Skip to content

Security & data handling

This page explains, in plain terms, what happens to your data when you use VitruAI — what leaves your machine, which AI models process it, whether it's used to train anything, how firms are isolated, where data is stored, and the options available to security-sensitive and enterprise customers.

The short version

  • Your .rvt model file never leaves your machine. VitruAI reads the open model locally and sends only the specific data a check needs.
  • Model access is read-only — no VitruAI tool can modify your Revit model.
  • Your data is not used to train AI models.
  • Each firm's data is strictly isolated from every other firm.
  • You choose your model (Anthropic Claude by default; others or self-hosted), and can bring your own key.
  • For strict data-governance, an enterprise deployment runs in a dedicated region, or fully on your own infrastructure — so nothing leaves your environment.
  • A licensed professional stays responsible for every decision; findings are explainable and auditable.

What leaves your machine (and what doesn't)

VitruAI is split between a local add-in inside Revit and the cloud service your agents run in. The boundary matters:

Stays on your machine Leaves your machine (to run a check)
The .rvt project file — never uploaded Extracted, task-specific data: element parameters, categories/types, geometry summaries (bounding boxes, centrelines, areas), room/spatial data, clearances, sheet names
Local logs (%APPDATA%\VitruAI\Logs) Rendered images of the active view or sheets — only when a rule needs visual evidence
Reports cached locally (%APPDATA%\VitruAI\Reports) Computed rule results (a rule can compile and run locally; the verdict flows up, not raw geometry)

The add-in reads your model through the Revit API over a local loopback bridge and streams only the results of each requested command to the cloud service over HTTPS. There is no step that uploads or copies your project file to our servers. In an enterprise deployment, even that extracted data can stay entirely within your infrastructure.

Model access is read-only

No agent tool can change your model. Rule code is safety-checked before it runs to forbid model mutation and file, network, and process access. See Agents, tooling & delegation and Authoring checks.

The AI models we use

VitruAI is an agentic system built on frontier models, used over their commercial APIs (or self-hosted — see enterprise deployment):

Role Model
The agents' reasoning Anthropic Claude by default — switchable to OpenAI, Google Gemini, Kimi, or self-hosted open models
Document search (embeddings) OpenAI
OCR (scanned documents) Mistral

The checks are deterministic — the model doesn't 'guess' compliance

The AI decides which checks to run and reads the model, but the pass/fail measurement itself is executed as deterministic compiled rules, not a language-model opinion. That's what makes findings repeatable and auditable.

Your data isn't used to train AI

VitruAI does not train models on your data. The providers we use commit — for business/API use — not to train on the content you send:

  • Anthropic (Claude): "Anthropic may not train models on Customer Content from Services." (Commercial Terms)
  • OpenAI: business and API data is not used to train models by default. (Business data)
  • Google (Gemini, paid): "Google doesn't use your prompts … or responses to improve our products." (Gemini API terms)
  • Mistral (OCR, paid API): "we do not use your Input and Output to train our artificial intelligence models when you use … the paid version of our APIs." (Privacy Policy)

In an enterprise deployment on self-hosted open models, your prompts and outputs never leave your environment, so there is no third-party training exposure at all.

Bring Your Own Key (BYOK)

Firms can run VitruAI on their own model API key. When enabled, every model call for the firm routes through the firm's own provider account, billed to the firm and governed by the firm's own terms. The key is stored per-firm, encrypted at rest, and the system fails closed — it never silently falls back to a shared key.

What BYOK is — and isn't

BYOK is an account-of-record control: it changes whose account and terms your model calls fall under. It does not by itself change where the hosted application runs. To also keep the data path inside your environment, see enterprise deployment.

Where your data is processed and stored

  • Hosted (SaaS) region. The shared service runs in the United States (US-East) on Hetzner Cloud. Cloudflare provides authoritative DNS; TLS is terminated at the origin (Let's Encrypt via Caddy). Enterprise customers can select a region or deploy on their own infrastructure.
  • What we persist, per firm: your chat transcripts; verification reports and findings (saved to your firm's version-controlled workspace when a check is scoped to a project); documents you upload and their search embeddings; and billing/usage records.
  • What we don't store: your .rvt files (they're never uploaded).

Data retention & deletion

Your account and model data are retained for the life of your firm's account. You can request an export or deletion of your data in writing; we action it within 30 days. Encrypted off-site backups roll on an approximately 30-day cycle.

Tenant isolation

Your account belongs to a firm, and every firm's data is isolated from every other firm's. Isolation is enforced in code at the database layer — each read and write is scoped to your firm, and the system fails closed if a request lacks a firm context. See Concepts → Firms, members, and roles.

Authentication & access

  • Sign-in is email + password with mandatory email verification; optional two-factor.
  • The Revit add-in pairs as a connected device with its own token; owners and admins can see and revoke any device.
  • Project access is granted per member (viewer / contributor / maintainer) — see the Firm-admin guide.
  • Single sign-on (SSO / SAML) is available for Enterprise customers.

Encryption

  • In transit: TLS on all connections.
  • At rest: API keys and secrets are encrypted; off-site backups are encrypted.

Enterprise & sovereign deployment

For security-sensitive firms, VitruAI can be deployed well beyond the shared cloud. Talk to us ([email protected]) or see vitruai.com/enterprise:

Option What it does
Dedicated single-tenant instance + region selection Your own isolated instance, in a region you choose.
On-premise / sovereign deployment Runs in your own datacentre or VPC with self-hosted open models — your data never leaves your infrastructure.
Model choice & BYOK Claude, OpenAI, Gemini, Kimi, or self-hosted open models; or your own key.
SSO / SAML, audit logs, DPA Enterprise identity, logging, and a signed Data Processing Agreement.
Trust Center pack Sub-processor list and data-flow inventory on request. SOC 2 is in progress.

Security FAQ

Is my Revit model uploaded to VitruAI? No. Your .rvt file never leaves your machine. VitruAI reads the open model locally and sends only the specific data a check needs (parameters, geometry, room data, and sheet images where a rule needs visual evidence).

Is my data used to train AI models? No — see above. VitruAI doesn't train on your data, and the providers commit not to train on business/API content.

Can VitruAI change my model? No. Every model tool is read-only, and rule code is safety-checked to forbid mutation and file, network, and process access.

Can it run on-premise or fully in our own environment? Yes, in an enterprise deployment — a dedicated instance with region selection, or a full on-premise/sovereign install on your own servers with self-hosted open models. ("Local" means your server or VPC, not each workstation.)

Which models can we use? Anthropic Claude by default; you can switch to OpenAI, Google Gemini, Kimi, or self-hosted open models, and you can bring your own key.

Where is my data hosted? The shared SaaS runs in the United States (US-East). Enterprise customers can select a region, a dedicated instance, or an on-premise deployment.

Do you have a DPA / can you complete a security questionnaire? Yes — contact us for the Trust Center pack.

See also