Can AI Improve Bid Accuracy?
Australian infrastructure estimates have made too little allowance for bad outcomes: their worst cases sit far too close to their medians. Models built on completed projects may help correct that where past jobs are well recorded and alike. They cannot price scope that nobody has designed yet.
On this page
Better: Note and confidence
Low to moderate confidenceTighter on comparable, well-recorded work; untested on immature scopeIn one peer-reviewed UK study, read in abstract, 92% of 100 validation predictions from neural network models fell within 10% of final cost (measured). The authors' earlier conference paper shows a hand-selected set of similar water projects with the target cost as an input. We found no published test of such models against Australian tender prices or final costs.Faster: Note and confidence
Low confidenceQuick to apply, slow to buildIn our reading, once a reference class exists, applying its uplift is arithmetic. Building one is not; in a Hong Kong roadworks study, much of the effort went into retrieving and verifying project data. We found no study measuring estimating time with machine learning.Cheaper: Note and confidence
Low confidenceBetter estimates allocate money better; savings are unmeasuredGrattan judged Australian worst-case allowances likely too low, while early Hong Kong estimates with contingency showed pessimism bias, which the authors link to inefficient allocation of funds. In our reading both errors cost money, but we found no study measuring savings from model-based estimating.Safer: Note and confidence
Governance-gatedA model's number still needs a signatureAn estimate that sets a bid, a budget or a contingency carries commercial and public consequences. Which percentile is used, which past projects were excluded and who signs the estimate are decisions for people, and they should be recorded.The problem, in numbers
A bid or a budget is a forecast, and in Australian public infrastructure the worst outcomes have run well beyond what estimates allowed for. Inland Rail's estimated cost was $16.4 billion in 2020; two years later the operator's latest estimate was $31.4 billion. The Commonwealth's independent reviewer, Kerry Schott, called that near doubling astonishing and traced it mainly to scope growth caused by immature preliminary designs and approval requirements, to delays from prolonged approvals and to recent escalation.Source 3 About 90% of the project's scope still had low or moderate cost certainty, by the operator's own account.Source 3
The pattern is wider than one project. The Grattan Institute found that recent Australian business cases put the worst-case (P90) cost about 7% above the median (P50) estimate; on the experience of projects completed over the past two decades, the observed difference averaged 49%.Source 2 Inland Rail's 2015 estimates, as Grattan tabulates them, had put the worst case 7.8% above the median.Source 2 Grattan's reading is that either median estimates are generally too high or, more likely, worst-case estimates are generally too low.Source 2
Infrastructure Australia notes that its pipeline forecast uses cost estimates with limited certainty about future escalation, and in a like-for-like comparison across one year, changes in investment estimates added 10% to the five-year pipeline.Source 1 The Australian Construction Industry Forum projects that work done could contract by 0.8% in 2026, with margins squeezed in an industry that already tops the insolvency count.Source 8
What improves, and why
Conventional estimates take an inside view, building up from the project itself. Reference class forecasting, which Bent Flyvbjerg applied to infrastructure, adds an outside view. It finds completed projects like this one, measures how far their final costs strayed from their estimates, and applies that spread as an uplift chosen for the risk the owner will accept.Source 4 In uplifts developed for the UK Department for Transport, a reference class of 172 completed road projects implied a 15% uplift for a 50% risk of overrun and 45% for a 10% risk.Source 4
In our reading, machine learning applies the same idea at a finer grain. Instead of one uplift per project type, a model learns how features such as scope, purpose, region, duration and target cost relate to final cost.Source 7
A model trained on finished projects can tell an estimator how wrong estimates like theirs have been. It cannot price scope that nobody has drawn yet.
What the evidence actually says
The strongest machine-learning result we found comes from UK water infrastructure. Dominic Ahiaga-Dagbui and Simon Smith built neural network models on 1,600 completed projects; in the journal abstract, 92% of 100 validation predictions fell within 10% of actual final cost and 77% within 5%.Source 6 Their earlier conference paper, which we take to use the same data, shows the conditions. The 1,600 were one cluster hand-picked from almost 5,000 projects in a UK water industry partner's database, chosen to be similar and free of significant missing data or extreme values, and the inputs included each project's target cost. Its best model was validated on 15 projects.Source 7 In our reading, that is a model correcting an existing estimate on clean, comparable work: a best case, not a typical tender.
Reference class forecasting has been tested more directly against outcomes. In a feasibility study commissioned by Hong Kong's Development Bureau, data from 25 completed roadworks projects were used to build uplift curves by planning stage, after excluding projects started before 1993. At the earliest stage, an 80% chance of staying within budget needed a 44% cost uplift; once projects were ready to tender, it needed 14%, and the P50 uplift was minus 1%, which the authors read as the median estimate tending to match final cost.Source 5 When each project with an earliest-stage estimate was left out in turn and the curve rebuilt from the rest, the P80 uplift would have covered 14 of 18 projects.Source 5
Hong Kong roadworks projects by planning stage: Category C (accepted in principle), Category B (resources earmarked) and Category A (ready to tender). Uplift on base cost, excluding inflation and bottom-up contingencies. Bar widths are our calculation, relative to Category C.
The study cuts the other way too. Early estimates that included contingency showed pessimism bias, and half the projects underspent the budget set at the earliest stage.Source 5
Grattan links the Australian gap to missing data on completed projects. The federal infrastructure department has pointed to estimators lacking the historical cost databases available to US and European counterparts, and the only progress Grattan found was a federal benchmarking pilot covering 45 road projects in 2015 and 32 in 2017, which was not repeated in 2019 as recommended.Source 2 We found no published study testing machine-learning estimates against Australian tender prices or final costs.
Where it falls short
Scope it cannot see. The reviewer tied much of Inland Rail's increase to scope growth from immature designs, and found estimates had substance in the few sections where designs were final, approvals made and bids received.Source 3
Clean data is the exception. The best machine-learning result used projects selected to exclude incomplete records and outliers, and its authors name a poor culture of data warehousing in construction as a major challenge.Source 7
Small reference classes. Hong Kong's study rested on 25 projects (its validation on 18), a sample size its authors chose as a trade-off between statistical validity and the effort of collecting archival data.Source 5
Uneven readiness. In Infrastructure Australia's 2025 survey of 200 firms, 64% said they had invested in data and digitalisation, across BIM, AI, robotics and automation (self-reported; AI is not broken out), but 26% made no digital investment, and small businesses lagged on AI awareness.Source 1
Where governance decides the outcome
A model-based estimate should be traceable: which reference class or training set was used, which projects were excluded, which percentile was chosen and who signed it off. Choosing P50 or P90 is a decision about risk appetite for the client or board, not a setting inside the model. The Hong Kong study proposes holding contingency in tiers, at contract, project and portfolio level, so that a realistic uplift does not simply become a larger budget to spend.Source 5 Grattan recommends that states require completed project data, from first announced cost to final outturn, for every project of $20 million or more.Source 2 A model-assisted bid is still the contractor's offer, and its sign-off should stay with a named estimator.
How to think about adoption
Start with your own closed jobs: the estimate at each stage, the tender price, the final cost and what changed in scope. Without that record, no model has anything to learn from. Test any model on past projects it has not seen, and report the share of predictions within 5% and 10% of final cost, as the UK study does. In our reading, repeatable work with stable scope is the place to start. State early estimates as ranges with an explicit percentile, and treat vendor accuracy claims as hypotheses until they are tested on your own completed projects.
BFCS.ai does not fabricate figures. The machine-learning accuracy figures come from one UK industry partner's selected projects, read from a journal abstract and an earlier conference paper. Reference class results come from UK transport and Hong Kong roadworks, read in the authors' arXiv manuscripts. We found no Australian study testing machine-learning estimates against tender prices or final costs, and no vendor data is used here.
Sources
- 1
Infrastructure Australia, "Infrastructure Market Capacity 2025 Report," November 2025. infrastructureaustralia.gov.au
Back to textWhat we took from it
Page 19: its forecast "uses cost estimates with limited certainty about future escalations", unlike ABS data, which reflect actual costs. Page 23: in a like-for-like comparison of the five-year Major Public Infrastructure Pipeline ($189 billion of the 2024 forecast), changes in investment estimates added 10%, broadly in line with prior years. Page 6: growth in key materials costs such as cement and timber has eased to around inflation levels. Pages 14, 17 and 63: in the 2025 Industry Confidence Survey (n=200 Australian building and construction businesses, self-reported; page 17 gives the base), 64% had invested in data and digitalisation in the past year, with investments across BIM, AI, robotics and automation; 26% made no investment in digitalisation, with small business lagging in adoption and awareness of AI tools. AI is not broken out separately.
- 2
Terrill, M., Emslie, O. & Moran, G., "The rise of megaprojects: counting the costs," Grattan Institute Report No. 2020-15, November 2020, section 6.2 (pp. 37 to 41). grattan.edu.au
Back to textWhat we took from it
In recent business cases the difference between P50 and P90 cost estimates is generally about 7% (Table 6.1 average 7.3%, including Inland Rail at $9,889 million median and $10,657 million worst case, 7.8%, sourced to ARTC 2015); across the experience of the past two decades (Table 6.1: projects completed in the past 19 years) the difference is 49% on average, based on final cost compared with the estimate when first committed (ten per cent of projects overran by more than 49% while the median overrun was zero; Grattan analysis of Deloitte Access Economics Investment Monitor). Grattan concludes that either median estimates are generally too high or, more likely, worst-case estimates are generally too low. The federal infrastructure department has pointed to estimators "not having access to historical cost databases" available to US and European counterparts; a BITRE benchmarking pilot covered 45 road projects in 2015 and 32 in a 2017 update, and its recommended 2019 repeat did not happen. Recommendation 6: states should require completed project data, including first announced cost, contracted cost, cost estimate at the start of construction and final outturn cost, for any infrastructure project valued at $20 million or more, to be provided by each state's infrastructure minister to its infrastructure advisory body.
- 3
Schott, K., "The Delivery of Inland Rail: An Independent Review," Commonwealth of Australia, January 2023 (released by the Department of Infrastructure, Transport, Regional Development, Communications and the Arts). infrastructure.gov.au
Back to textWhat we took from it
The estimated total cost in 2020 was $16.4 billion; the latest ARTC estimate is $31.4 billion, "an astonishing estimated cost increase of almost double in just two years", mainly from scope increases caused by immature preliminary designs and approval requirements, prolonged approvals and recent escalation (Table 6.3: base scope growth $3.5 billion, new scope $3.0 billion, prolongation $2.5 billion, escalation $2.0 billion, among others). About 90% of the scope has low or moderate cost certainty. In the few sections where designs are finalised, approvals made and bids received, the scope is stable and estimates have substance; this is not the case for most of the route. Recommendation 18 calls for the Commonwealth to appoint an independent value engineer/cost estimator; the review also says the estimate's structure should be better aligned with standard P50 and P90 practice.
- 4
Flyvbjerg, B., "From Nobel Prize to Project Management: Getting Risks Right," Project Management Journal 37(3), August 2006, pp. 5-15 (peer-reviewed), read in the author's manuscript posted on arXiv, not the publisher's typeset version. arxiv.org
Back to textWhat we took from it
For transport infrastructure, accuracy in cost forecasts has not improved over the 70-year period for which cost data are available. Reference class forecasting takes an outside view: a reference class of comparable completed projects, its distribution of cost overruns, and an uplift set by the acceptable risk of overrun. In the uplifts developed with COWI for the UK Department for Transport, a reference class of 172 completed and comparable road projects gives a required uplift of 15% at a 50% acceptable risk of overrun and 45% at 10%.
- 5
Flyvbjerg, B., Hon, C. K. & Fok, W. H., "Reference Class Forecasting for Hong Kong's Major Roadworks Projects," Proceedings of the Institution of Civil Engineers: doi.org
read in the authors' manuscript posted on arXiv, which gives the full reference to the published version; the publisher's typeset version was not read. arxiv.org
Back to textWhat we took from it
Civil Engineering 169(CE6), November 2016, pp. 17-24,
Feasibility study commissioned by Hong Kong's Development Bureau in September 2012; data on the 25 most recently completed Highways Department roadworks projects over HKD 100 million, constructed 1993 to 2011, a sample size chosen as a trade-off between statistical validity and the effort of collecting archival data. Projects started before 1993 were excluded from the reference class. Base cost uplifts (excluding inflation and bottom-up contingencies), Table 3: P80 +44% at Category C (accepted in principle), +34% at Category B (resources earmarked), +14% at Category A (ready for tender with an approved estimate); P50 -1% at Category A, which the authors read as the median estimate tending to match final cost. Leave-one-out validation on Category C estimates for 18 projects: the P50 uplift would have prevented overrun in 9 and the P80 uplift in 14. Full estimates including contingency showed pessimism bias at Categories C and B, and half the projects underspent their Category C budget. Much of the effort went into retrieving and verifying project data. The paper proposes tiered contingencies held at contract, project and portfolio level.
- 6
Ahiaga-Dagbui, D. D. & Smith, S. D., "Dealing with construction cost overruns using data mining," Construction Management and Economics 32(7-8), 2014, pp. 682-694 (peer-reviewed; abstract only, read on RePEc). doi.org
Back to textWhat we took from it
Bootstrapping and ensemble artificial neural networks built on 1,600 completed projects to forecast final cost; 92% of 100 validation predictions were within ±10% of actual final cost and 77% within ±5%; the models are being deployed by the industry partner to improve initial estimates. The abstract does not name the sector or describe how the projects were selected.
- 7
Ahiaga-Dagbui, D. D. & Smith, S. D., "My cost runneth over: data mining to reduce construction cost overruns," Proceedings of the 29th Annual ARCOM Conference, Reading, 2-4 September 2013, pp. 559-568 (conference paper, the earlier stage of the work in footnote 6; ARCOM papers are refereed by its scientific committee, a practice we confirmed for other years but not for 2013 specifically; the two authors also edited the 2013 proceedings). arcom.ac.uk
Back to textWhat we took from it
Data from a UK water infrastructure industry partner: a database of almost 5,000 projects completed 2000 to 2012, from which cluster analysis and purposive sampling selected about 1,600 completed 2004 to 2012 that were similar and without significant missing data or extreme values (the authors describe the cluster as representative of the dataset). The initial predictors listed were delivery partner, purpose, scope, target cost, duration and region. Best model: average absolute percentage error 3.67% across 15 validation cases, 87% within 5%. The authors identify a poor culture of data warehousing in construction as a major challenge.
- 8
Australian Construction Industry Forum, "May 2026 Forecasts released: ACIF Forecasts Industry could be bought to Standstill," 3 June 2026. acif.com.au
Back to textWhat we took from it
The May 2026 forecasts project that growth could be brought to a standstill after 3.6% growth over 2025, which could lead to a 0.8% contraction in work done in 2026 (projection); ACIF's chief forecaster says exposure to interest rate changes, which affect demand and costs, compresses margins in an industry that already tops the count for insolvencies.