- Privacy means your model and your data never have to leave your building. On the on-premise tier, the application itself runs on your own server, not just a private link to someone else’s cloud.
- A common data environment is one shared system, not a hundred separate chat threads. Your team’s projects, findings, and history live together, searchable, instead of scattered across everyone’s private tabs.
- Real collaboration is already shipped for your internal team. Agents working alongside outside partners in shared, governed spaces is the obvious next step, and it’s on the roadmap, not in the product yet.
- Your own libraries — your rules, your standards — live in a workspace you own. Teach it once, and upgrades are built to never overwrite what you taught it.
- A general AI assistant can’t answer any of these four questions. Not because it’s missing a feature — because none of them were what a model endpoint was built to do.
You started experimenting with ChatGPT or Claude, and you liked it. Ask it to draft an RFI response, summarize a spec section, or sketch three ways to frame a lobby, and it answers fast, fluently, without ever getting tired of your questions.
Your clients noticed the same thing. They’re already using AI to generate visualizations and talk through ideas — showing up to a kickoff meeting with a rendering pulled from a prompt the night before, asking your team to make it real.
Definitely, AI is the future for AEC. Nobody serious is still debating that part.
But the more you actually use ChatGPT, the more a different kind of question starts creeping in. Not “can it write me something decent” — you already know it can. It’s the harder ones: privacy. A common data environment. Real collaboration. A library that’s actually your own.
And external AI providers can’t answer these questions. Not because they haven’t gotten around to it yet — because none of it is what a model endpoint was built to do.
We’ve been building VitruAI with all of this in mind, because we heard it from you — from firms living inside these tools for months, asking exactly these questions on exactly these calls.
So let’s unfurl it.
Privacy: where does your data actually go?
Every time you paste a description of a model into a general AI tool, that text leaves your building and lands on somebody else’s server, under somebody else’s terms. Most of the time you don’t get to see what happens to it next.
With VitruAI, the model file itself never leaves your machine, on any tier, including our standard cloud product. Our Revit add-in reads the open model locally and sends only what a specific check needs — never the file itself, never a copy staged anywhere. Access is also read-only by design: the system can’t edit your document, because any code that would touch the file, reach the network, or launch a process gets rejected before it’s ever allowed to run, not after the fact.
On the on-premise tier, we take that further. The application itself — not just the model — lives on a server inside your own building, under your firm’s control. The only thing that leaves is the specific data a given check needs, sent to the model provider you choose, under your own account and your own contract, with a documented commitment that it won’t be used to train anything. Your projects, your reports, and your history stay exactly where you put them.
A common data environment a chat window can’t give you
A conversation in ChatGPT disappears into its own thread the moment you open a new tab. There’s no shared memory across your team, no single place where a project’s findings, models, and history live together — just a growing pile of separate conversations that nobody else on the job can see.
VitruAI’s on-premise tier is built to be the opposite: one system, on your own server, holding your team’s projects, findings, reports, and rules together, searchable across the whole firm. Every database read and write is scoped to your firm and fails closed if a request ever arrives without the right context, so the environment stays exactly as contained as you need it to be. It’s the same system every person on your team opens, not a private thread each of them started separately.
Real collaboration — today, and what’s next
Today, that means your whole team already works inside one shared system: web access for everyone, no requirement to have Revit open, and a coordinator agent that dispatches discipline specialists to run checks in parallel, each with its own permissions and its own transcript you can open and read. Project access is granted per person, and a device that leaves the firm can be cut off in one click.
Where we’re going next is the obvious extension of that: agents as full participants in your firm’s coordination, sitting alongside your people and your external partners in shared, governed spaces — not just a tool one person opens in a private tab. That collaboration fabric is still ahead of us. We’re telling you that plainly rather than letting today’s demo imply it’s already here.
Your own libraries, not ours
The best thing you can teach a general AI model is a prompt. The moment you close the tab, that teaching is gone, and the next person on your team starts from zero.
Every firm on VitruAI gets its own version-controlled workspace: its QA standards, its rules, its project templates, and the instructions that shape how each agent behaves. It’s browsable inside the app. Edit a file and the change applies on the next run — no deploy, no ticket to us, no wait for our next release. Every edit is committed, so you can show exactly which version of your standard applied to a check that ran six months ago.
When we packaged our first enterprise on-premise deployment, the box shipped with that firm’s own body of knowledge already loaded.
None of that is our engine. It’s the firm’s standards, written in a form the system can read.
And when a new release lands, our upgrade process reconciles the platform files and leaves everything the firm authored alone — not by policy, but because that content sits outside the part of the file tree an upgrade is allowed to touch. Nothing is locked in, either: the workspace is a git repository, so if a firm ever leaves, it takes its rules with it, in a form another system could read.
Why a general AI assistant can’t answer any of this
We build on frontier models, and we say that outright — Claude and GPT-class models reason better than anything we could run ourselves on a CPU box today. But renting the smartest model in the world still doesn’t hand you privacy controls, a shared environment, real collaboration, or a place to keep your own standards. Those aren’t missing features a general assistant will add in its next update. They require a desktop integration that opens your live model, a workspace built to hold a firm’s knowledge, and an access model built for a team, not a single person’s account.
A general AI assistant is a very capable person who has never seen your building, can’t open your model, doesn’t know your standards, and gives a slightly different answer every time you ask.
What we’re building next
Today, the reasoning runs on the model provider you choose, under your own account and your own terms, and nothing else leaves the building. For firms where even that outbound call isn’t acceptable, a fully self-hosted model configuration is next on our roadmap — not available yet, and we’re not going to pretend otherwise.
Behind that, a tamper-evident, chained audit log with a governance console is in progress: the write seam that records each governed decision already exists, but the queryable, exportable chain around it doesn’t yet. Both get roadmap language here on purpose, not present-tense marketing.
Our own access to a customer’s box, meanwhile, is already as small as it sounds: sudo on that one VM and a VPN peer, expected to be revoked at handover and re-granted only for the next upgrade. We had it. They revoked it. The system kept running. That’s the whole argument, in three sentences.
Ask these four questions before you trust any AI vendor with your firm
- On privacy: ask exactly what data leaves your building, where it goes, and under whose contract. “It’s secure” isn’t an answer. A specific data flow is.
- On a common data environment: ask whether your whole team can work from the same system on the same project, or whether what you’re really buying is one person’s private chat history.
- On collaboration: ask what happens the day you need to loop in a structural consultant or an outside partner. If the answer is “email them the PDF,” you already have your answer.
- On your own libraries: ask what happens to your custom rules and standards the next time the vendor ships an update. If they can’t tell you clearly, assume the answer is that they get overwritten.
Bring your privacy, collaboration, and standards questions. We built the product around exactly them — so see how VitruAI answers your four questions against your own model.
Book a 1:1 call → See how Vitru works
FAQ
What does “on-premise AI” actually mean for an architecture firm?
It means the whole application — not just a private connection to someone else’s cloud — runs on a server your firm owns and operates. For VitruAI, that’s the agents, the database, the reports, and the admin console, all on one server, with the only required outbound connection going to your model provider on your own API key.
Why can’t we just get these answers from ChatGPT or Claude directly?
Because a general assistant can’t open your live Revit model, can’t guarantee where a pasted description ends up, has no shared environment for your whole team, and has nowhere to permanently store your firm’s standards. VitruAI runs on top of those same frontier models but adds the system around them — the model access, the shared workspace, and the governed library that holds what makes your firm’s standards yours.
Do we need a data center or a GPU cluster to run this?
No, and this trips up almost everyone the first time they ask. The reasoning runs on a frontier model over an API, on your own key. Your server only runs the application, and that fits on one modest virtual machine — 8 vCPU, 16 GB RAM, 300 GB of disk, no GPU. It’s a hardware conversation your IT team can close in one meeting, not a data-center build.
Can our whole team work in the same system, or is this a one-person tool?
Your whole team, by design. Every person gets web access with no requirement to have Revit open, project access is granted per person, and a coordinator agent dispatches discipline specialists that run in parallel, each with its own permissions and its own transcript. It’s one shared environment for the project, not a private chat window per employee.
What happens to our own rules and standards when you ship an update?
They’re left alone. Your rules live in a version-controlled workspace that sits outside the part of the file tree our upgrades are allowed to touch, so an update can change the platform without touching what your firm authored. The workspace is also a git repository, so your content is portable even if you ever decide to leave.
The model call still has to go out — is that a privacy problem?
Here’s exactly what leaves: the specific data a given check needs, sent to your own account, under your own contract with the model provider, with a documented no-training commitment. Your project files, your rules, your reports, and your history stay on your box the entire time.
What happens to the system if VitruAI goes out of business?
The box keeps running. There’s no license server, no phone-home, and no registry dependency it needs from us. Your rules live as text in a git repository you hold, not in a database only we can read.
About VitruAI
VitruAI is an AI agent platform for architecture and engineering firms, developed by ADAIA. QA/QC and standards checking for Revit is live today, with web access for the whole team and no requirement to have Revit open. From there, the same agent layer is extending across the project lifecycle: edit mode that fixes issues instead of just flagging them, company-level rule libraries, image-to-model, a worldwide building-code catalog, document ingestion, an ArchiCAD plugin with pure IFC checking, and a project coordinator agent handling emails, project plans, and Gantt charts. The founder first worked with AI in 2007, helped build technology later acquired by Dell for $130 million, and has since helped accelerate 500+ companies using AI as a core part of how they operate.
Bring your four questions. Book a 1:1 call and we’ll answer them against your own model.