AI in construction · Assistants
A construction AI assistant
should cite its sources.
If you are picturing a chatbot, this page will talk you out of it, politely. The assistant a building business actually needs is grounded recall inside the software that runs the job, answering from your own records, showing its evidence, admitting when nothing is on file, and never changing a record without you.
01 / The direct answer
The assistant you want is not a chatbot
A construction AI assistant, done properly, is a way to ask your job a question and get a truthful answer. When is practical completion due, which invoices are overdue, what is still committed against the frame stage, what variations are waiting on the client. The answers exist in your records, the assistant's job is to retrieve them, cite them, and stop there. In VIABUILD that capability is Ask Oryn, part of Oryn, the structured intelligence layer, and it is deliberately not a chatbot.
The distinction is worth being blunt about. A chatbot generates plausible text, and plausible is the failure mode, an answer about your cash position that sounds right and cites nothing is worse than no answer. Structured intelligence works the other way around, deterministic retrieval from your own records first, a language model only to compose, restricted to what was retrieved, with sources shown. The wider map of which engine belongs to which task is in the AI in construction hub, and if the tool you are weighing up is a general one, the honest comparison is ChatGPT for builders.
02 / The standard
Six things to demand from any assistant
Whatever the vendor, hold the product to these. Each one is testable in a demonstration with your own job data, and each one separates recall from performance.
Answers from your job, not the internet
The question is about your budget, your claims, your overdue invoices. A worthwhile assistant retrieves from your own records and screens, never from general knowledge, so the answer is about the Henderson job, not a hypothetical one.
A source on every answer
Each answer carries source chips to the document and page, or the exact in-app screen it read. Checking takes one click, and an answer that cannot show where it came from should not be trusted with a building business.
Your shorthand understood
Builders do not ask in full sentences. DA, CDC, PC, VO, EOT, GFA. An assistant built for Australian construction understands the shorthand the industry actually uses, so the question lands the way you asked it.
An honest blank
When nothing is on file, the right answer is that nothing is on file. A grounded assistant says so rather than inventing something fluent, which is precisely the failure a general chatbot cannot avoid.
Suggests, never commits
An assistant that can quietly change records is a liability wearing a convenience. The standard to demand is suggest only, with anything touching money or legal waiting for an explicit human commit.
Inside the workflow, not beside it
The useful assistant is embedded where the work happens, the command palette, the docked panel, the review queue, reading the same live numbers you see on screen, so its answers can never drift from the system of record.
03 / Under the hood
How Ask Oryn answers a question
Four steps, deterministic first, a model only when it earns its place, and evidence on everything that comes back.
- 01
You ask, from wherever you are
Open the command palette or the docked Oryn panel and ask in plain words. When is PC due on Riverview, which invoices are overdue, what variations are still pending approval.
- 02
Instant answers surface first, for free
As you type, matches from what is already on file appear immediately. Deterministic retrieval that reuses the app’s own screens, so what you see matches the Budget tab exactly, with no model call involved.
- 03
Press enter for a considered answer
If the instant answer misses, a language model composes one, restricted to the material retrieved from your job. It is forbidden from using outside knowledge, and it drops anything it cannot source.
- 04
The answer arrives with its evidence
Source chips point to the document and page or the in-app screen behind every claim in the answer. If nothing relevant is on file, Oryn tells you that, plainly, instead of guessing.
04 / The hard line
Why the assistant cannot touch your records
The most important design decision in Ask Oryn is what it is not allowed to do. Answering and acting are separated on purpose. Answers may be composed by a language model from retrieved material, but the part of the system that interprets an instruction to change something is deterministic, not a model, so a misread question can never quietly trigger a money action. When you do want Oryn to act, it stages the action as a frozen, human-readable sentence describing exactly what will happen, and nothing happens until you confirm it. Amber then green, on everything that matters.
What the assistant cannot do is also worth stating. It does not browse the internet, does not answer general questions, and does not opine on compliance, software can surface your approval and warranty paperwork, but obligations change and differ by state, so verify them with the relevant authority rather than any software's reading, including ours. It answers about your job, from your job, or it tells you nothing is on file. That narrowness is the feature. The same grounding discipline runs through the whole product, from document management to the accounts payable module, and the case for intelligence living inside the operating system rather than beside it is made in the AI construction operating system.
05 / FAQ
Common questions.
The phrase usually conjures a chatbot, a text box that answers building questions. The version actually worth having is different, grounded recall built into your construction software, answering questions from your own jobs, budgets, claims, orders, invoices and schedule, with a source attached to every answer. The distinction matters because a building business runs on specific numbers, and only an assistant with structured access to your records can answer about them truthfully. A general chatbot produces fluent text about jobs it has never seen.
Three ways. Grounding, Ask Oryn retrieves deterministically from your own facts, documents and the app’s own screens, and when a language model composes the final answer it may only use that retrieved material, with sources shown as chips. Honesty, if nothing is on file it says so rather than inventing an answer. And authority, it cannot change your records from a conversation. The only way an instruction leads to a change is a staged action, a frozen human-readable summary of exactly what will happen, which you confirm. A chatbot has none of those properties.
You can ask about what the system holds, which on a job running in VIABUILD is most of what matters. The job overview and dates, the budget and how it differs from the contract price, progress claims, contacts, the schedule including overdue tasks, variations approved and pending, purchase orders and committed money, invoices including overdue ones, and takeoff quantities. Australian shorthand works, DA, CDC, PC, VO, EOT, GFA. What it will not do is answer general knowledge questions, that is deliberate, an assistant grounded in your job stays truthful by refusing to leave it.
Because misreading a question must never trigger a money action. When you tell the system to do something, the part that interprets the instruction is deterministic, exact and repeatable, not a model inferring intent. A language model is allowed to help compose answers from retrieved material, where the worst failure is a wrong sentence you can check against its source. It is kept away from the path that changes records, where the worst failure is a wrong action. Different risk, different engine, which is the whole design philosophy.
Yes, because retrieval reuses the same live screens and records you work in rather than a separate copy. When an invoice is committed or a claim goes out, the next answer reflects it, the assistant reads from the same numbers on your screen, so it cannot drift from the system of record. That is a structural property of building the assistant inside the operating system rather than beside it, and it is the practical reason embedded intelligence beats a bolted-on chat window.
06 / Keep reading
Keep reading on assistants and grounding
The neighbouring cluster pages, and the references behind grounded recall.
Ask your next job a question.
Ask Oryn ships inside VIABUILD, the Construction Operating System for Australian residential builders. Every answer grounded in your own records, every answer with its sources showing, and your records untouched until you commit.