AI in construction estimating: what is actually working, what is failing, and what is overhyped

Every two years a new wave of construction technology promises to revolutionise estimating. BIM did, partially. Cloud-based job management did, partially. AI is in its second wave now, and the gap between the marketing and the reality is wider than usual. Here is what is actually moving the needle on a real Australian remediation crew, and what is still a demo.

01What is working: assistive drafting

AI is genuinely useful as a drafting assistant for scope of works. Feed it three site photos, a 90-second voice note, and a building manager's emailed brief, and it will produce a 30-line scope draft. Materials, trades, sequence, references. The estimator then edits, prices, and signs off.

The save is roughly ninety minutes per quote. It is not magic. It is the AI doing the part of estimating that is mechanical: transcribing notes, structuring line items, naming trades the photos suggest, attaching the standard references. This is the part that bores estimators and where they make tired-Friday-night mistakes. It is also the only part the AI is reliably good at.

02What is failing: full autonomy

An AI that sends quotes without a human in the loop will, sooner or later, hallucinate a trade, miss a compliance clause, or under-quote an access scenario. The model does not know that the building in Glebe has no lift, or that the strata manager is litigious, or that the client's last builder went bankrupt mid-job. The judgement that handles those is not in the model.

Vendors who pitch full autonomy are pitching to investors, not to builders. The crews we know who have tried it have all walked it back to assistive drafting within three months.

03What is failing: generic models

A general-purpose AI does not understand AS 3740, NCC clauses, or what a strata panel expects to see in a defect rectification scope. It will produce text that looks plausible to someone outside the industry and falls apart to anyone inside it.

Fine-tuning on real Australian scopes is the difference between a toy and a tool. The training data has to know that strata committees skim cover pages, that NCC clauses matter, that the difference between "remedial" and "rectification" carries legal weight. Generic foundation models do not learn that from internet text.

04What is overhyped: photo-to-quote in 30 seconds

The demo videos show a builder snapping a photo and watching a polished quote materialise. The reality is that any photo-to-quote pipeline that works in 30 seconds is producing a quote that needs another three hours of human refinement. The 30-second demo is the easy part. The trustworthy quote is the hard part.

The honest version: photo and voice input gives you a draft in under ten minutes. The draft is good enough to edit, not good enough to send. A senior estimator gets it to send-quality in another twenty minutes. That is still a 5x speed-up. The marketing version oversells, the actual version still delivers.

05What is next: variation-aware estimating

The next genuine frontier is scope variations. Same site, three pricing scenarios: cheap-and-quick, standard, premium. An AI that can produce three internally consistent variations of a scope, with consistent rate logic, is going to change how builders pitch.

Right now this is a manual job that most builders skip. They quote one scenario and let the client decide. The crews that consistently win premium work present three options, anchored so the middle one is the easy "yes". An AI that makes that workflow as fast as quoting a single scenario is the next compounding advantage.