Can AI produce an early cost view a quantity surveyor would stand behind?
An early estimate decides whether a project goes ahead, and it is made when almost nothing is settled. AI can produce a number quickly, but a number is only useful if a professional will put their name to it. This note is about how AI can speed the estimate up while the quantity surveyor still owns it. It is grounded in New Zealand cost data and in the published record on how estimates actually behave.
The first cost view is a judgement, and it sets the budget.
Before much is drawn, a quantity surveyor turns a sketch and a floor area into a number. Its purpose is plain: to establish whether the project is affordable and to set a realistic cost limit that then disciplines the design.1 It is also where a project is quietly won or sunk, because a figure set too low commits everyone to something that cannot be built for the money, in a market where build costs have risen far faster than general inflation over the past decade.19
Whether the project proceeds at all, on the figure put in front of the client.
The budget the design then has to be kept inside as it develops.
What the client can realistically ask for on this budget, and what has to give.
What must be raised or borrowed, and the basis on which lenders and boards commit.
An early estimate is wide by nature, and pretending otherwise is the mistake.
The uncertainty is not a sign of a poor estimator. It is structural. With the design only a few percent defined, the international cost-classification standard puts the realistic accuracy of a concept estimate as wide as minus 50 to plus 100 percent.2 The historical record agrees: across a large sample of public projects, costs were underestimated in almost nine out of ten, by an average of about 28 percent, and a later database of more than 16,000 projects found over 90 percent ran over budget, over schedule, or both.34
The realistic accuracy band of a concept estimate, at 0 to 2 percent design definition, under the AACE classification.
AACE 18R‑97Public infrastructure projects whose costs were underestimated against the decision-to-build figure.
Flyvbjerg, JAPA · 2002Average cost overrun across 258 infrastructure projects, in constant prices, with a wide spread around it.
Flyvbjerg, JAPA · 2002In New Zealand the cost of an early number set too low is on the public record.
These are not abstract risks. The country's recent major projects show what happens when the early figure is optimistic: the funder tops up, the scope is cut, or the project is cancelled. The Auditor-General found several were announced with limited or no business case, after officials had warned of a real risk of cost overruns.5
| Project | Early estimate | Where it landed | Outcome |
|---|---|---|---|
| City Rail Link, Auckland | $3.4b · 2014 | $5.49b · 2023 | +62% |
| Transmission Gully | ~$850m · 2012 | $1.25b · 2022 | +47% |
| New Dunedin Hospital | $1.4b · 2018 | ~$2.05b · 2025 | +47% |
| NZ Upgrade, transport | $6.8b · 2020 | +$1.9b top-up · 2021 | Over |
| Ellesmere College rebuild | ~$30m | $60m | Doubled |
| America's Cup infrastructure | ~$254m approved | $238m spent · 2021 | Under |
The pattern is consistent. An early underestimate is resolved later by more money, less building, or no building. Across 111 major school projects, about one in four came in at least a third over budget. The one that held, the America's Cup infrastructure, is shown for balance.678910
AI can take the grind out of the estimate, not the judgement.
The useful question is not whether AI can produce a number, it plainly can, but which parts of the work it genuinely improves. The honest answer is the mechanical parts: measuring quantities off drawings, pricing them against current data, benchmarking against comparable projects, and keeping the figure live as the design moves.1420
Quantity takeoff
Measuring areas and counts off drawings, where tools report large time savings against doing it by hand.
Live pricing
Pricing those quantities against a current cost library, and updating as rates and the design change.
Benchmarking
Comparing the figure against the outturn of similar past projects, which is the basis of a sound early estimate.
The range, not a point
Presenting the estimate as a band with its assumptions, instead of a single number that reads as settled.
The limit is the input, not the model.
At concept stage the design is two percent defined, and good structured cost history is scarce, so the number stays wide whoever produces it. AI changes the speed and the honesty of the estimate, not the physics of how little is known.2
Two ways to get a number, and only one of them holds up.
As with any AI output, there are two ways to produce the estimate, and the evidence is clear about where each fails. One is to let a general model reason its way to a figure. The other is to keep the rates and the arithmetic in structured tools and use AI to orchestrate them. Large models are documented to be unreliable at exact calculation: their own builders route arithmetic to a calculator, and one irrelevant added sentence has been shown to cut a model's maths accuracy by up to 65 percent.1716
Drop in a model's mathematical accuracy from adding a single irrelevant sentence to the problem.
Apple, GSM‑Symbolic · 2024Typical error of academic machine-learning models predicting conceptual building cost, the honest state of the art.
Peer-reviewed · 2002–2025Vendor takeoff "accuracy", which measures geometry off a drawing, not the cost of the building.
Vendor claimA general model estimates
Ask a capable model to reason over the project and the open web and arrive at a cost figure.
- Drafting and explaining the estimate in plain words
- Surfacing comparable projects qualitatively
- Speed, and flexing to an unusual brief
- A readable first pass with nothing to build
- Exact arithmetic, which it gets wrong as numbers grow
- Current local rates, which it does not hold
- Provenance: which rate came from where, and when
- Returning the same figure twice
Structured cost data, AI on top
Keep rates and sums in cost tools and your own project history, and use AI to orchestrate, draft and explain.
- Current, authoritative rates with a source attached
- Deterministic, checkable arithmetic
- Benchmarking against real project outturns
- An auditable trail behind every figure
- Only as good as the cost data behind it
- Build and upkeep of the data and tools
- The genuinely novel project with no comparable
- Questions outside the rates it holds
The strongest result is not one approach or the other. It is structured tools carrying the rates and the arithmetic, with AI doing the drafting and the surveyor doing the judging.
The arithmetic
Rates and sums sit in structured tools, so every figure is current, sourced and reproducible.
The judgement
AI drafts, benchmarks and surfaces the range and the assumptions for the surveyor to weigh.
The signature
The quantity surveyor sets the uplift and owns the number, now reached faster and shown honestly.
A precise-looking number is the most dangerous output of all.
The failure mode here is not a wrong figure, it is a confident one. An AI estimate returned to the dollar reads as certain when it is a concept-stage figure with a wide band around it. The discipline that keeps it honest is well established: show the range, carry every rate's source and date, name the assumptions, and add an explicit allowance for optimism, which the evidence says is endemic.3
A figure like $4.24 million reads as settled, invites the client and the board to plan against it, and hides the wide band it actually carries. When the real cost arrives, the gap becomes the project's problem rather than the estimate's, and that is how budgets are quietly broken.
The estimate is a band, every rate shows where it came from and when, the assumptions are named, and an optimism-bias uplift is applied from comparable projects. New Zealand's own guidance for public investment expects exactly this, and the United Kingdom Treasury's evidence-based uplift for a non-standard building at this stage reaches about half the capital cost again.111213
A faster, honest estimate changes where the surveyor spends judgement.
When the takeoff and the pricing are automated and the range is shown plainly, the quantity surveyor's time moves to where it matters: the comparables, the assumptions, and the uplift. The timing suits it. From early 2026 the RICS standard on AI requires that a named, qualified surveyor owns and documents any AI output that affects the advice, which is precisely the posture this argues for.1822
Judgement where it counts
Time moves from takeoff grind to the assumptions and the uplift.
An honest range
The client hears a band and its risks, not a single false-precise figure.
A defensible trail
Every rate and assumption carries its source, so the number can be stood behind.
Faster feasibility
A sound early view forms in a sitting, not over a fortnight of takeoff.
Price from structured data
Keep rates and sums in cost tools, not in a model's head.
Show the range
Lead with a band and its assumptions, never a single number.
Benchmark against real projects
Anchor the figure in comparable outturns, and uplift for optimism.
Keep the surveyor's name on it
Use AI to prepare the estimate, not to own it.
The goal is not an AI that prices the job.
It is a quantity surveyor who can stand behind the early number sooner, and show it honestly. AI does the measuring, the pricing and the benchmarking, and lays the range and the assumptions out in the open. The number still belongs to a person. They just reach it faster, and with the uncertainty shown rather than hidden.
- RICS. NRM 1: Order of cost estimating and cost planning for capital building works. Effective 1 Dec 2021. rics.org
- AACE International. Recommended Practice 18R‑97 / 56R‑08: Cost Estimate Classification System. 2016–2020. aacei.org
- Flyvbjerg, Holm & Buhl. Underestimating Costs in Public Works Projects: Error or Lie? Journal of the American Planning Association 68(3). 2002. arxiv.org
- Flyvbjerg & Gardner. How Big Things Get Done (database of 16,000+ projects). 2023. prh.com
- Office of the Auditor-General. Making infrastructure investment decisions quickly (NZ Upgrade and Shovel-Ready). 2023. oag.parliament.nz
- RNZ. Auckland's City Rail Link cost climbs to $5.49b. 2023. rnz.co.nz
- RNZ. Transmission Gully: original estimate and the Richards review. 2021–2022. rnz.co.nz
- RNZ. New Dunedin Hospital: approved budget higher than government claimed (Treasury QIR). 2026. rnz.co.nz
- RNZ. School building projects face delays and cost blowouts (MoE OIA data). 2023. rnz.co.nz
- NZ Herald. America's Cup infrastructure cost and economic cost-benefit. 2021. nzherald.co.nz
- HM Treasury. Supplementary Green Book Guidance: Optimism Bias (uplift table). gov.uk
- NZ Treasury. Better Business Cases and Guide to Social Cost Benefit Analysis (optimism bias; quantitative risk assessment). 2015–2023. treasury.govt.nz
- Waka Kotahi NZ Transport Agency. Cost Estimation Manual SM014 (consider optimism bias; expected accuracy by method). Eff. May 2025. nzta.govt.nz
- Togal.AI. Automated quantity takeoff: accuracy and time-saving claims (vendor). 2026. togal.ai
- Emsley et al. Data modelling and the application of a neural network approach to the prediction of total construction costs. Construction Management and Economics 20(6). 2002. tandfonline.com
- Mirzadeh et al. (Apple). GSM‑Symbolic: understanding the limitations of mathematical reasoning in LLMs. 2024. arxiv.org
- Cobbe et al. (OpenAI). Training Verifiers to Solve Math Word Problems (GSM8K; calculator injection). 2021. arxiv.org
- RICS. Responsible use of AI in surveying practice (a named surveyor must own AI outputs). Effective 9 Mar 2026. rics.org
- QV (Quotable Value). Construction costs up 61% since 2015 versus CPI 33%. 2025. qv.co.nz
- QV CostBuilder. New Zealand construction cost database (elemental and per-m² rates for preliminary estimating). 2026. costbuilder.qv.co.nz
- Rider Levett Bucknall. Riders Digest New Zealand 2025 (indicative cost rates, excl GST). 2025. rlb.com
- NZIQS. Code of Conduct (members accountable for negligent or incompetent quantity surveying services). 2025. nziqs.co.nz