Can AI read construction plans?
Some of a plan, reliably. Not the way you do.

Here is the answer without the marketing. Software can reliably read specific, bounded things off a construction plan today. It can measure the drawn geometry with deterministic computer vision, one click fills a room and measures it, the same way every time. It can read the dimensions the architect printed and trust them over its own inference. It can extract the title block, the drawing number, title and revision, and file the sheet accordingly. And in VIABUILD there is one genuine model-vision capability on plans shipping today, Oryn reads the council approval stamp off a drawing, the DA, CDC or BA number, the approval date and the issuing council, even when the stamp is a rasterised image sitting on an otherwise digital sheet.

What software cannot do, ours included, is comprehend a full drawing set the way a person does. There is no shipped, dependable capability anywhere in this market that reads a set of residential plans, understands the design intent, connects the notes to the details to the schedules, and can be trusted with the result. We do not market plan understanding, because it is not a product, it is a demonstration. The line between those two paragraphs is exactly where a builder's scepticism should sit, and the wider context for it lives in the AI in construction hub.

Three capabilities are routinely implied and should not be bought. Whole-plan comprehension by a language model, a system that reads the set and tells you what the job is, is not shipped anywhere in a form a builder should rely on, and it is not in VIABUILD. Symbol counting from an example, click one downlight and it counts the rest, defeats current tools on real drawing sets far too often to trust, we do not claim it. And automated compliance checking, software reading plans against the NCC or BASIX requirements and passing them, is beyond what any reading technology should be trusted with, compliance requirements change and differ by state, so verify with your certifier and the relevant authority, never with a software reading, including ours.

The pattern behind all three is the same. Reading a bounded fact off a sheet can be verified in seconds, so it can be trusted in a workflow. Comprehension cannot be verified without doing the comprehension yourself, so it cannot be delegated safely at any price. This is the deterministic-first argument made throughout the document intelligence reference, applied to the hardest document class a building business owns. Drawings are also the document class where being subtly out of date costs the most, which is why the discipline in drawing revision control matters more than any extraction feature built on top of it.

Everything VIABUILD reads off a plan follows the same trust rules as everything else Oryn touches. An extracted value, a stamp number, a revision, a title, lands amber, a candidate carrying the document and page it came from, and becomes part of the job only when a person commits it to green. A suggested revision match is a suggestion, the superseded sheet is never silently overwritten. Measured quantities are visible as shapes on the drawing, so checking them is a glance, the takeoff workflow around that is on the AI takeoff software page.

Review is not a disclaimer here, it is the design. On a residential job, a misread plan becomes a wrong quantity, a wrong quantity becomes a wrong price or a wrong order, and the error surfaces months later as a variation argument or a supplier dispute. Software earns its place by making the human check fast and the evidence one click away, not by promising the check is unnecessary. The same standard applies to the money side of the question, whether a model can price the job it read, and that answer is at can AI estimate building costs.