs, usually the number of hours worked, to output, the value of production minus the cost of inputs. To improve productivity the relationship must change, for example increasing output faster than hours worked or decreasing hours worked while maintaining output will both increase productivity. Although that seems pretty straightforward, there are many issues that complicate measurement, ranging from how data is collected and managed to adjustments for price and quality effects. Those data issues are more difficult for capital productivity, which measures the contribution of buildings, structures, plant, machinery and equipment to output, adjusted for age and depreciation. 

 

Also, productivity does not change much from year to year, currently increasing by less than one percent a year for the Australian economy as a whole, and so is gradual and cumulative, and works over the long-run. Even in high growth periods and economies, annual productivity increases of around three percent are common. Therefore, productivity is not a good target for industry policy. Finally, the statistics used for construction include onsite work and people employed by contractors and subcontractors only, and thus exclude important activities like offsite manufacturing and prefabrication, and design, technical and professional services.

 

There are three factors that will affect improvement in industry productivity from AI. First is the adoption rate of AI by firms, which will be the key driver because the more firms that use AI the bigger the effect. However, second, AI productivity gains may be concentrated in a few firms, lowering the overall level of industry productivity improvement. Third, some or many physical or manual tasks may be affected by AI through robotics, which is outside the discussion here. Onsite construction robots were covered in this recent post. 

 

A November 2024 OECD publication on the productivity effects of AI over the next decade included industry estimates of the share of tasks affected by AI using two measures, a baseline exposure for the share of tasks that time for completion substantially decreases using Generative AI, and an expanded capabilities measure that includes tasks where gains are achievable if other software is developed on top of current systems. The largest gains are in ‘knowledge-intensive services that rely strongly on cognitive tasks, such as Finance, ICT services (including telecoms), Publishing and Media, and Professional services. The least exposed sectors include sectors with a strong manual component, such as Agriculture, Mining and Construction.’

 

Figure 1. Share of tasks affected by AI

 

Source: Miracle or Myth? Assessing the macroeconomic productivity gains from Artificial Intelligence. OECD Working Paper. 2024: 19.

 

The OECD study looked at tasks and the extent to which AI can do or help with. High exposure means it has a large share of tasks that AI can assist with, and there are only half a dozen high exposure industries. Although construction is at the lower end of the range, 20 to 30 percent of tasks could be AI assisted, which is comparable to metal and plastic manufacturing and warehousing, and not far behind transport, fabrication, motor vehicles, and water and waste. There are, therefore, plenty of potential productivity gains for construction from AI.

 

How AI Might Improve Construction Productivity

 

There are two ways AI might affect construction productivity: it could reduce the total number of hours worked on a project or it could decrease the time taken to complete a project. Note, however, if AI had both of these effects they would cancel out and productivity would not change.  

 

The most straightforward use of AI is for equipment management and predictive maintenance. If these increase the availability and reliability of machinery and equipment there will be an increase in productivity from less downtime for operators, hours worked might not change but output would increase, and that would also decrease time taken to complete. The impact and importance of this on productivity depends on how unreliable equipment currently is. If contractors and operators are aware of the liabilities and costs of breakdowns and do adequate maintenance, any improvement from AI will be at the margin, however both labour and capital productivity would improve. 

 

In preconstruction and estimating, AI could cut time and reduce mistakes. All construction firms do estimates, and using AI and image recognition for automated quantity take-offs from drawings, PDFs and BIM, and scanning drawings to count items and check for missing information and risk assessment seems an obvious way to save time. That data can be given to estimating systems that also do BQs, cost plans, bidding automation, indirect costs, KPI analysis, and predictive analytics. For larger firms doing design and build work, AI systems that do generative design, site layouts, materials selection, planning and code compliance checks would be useful. Although important for individual firms, the effect on overall industry productivity of reducing hours worked in these offsite functions would be offset by the much larger number of hours worked onsite. AI could make a major difference to productivity if it improves project management.

 

Because of the large number of different project documents used, copilots, AI assistants and chat bots that search, summarise and analyse those documents on request is another obviously useful task for AI that would be relevant to most firms in the industry. If site and project managers spend much time searching for information and retrieving documents, and suppliers and subcontractors have to wait for instructions, this would be an efficiency gain that could reduce both hours worked and completion time for projects. 

 

Document management can also include project management functions like bidding, billing, payments, RFIs, claims and compliance management. Other AI assisted PM offers are task plans and schedules, risk assessments, status reports, workflows, resource management, ITPs and quality assurance, and site management. Some systems have collaboration and communication tools. How much difference could using AI for these make to PM and productivity? Tasks like scheduling, work plans and status reports are time intensive, so AI can improve productivity by reducing the hours required. Better onsite collaboration and communication can potentially reduce the time to complete, particularly on large projects where coordination is more important. 

 

For AI enhanced PM to improve industry productivity it would have to become widely used, and that depends on how much it improves the capability and competitiveness of firms using AI. If there are significant benefits to early adopters of AI they will be able to out-compete other firms, that will then have to become fast followers to survive. Whether or not AI will deliver meaningful reductions in hours worked, and those significant benefits are available and achievable, is an open question at this point. 

 

Finally, site monitoring and safety. Reality capture and computer vision systems match site work to digital twin and BIM models to track progress.  These design vs as-built comparisons can be done by drones or hand-held cameras and provide daily or real-time progress tracking, with time saved on reports and documentation a useful improvement, and possibly leading to an important reduction in disputes and defects. However, the main effect on productivity will depend on how well this information is used to improve project management, particularly collaboration and communication. 

 

Safety systems monitor sites and detect hazards, and dangerous behaviour by people, and can prepare safety plans. While important, this will not have much effect on productivity, although reducing injuries and time off work for recovery would have an indirect effect. Some of these systems also identify people using sensors, cameras and wearables like beacons and badges, and can also be used for access control, headcounts, timesheets, task and workforce management. Again, there can be time savings in these administrative functions, however, worker and workplace monitoring raise a number of important issues. 

 

Monitoring and Algorithmic Management

 

In Australia workplace monitoring and surveillance is covered by a mix of federal and state laws. The Privacy Act 1988 doesn't explicitly address workplace surveillance, but has 13 Australian Privacy Principles that employers must follow when handling personal information and data collected through monitoring. Employers can monitor employees to ensure they are doing their work and using resources appropriately, but should only monitor employees to the extent necessary to achieve a legitimate business purpose. Employees have to agree to have their data shared. The states have two laws that regulate workplace surveillance, one is a Privacy Act requiring compliance with the Australian Privacy Principles (in NSW a Workplace Surveillance Act), and the other a Surveillance Devices Act that regulates the use of surveillance devices in public places and workplaces [2].  

 

Monitoring with AI becomes controversial when it is used for ‘algorithmic management’, where workers are monitored, ranked, rated, given instructions, potentially harmed or auto-fired by an AI system. In the current dispute between Woolworths and the United Workers Union a worker performance management program called ‘The Coaching and Productivity Framework’ that was introduced across warehouses is a major issue. The December 7^th Sydney Morning Herald article on the dispute said the ‘point of contention is the disciplinary action workers face if they fail to achieve 100 per cent adherence to pre-determined pick rates. What was previously a non-enforced goal is now a mandatory requirement.’

 

In the US the Federal Trade Commission has authority to block unfair trade practices, and a recent speech by commissioner Alvaro Bedoya argued algorithmic management is an illegal unfair trade practice. Because it causes substantial injury, can’t be reasonably avoided, and isn’t outweighed by a countervailing benefit, Bedoya says algorithmic management satisfies the three criteria for an unfair trade practice.

 

He gives three examples. First, on substantial injury, Bedoya describes a warehouse worker injured while working for ecommerce sites monitored by an AI that required him to pick and drop an object off a moving belt every 10 seconds, for 10 hours a day. Workers are tracked by a leaderboard, supervisors punish workers who don’t make quota, and the algorithm auto-fires them if they fail to meet it. Second, Bedoya describes the experience of New York rideshare drivers when the apps start randomly suspending them, telling them they can’t book a ride. Drivers who stop for coffee or use a bathroom are locked out for hours as punishment. Third, on average a call-centre worker is subjected to five forms of AI, with AI video and voice monitoring to measure empathy, an AI timing calls, and two more to analyse calls. AIs produce transcripts of calls, but workers with accents find them ‘riddled with errors’ and their performance assessment is based on the transcripts. 

 

Given these considerations, construction employers who hope AI monitoring of workers will improve productivity should approach this idea with great caution. Workers have the right to know what data is being collected, who it’s being shared with, and how it’s being used. 

 

Conclusion

 

Industry productivity is the combined effect of both capital and labour productivity, and AI adoption can lead to broad and long-run gains in construction productivity as AI assistants take over time consuming tasks like report writing and document management, and AI systems improve resource and project management. The potential productivity gains from AI are significant, but there are factors that could limit the effect, such as the willingness of firms to adopt AI, their capabilities, skills, and access to data. Also, AI adoption across the industry is likely to be uneven, and the productivity benefits of AI will tend to go to firms that use other advanced technologies and have IT skills. 

 

Two construction tasks that AI could significantly improve are estimating and project management, because it can synthesise, summarise and interpret data, and provide insights and suggestions, although it cannot replace expertise and requires supervision and checking of results. AI would also make a major difference to productivity by saving time used for document management, through quicker access to information, by reducing mistakes, disputes and defects, and could lead to more and better collaboration and communication. 

 

To improve construction industry productivity, AI has to increase efficiency and reduce the hours worked or decrease the time taken to complete a project, although if AI had both of these effects they would cancel out and productivity would not change. Therefore, productivity is not the most appropriate metric to use for industry-wide effects of AI, rather it should be used to measure productivity improvements in specific tasks where AI can assist workers. 

 

In addition to productivity there are other labour market implications not discussed here, such as the effects of AI on recruitment, retention and wages, and ongoing demographic changes such as the ageing workforce. Also, the history of technologies like electricity, internal combustion engines, computers and the internet show it typically takes two to three decades for a new technology to become widely used and to significantly affect the economy and productivity. AI may be different, but probably not. Modern AI began with Alpha Go in 2016, so we are almost a decade on already and the next decade could see rapid uptake of the AI systems surveyed in the previous post.

 

At this stage it is impossible to know whether AI will deliver on its potential productivity gains. That will depend on how the technology continues to develop, how quickly AI is adopted and successfully applied, associated innovation-boosting effects, and how government policy on incentives, digital infrastructure, AI regulation and data develops. 

 

                                                                     *

 

[1] Progress in AI may be slowing down. GPT-4 is nearly two years old, but The Information said OpenAI’s new models have only incremental improvement, and Reuters quoted Ilya Sutskever, co-founder of OpenAI, saying results from scaling up pre-training have plateaued. 

 

[2] For more details see the Office of the Australian Information Commissioner about workplace surveillance: https://www.oaic.gov.au/privacy/your-privacy-rights/surveillance-and-monitoring/workplace-monitoring-and-surveillance

 

 

 

 

 

 

The Four Cs of Construction AI: Cost, Capability, Competence and Competition

 Construction artificial intelligence in 2024: Part 1

 

 

 

 

This year has seen artificial intelligence (AI) enter the mainstream. After decades of development, new methods of machine learning (ML) led to the release in November 2022 of Chat GPT by OpenAI, followed by GPT-4 and other large language models (LLMs) like Anthropic’s Claude and Mistral in 2023. At the end of 2024 these are available through widely used software like Microsoft’s Edge and Office (with Copilot, using Chat GPT), Google’s Gemini, and Llama from Meta. New computers come with AI installed, such as Apple Intelligence, another version of Chat GPT, on Macs and iPhones.  

 

Cloud service providers provide web access to AI as Models as a Service. Amazon Web Services has Bedrock, offering models from AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon for text generation, summarisation, question answering, and image recognition and generation. Companies can use their own labelled datasets or do pretraining with unlabelled data to adapt models to a specific domain or industry. Amazon SageMaker Studio has tools for preparing data and building, training, deploying and managing a ML model. Microsoft Azure offers nine models from OpenAI (Microsoft is a major investor), Phi, Cohere, Meta, Mistral, Hugging Face, Nixtla, Core 62 and Stability AI. Available in Azure AI Studio or Azure Machine Learning. Google’s Vertex AI Studio has 150+ first- party, third-party, and open-source models [1]. 

 

The actual extent of use of AI by companies rather than individuals is unclear. A US Census Bureau survey in March 2024 found ‘the fraction of firms using AI is relatively low but rising: from about 3.7% at the start of the collection in September 2023 to about 5.4% at the end of February 2024. AI use is expected to rise further to about 6.6% by Fall 2024. There is enormous variation in current use by sector from a low of 1.4% in Construction and Agriculture to a high of 18.1% in Information.’ The survey also found a much higher rate of AI use by people of 39% than by firms. 

 

Other surveys have smaller samples. For example, the annual McKinsey chief experience officer (CXO) survey in October 2024 surveyed 276 people in five corporate functions across 18 industries in North America and Europe [2]. McKinsey claimed: ‘In less than two years, generative artificial intelligence (gen AI) has become a mainstream tool with applications across almost every area of the economy.’ However, as Figure 1 from their report shows, that claim is nonsense, as the actual level of use is 22% in these large companies for functions where AI should be most applicable (80% of companies in the survey had revenue over USD$1 billion). For small and medium size companies the use rate will be much less. 

 

Figure 1. McKinsey CXO survey

 

The Deloitte August 2024 State of Generative AI report with 2,770 respondents also found limited use: ‘Generative AI efforts are still at the pilot or proof-of-concept stage, with a large majority of respondents (68%) saying their organization has moved 30% or fewer of their Generative AI experiments fully into production’.  The survey included use cases for customer service and report writing from eight industries, and 42% of respondents said productivity and efficiency gains or cost reductions had been the main benefit. The survey also found employees with AI skills were paid more. 

 

This post is divided into two parts. Part 1 below is a discussion of the current state of use of AI based software systems in construction, at this point in their deployment. It attempts to get some perspective on what is available and how construction firms have to balance the potential benefits of AI against the cost of subscribing to AI systems, how AI might improve the capability of firms, and the importance of employee competence in the skills and expertise required to use AI. It suggests AI can provide a competitive advantage for firms that successfully reduce costs with AI, and that could be from one or more specific point solutions rather than a platform. Firms will have to work with AI to survive.

 

Part two is a survey of companies with construction AI systems, divided into six areas: preconstruction and estimating; project and document management; site monitoring and safety; equipment management and maintenance; design and planning; and materials. It has thumbnail outlines of companies ranging from large and recognised technology leaders to startups [3]. It does not attempt to cover every new company everywhere with a construction related AI system, although this is a large and representative sample with nearly 100 companies listed. 

 

 

Will Construction AI Deliver Productivity and Efficiency Gains or Cost Reduction? 

 

Much of the commentary on construction AI is little more than marketing attached to the idea of industry ‘transformation’, often from industry observers rather than practitioners. Unsurprisingly, the websites of construction AI developers have glowing testimonials on their products. There are extravagant and exaggerated claims made, like this: Building the Future: Bluebeam AEC Technology Outlook 2025 found that 74% of AEC firms are now using AI in at least part of their projects, with 35% of respondents reporting cost savings between $100,000 and $500,000 through the use of new technologies. (BlueBeam is part of Nemetschek Group, a major software provider with Archicad and Vectorworks)

 

Unfortunately, there is no serious publicly available discussion from users of construction AI on the costs and benefits, beyond generalities about the potential for AI to manage tasks and solve problems like excess waste, energy use and defects, while increasing communication and coordination and reducing claims and disputes. The problem is, people in companies that are actively investigating and applying AI on their projects, if AI is delivering for them, aren’t going to tell their competitors. This could be taken as a sign that AI is in fact making a difference.

 

There are repetitive tasks that AI could do for almost all construction firms, like takeoffs for estimates, code compliance checks with image recognition, or using a text reader for document management. These would not require changing firms’ work processes and complement current skills. For contractors, progress tracking and site monitoring could deliver cost savings on workforce and safety management, and subcontractor supervision and payments, and reduce reporting cycles for quality control. 

 

The important measure would be cost savings, and a contractor, supplier or subcontractor could target one task where AI can reduce costs through automating repetitive and time consuming work. In an industry with fine margins and competition for work, reducing costs with AI in one task is an advantage even if the reduction is not great, because the competitors are similar firms with the same costs. Firms will have to start working with AI somewhere, sooner rather than later, if they are to survive. 

 

Productivity and efficiency gains are more difficult to measure. Improved communication and coordination would be a positive, and many AI systems offer that, but the actual effect on costs and project delivery may not be much. Then there is the promise that a professional (like an estimator, engineer, PM, or claims manager) would be more efficient with an AI assistant, but there is a learning curve associated with working with AI, and an assistant still has to be monitored for accuracy. Realising the potential efficiency gains from AI may take some time and require some degree of reorganisation. 

 

This is an argument for incremental gains from adopting and learning how to use AI, rather than the breakthrough leaps in performance advocates typically claim for AI. There are two reasons why this may be the case: first, there are many elements in the final cost of a project, so improvements in one or some of those can only affect total costs at the margin; and second, while AI might deliver a cost reduction that will not come for free, because there are costs for AI subscriptions and IT resources. 

 

That said, the extent and range of systems found in the companies included in the survey of construction AI in Part 2 of this post suggests that in the near future there will be significant productivity and efficiency gains and cost reductions for firms, regardless of size, that get their AI adoption and implementation strategy right. 

 

 

Diversity Across Six Areas of Application 

 

The diversity of construction AI offerings is striking. There are large, enterprise scale systems from major software developers like Autodesk, Bentley, Nomitech, Procore and Trimble that integrate with other software. Global contractors like AECOM, Bechtel and Skanska have been developing AI assistants and skills for several years. Then there are specialised, specific task systems, like those from equipment manufacturers for predictive maintenance, image recognition for cost estimating, reality capture for progress tracking and safety management, and generative design for architects. 

 

The companies outlined in Part 2 are divided into six areas. The divisions are not always neat, particularly for the larger systems because there are overlapping functions, and there are subsets within the six divisions. However, this is a reasonably comprehensive if not complete survey of construction AI at the end of 2024 with a large and representative sample of nearly 100 companies. Excluded are onsite automated and robotic equipment, covered here in a previous post, and 3D printing, covered here in a 2023 post.   

 

Preconstruction and estimating - AI is used for site layouts, submittal logs, compliance checks, takeoffs, estimating costs, bids and tenders. Two issues that AI and ML has to address are the lack of standardised industry data and extracting data from different sources, such as PDFs, BIM models and drawings. The first means companies have to build databases and rely on their own projects for data, and there are providers offering assistance with those. These systems use AI and historic data from projects for estimating costs. The second has seen development of AI for image recognition to automate quantity takeoffs from drawings, PDFs and BIM, and scanning drawings to check for missing information and risk assessment.  

 

Project and document management – As well as AI assistants for PM and construction management there are AI systems for workforce and equipment management, claims, code and contract compliance, attendance and access, and risk management. Some PM systems combine project, site and workforce data for planning, scheduling, and task assignments. 

 

Site monitoring and safety - Safety systems monitor sites and people using sensors, cameras and wearables like badges and detect hazards. Reality capture and computer vision matches site work to BIM models to track progress and defects, and provides access control, headcounts, timesheets and other analytics. 

 

Predictive maintenance of construction equipment - Called telematics, these systems integrate sensors and wireless to collect and record equipment use and performance, and use AI for analysis and predictive maintenance. Typically available as a subscription service, manufacturers now install them on most new equipment and they can be retrofitted to older machines. 

 

Design and planning - One of the first applications of construction AI was in design, which was already software based. Generative design systems create multiple design options according to specific criteria, by providing alternative solutions using a given set of parameters. Used to optimise a design, examples are crane positioning in site plans, site and floor plan layouts for office and apartment buildings.  

 

Materials – there are AI applications for concrete quality control, and some country-based ones for selection of materials by designers. 

 

It is not possible to get a good grasp on the cost of most of these systems. All are subscription based, usually but not always on the number of users, but many do not have prices on their websites. Where available, advertised monthly payments range from tens to hundreds of dollars, and subscriptions often have different options. All this makes assessing the value of different offerings difficult. 

 

 

The Four Cs of AI: Cost, Capability, Competence and Competition

 

At this stage, for the majority of small firms in construction, AI is irrelevant for three reasons. First is the cost, which no matter how low will be too high compared to any efficiency or productivity gains or benefits they might get. That is because, second, the work they do is so small in scale and simple that AI will not improve their capability to do that work. And third, these firms typically do not have people with the skills and technological competence to use AI.

 

At the other extreme are the largest global construction firms. They can pay for AI systems and employ people with the necessary skills, and some have several years of experience with AI systems they have developed internally (e.g. Accionia, AECOM, Balfour Beatty, Bechtel, Skanska). For these firms, using chatbots for document search, image recognition for design and estimating, or vision systems for site monitoring and progress tracking could deliver significant efficiency and productivity gains, improving capability. If AI can increase margins or winning tenders for the megaprojects the firms compete on, their AI use and expertise will improve rapidly. Also, on large projects the cost of using one or more AI systems is not prohibitive, allowing learning by doing and trials of systems on a project by project basis. 

 

For large national or regional contractors with a portfolio of projects, the equation is different. Although project management platforms offered by providers like Aconex, Autodesk and Procore are not cheap, these firms can afford them and most will probably already be using one or more of them, or be using other systems for design, estimating and progress tracking. Although they are unlikely to be developing AI applications internally, they have well established capex programs and IT capability. Importantly, there will be relationships with the major software providers who are incorporating AI into their products that can be built on. These firms have the scale to adopt AI and it should improve their capability, and they will be conscious of rival firms that could gain a competitive advantage through faster adoption of AI. However, they may struggle with competence because there is fierce competition for people with AI skills, and they may be reluctant to pay a premium for those skills. 

 

There are some large US contractors that have technology strategies. Suffolk Constructions has Suffolk Technology, a subsidiary that acts as a venture capital arm and has invested in several construction AI startups [4]. DPR Construction has an internal task force working on AI and uses ML on their two decades of project data. 

 

For medium size construction firms, AI is a more difficult proposition. Margins are thin and software expensive. They are often not using BIM or a PM platform. Employees are busy and focused on day-to-day project management. Trialling AI systems and learning by doing on their projects is challenging, and evaluating the number and diversity of offerings makes getting to a decision and committing hard. While there may be capability benefits in the medium-term, there are present competence and cost limits, and these constraints also apply to their competitors. Early adopters could gain a significant competitive advantage if they have the capital to invest in AI skills and systems. 

 

Automated takeoff and estimating systems should save significant time and effort in preparing bids for small and medium size firms. If these systems reduce cost and improve accuracy in bidding for work, they could significantly increase the capability of contractors and subcontractors that can both afford them and use them effectively. 

For larger specialised subcontractors like HVAC and electrical, there are AI systems specific to their work available. 

 

 

Point versus Platform

 

For all firms the difficult decision is what system or systems to invest in, because once committed there are significant sunk costs and switching can be painful and expensive. There is a basic choice between specific, point solutions that only do one or a few things, and a platform that does many things. The advantage of a platform is a unified solution, with a wide range of functions in one place. The problem with platforms is they are expensive, often complex, and lack flexibility. Point solutions can offer better value by addressing specific functions and doing them very well. However, for contractors where the number of point solutions used is increasing and issues of compatibility and complexity arise, a platform becomes an option.

 

Because platforms are large, have a range of functions, and can be integrated with other enterprise-scale software like accounting, resource management and customer management systems, they are also expensive. Major players like Autodesk, Bentley, Oracle/Aconex, Procore and Trimble have AI assisted systems that usually cover preconstruction, design, estimating and project management, and integrate with some other systems. These systems are platforms for large corporate users. Some of the new construction AI companies have developed platforms, examples are Buildpass and Smartbuild. 

 

 

Conclusion

 

It is a well-known characteristic of the construction industry that many innovations come from suppliers, from outside the industry. In the past this particularly applied to materials and manufacturers of equipment and components, but more recently it has included IT and software suppliers. Many of those suppliers are incorporating AI into their products, and have added AI assistants and analytics to their systems. There are also many new entrants with AI applications, some with platforms but the majority targeting a specific function. 

 

For some of these companies, their extent of use of AI is not obvious from their websites and in some cases may be more like advanced analytics and modelling, which could be sufficient for predictive maintenance for example. However, there are also systems based on chat bots and AI assistants that are clearly closer to the leading edge of AI. It is not possible to get a good grasp on the cost of these systems, many do not have prices on their websites, which makes assessing the value of different offerings difficult. 

 

Some of the large global contractors have developed AI systems internally. Examples are Accionia with their Brion system for water management, AECOM, Balfour Beaty, Bechtel, Skanska, and Larsen & Toubro, but there will be others that have not gone public. Companies that are successfully applying AI on their projects, aren’t going to tell their competitors. 

 

National or regional contractors are unlikely to be developing AI applications internally, but have relationships with the major software providers who are incorporating AI into their products. These firms have the scale to adopt AI and it should improve their capability, and they will be conscious of rival firms that could gain a competitive advantage through faster adoption of AI. However, they may struggle with competence because there is competition for people with AI skills. 

 

There are repetitive tasks that AI can automate, like takeoffs for estimates, code compliance checks with image recognition, or using a text reader for document management. Progress tracking could deliver savings on subcontractor supervision and payments, and reduce reporting cycles. Reducing costs with AI in one task is an advantage, because competitors are similar firms with the same costs. Firms will have to start working with AI somewhere, sooner rather than later, if they are to survive. 

 

Chat GPT was launched in November 2022 by OpenAI, and GPT-4 in March 2023.  In two years, AI has gone from being unreliable and error prone to a technology that allows firms to combine and analyse complex and diverse data. It can synthesise, summarise and interpret data, and provide insights and suggestions, however it does not and cannot replace expertise because it also requires supervision and checking of results. 

 

Construction companies looking at AI have to weigh up four considerations. First is the cost of subscriptions, and whether they have to workload to pay for a platform rather than one or more point solutions. Second is how much difference AI can make to their capability to win, organise and deliver projects. Third is competence in the skills required, and whether their people have or can be trained in these. Fourth is what their competitors are doing, and whether AI can provide a competitive advantage. Cost, capability, competence and competition are the four Cs of construction AI.

 

 

 

[1] Progress in AI may be slowing down. A recent report in The Information said OpenAI’s new models have only incremental improvement over previous ones. Reuters quoted Ilya Sutskever, co-founder of OpenAI, saying results from scaling up pre-training have plateaued, with delays and disappointing outcomes in the race to release a large language model that outperforms GPT-4, which is nearly two years old. 

 

[2] According to Wikipedia a chief experience officer is responsible for user experience of an organisation's products and services, including software and hardware design management. 

 

[3] Many of the startups came from Last Week in ConTech, an invaluable source of information on leading edge technology for the industry. 

 

[4] This is a particularly good interview with Suffolk Construction CTO Jit Kee Chin.