Why AI in Construction Needs a Builder, Not a Software Vendor

Connect Design Construct | Sydney NSW & Canberra ACT

Every developer and builder in the country is being sold artificial intelligence right now. Platforms that will read your drawings. Tools that will write your scopes. Systems that promise to find your defects, price your variations and manage your programme.

Most of these pilots will quietly die within a year. Not because the technology doesn't work — it works better than most people in this industry realise — but because the technology was never the hard part.

The hard part is knowing what to ask it, and knowing when its answer is wrong.

The thing nobody selling AI wants to say

Modern AI can read a full drawing set faster than any human alive. It can cross-reference a hydraulic layout against a structural set and flag that a drainage penetration at the Level 3 transfer slab appears on one and not the other. It will do that in minutes, across hundreds of sheets, without getting tired at 4pm on a Thursday.

That's real, and it's valuable. But look closely at what it has actually produced: a flag. An observation. A line item that says these two documents disagree.

Now consider the questions that flag raises:

• Does this actually matter, or is it a coordination note that will resolve itself in shop drawings?

• If it matters, who owns it — the hydraulic consultant, the structural engineer, or the head contractor under a design-and-construct scope?

• Does it need core drilling through post-tensioned slab, and has anyone checked the tendon layout?

• Is this a $15,000 fix or a $180,000 fix?

• Does it sit on the critical path, and what does two weeks of delay cost when liquidated damages and prolongation are both in play?

• Will the contractor price this at tender, or leave it out and claim it as a variation in month nine?

Not one of those questions can be answered by a language model. Every one of them can be answered in about ninety seconds by someone who has actually built a building.

The flagging is cheap. The judgment is the entire product.

Why this hits Tier 3 hardest

Building Commission NSW's own briefing to Parliament in 2024 set out the landscape plainly. Tier 1 developers and contractors — the ones building 500 to 3,000 apartments a year in towers over ten storeys — are the most likely to have adopted the post-2020 building reforms. Tier 2, delivering 200 to 500 apartments in six to ten storey developments, largely the same.

Tier 3 is different. These are developers delivering up to 200 apartments a year in three to six storey buildings, and the Commission's assessment of reform adoption in that cohort was blunt: less known. The Commission also noted that three to ten storey buildings will dominate the NSW residential development landscape by 2030, and that the sector needs 50 to 100 new players entering each year to hit housing targets.

So the fastest-growing, most numerous, most housing-critical part of the market is also the part with the thinnest technical infrastructure and the least experience of the compliance regime it now sits inside.

That regime is not forgiving. Under the Design and Building Practitioners Act, every regulated design must be prepared and declared by a registered practitioner and lodged on the NSW Planning Portal before the work it covers is built. Builders must then construct in accordance with those declared designs. A certifier cannot issue an Occupation Certificate until the lodgement obligations are met. Get the register wrong and the consequence isn't a fine — it's a settlement date that moves while your financier watches.

A Tier 1 developer absorbs this with a digital engineering team, a design manager and a compliance coordinator. A Tier 3 developer running four projects with six people cannot. This is precisely why AI represents a bigger relative advantage for Tier 3 than for anyone else in the market — it substitutes for headcount they were never going to hire.

But only if it's implemented by someone who knows what "correct" looks like.

Where AI earns its place, and where it doesn't

There is a clean line through this, and it comes down to whether the task is finding or deciding.

AI is exceptional at finding:

• Reconciling a drawing register against what was actually issued, revision by revision

• Detecting conflicts between architectural, structural, hydraulic, mechanical, electrical and fire documentation

• Comparing specification clauses against what the drawings show, in both directions

• Building a regulated design register and testing it against what has been produced

• Normalising six trade quotes with six different inclusion and exclusion structures into something comparable

• Reading two years of site correspondence and a programme to test whether an extension of time claim holds together

AI should not be deciding:

• What a scope gap will actually cost, without rates calibrated against real recent projects

• Whether a detail is buildable by the trades available in this market at this price

• Whether something complies — compliance is a determination, and determinations carry liability

• Anything that goes into a document with a licensed practitioner's name on it and hasn't been read by that practitioner

The failure mode we see most often is an organisation that has crossed this line without noticing. Output goes into a report. The report goes to a client. Nobody in the chain has verified the reasoning, because the output was fluent and fluency reads like competence. In construction, that's how you end up defending something in NCAT.

The three practical traps

Buying a platform instead of solving a problem. The right sequence is to identify the specific task that costs you money, solve it manually until you understand it completely, and only then automate the part you've done twenty times and resented every time. Organisations that start with procurement end up with licences nobody uses.

Ignoring where the documents go. Uploading a client's confidential drawing set to a consumer AI product is a commercial and professional risk that most people take without thinking about it once. Enterprise arrangements with defined data handling exist for a reason. If you cannot state your data position in a single sentence to a developer who asks, you aren't ready to handle their project.

Assuming standards are available to you. Australian Standards are copyrighted and licensed. AS 3740, AS 4654.2 and their equivalents cannot simply be loaded into a system and reproduced for clients. A great deal of enthusiastic construction AI work has run directly into this wall. It's navigable, but it needs to be designed for at the start, not discovered at the end.

The integration is the skill

None of this is an argument against AI in construction. It's an argument about who should be doing it.

A data scientist who has read about construction will build something impressive that flags four hundred issues, of which eleven matter — and will have no way of knowing which eleven. A builder who has watched a few webinars will either avoid the technology entirely or trust it in exactly the places it shouldn't be trusted.

What the work actually requires is one person, or one small team, holding both: enough construction experience to know that the waterproofing junction detail is the finding that matters and the four hundred drawing revision mismatches are noise, and enough technical literacy to build a workflow that surfaces it reliably and handles the data properly.

That combination is rare, and it's the whole game.

What it's worth

The arithmetic is not subtle. On a $30 million Class 2 project, variation and extension-of-time leakage driven by incomplete or inconsistent documentation typically runs between five and ten per cent of construction value. That is $1.5 million to $3 million, and most of it is visible in the drawings months before anyone breaks ground.

A thorough pre-tender documentation and buildability review is traditionally a three-week exercise, priced accordingly. Built properly — machine reading, human judgment — it becomes a ten-day exercise at well under a tenth of one per cent of construction cost.

The saving isn't the technology. The technology just made it fast enough to be worth doing before tender instead of arguing about after practical completion.

Connect Design Construct holds an unrestricted builders licence and brings over fifteen years and more than $600 million of delivered construction to design assurance, buildability and value management work for developers in Sydney and Canberra.

📧 info@connectdc.com.au | ☎ 02 8484 0313