AI in quantity surveying: what it can and cannot do
AI will not replace the quantity surveyor, but it is already changing which parts of the job take the time. To use it well you need a working picture of what a language model is doing when it answers a cost question, where that goes wrong, and what a trustworthy tool does differently.
By the Caprix AI team · · 8 min read
Key takeaways
- A language model predicts plausible text. On its own it does not look anything up, so a rate it quotes from memory is a guess that reads like a fact.
- The fix is grounding: the number comes from a database or a document, and the model's job is to find it, explain it and show its source.
- AI is strongest at language tasks: wording items, structuring bills, reading documents, explaining differences.
- Measurement, professional judgement and accountability stay with the surveyor.
- Judge any AI cost tool by one question: where did this number come from?
What a language model is doing
A large language model is trained on a very large body of text to predict what comes next. Ask it a question and it produces the continuation that best fits the patterns it has learned. That is why it writes fluent specifications and clear explanations: those are language problems, and language is what it models.
It also explains the main weakness. A model asked for the price of a material has no price list to consult. It produces a figure that looks like the figures it has seen. The answer can be out of date, from another country, in the wrong unit, or simply invented, and it will be delivered in the same confident tone as a correct one. This is usually called hallucination.
Two more limits matter for cost work. A model's knowledge stops at its training date, so it knows nothing of last month's prices. And it is unreliable at arithmetic over long tables unless it hands the calculation to a tool.
Grounding: where a trustworthy number comes from
The way round this is to stop asking the model to remember. In a grounded system the model is connected to a source of truth, such as a rate database, an index series or the document you uploaded. When you ask a question, the system retrieves the relevant records and the model works from those.
The division of labour changes. The data supplies the number. The model supplies the understanding of your question, the choice of what to retrieve, and a readable answer. A well built tool goes one step further and shows its working: the period, the location, the unit and the source, so that you can check it.
| Ungrounded answer | Grounded answer | |
|---|---|---|
| Source of the number | The model's memory of its training text | A database or document, retrieved at the time |
| Currency | Fixed at the training date | As current as the data |
| Can you check it? | No | Yes, the period, place and source are stated |
| Typical failure | A confident, wrong figure | No data found, and it says so |
Where AI helps today
Wording and structure
Drafting item descriptions to a standard method, organising work into trades, and spotting items a bill has left out.
Reading documents
Pulling the items out of a PDF bill, or finding the clause in a contract that a question turns on.
Checking
A second pair of eyes for missing units, unit mismatches, duplicated items and rates that look out of line.
Explaining
Turning a table of differences between two tenders into a paragraph a client can follow.
Searching data
Letting you ask for a rate in plain words instead of working through filters.
Scenario thinking
Setting out low, base and high cases with the assumptions behind each.
Where it does not
- Measurement. Taking quantities off drawings to the rules of a standard method is still specialist work. Treat any tool that claims to do it unaided with care, and check its output as you would a junior's.
- Judgement. Whether a rate is right for this site, this contractor and this programme is a professional call.
- Accountability. The surveyor signs the estimate. An AI output is a draft until someone qualified has checked it.
- Things it was never told. A model does not know the ground conditions, the client's priorities or the state of the local market unless that information is in front of it.
A checklist for judging an AI cost tool
- Where do the numbers come from? A named dataset, or the model's memory?
- Does every figure carry its period, location and unit?
- What happens when there is no data? A good tool says so. A poor one fills the gap.
- Can you see and edit what will run before you are charged or committed?
- Does it separate observed data from assumptions, especially in forecasts?
- What happens to documents you upload? How long are they kept, and are they used for training?
- Is the calculation done by code or by the model? Totals should never be left to prediction.
How to work with it
Be specific. A good cost question names the item, the specification, the state, the period and the unit. Vague questions get average answers.
Verify in proportion to the risk. A rate going into a feasibility note needs a sense check. A rate going into a tender needs its source read.
Keep the record. Note what was asked, what came back and what you changed. That is your audit trail, and it is what makes an AI assisted estimate defensible.
How Caprix AI approaches this
Caprix is built on the grounded pattern. Rates come from a Malaysian material, labour and machinery database, and each answer states its period, state and unit. A question opens a form that shows what Caprix understood, and nothing runs or is charged until you confirm it. Forecasts keep observed prices, outside evidence and assumptions apart. In the BOQ Builder, Draft writes the structure and wording of a bill and leaves quantities and rates blank on purpose, because measuring and pricing are yours to decide.