Security
Your firm’s data, sealed off by design.
Plain answers about where your data lives, who can reach it, what the AI providers keep, and the limits of what we offer today.
At a glance
Per-firm stack- Isolation
- Own app, databases and network
- AI regions
- United States only
- Training
- Never on your data
- Transit
- TLS, browser to server
- Backups
- Nightly, encrypted
- Admin 2FA
- Required
Isolation
A separate workspace for every firm.
Not one shared system with rules to keep customers apart. Each firm gets its own stack, and there is no path between them.
-
Own stack
Each firm runs in its own application container with its own MongoDB database, its own search index and its own vector database for documents, all on a private internal network that belongs to that firm alone.
-
No path between firms
Firms cannot reach each other’s networks or databases. None of these databases is reachable from the internet.
-
Separate domain
Workspaces live at yourfirm.zrico.app, a separate domain from our marketing site and dashboard at zrico.com, so the two never share cookies.
Data flow
What happens to a request.
From the moment someone presses enter to the answer on their screen, in four steps.
-
Browser
Step 1: Someone at your firm asks
The request travels over TLS through Cloudflare to your firm’s own address at yourfirm.zrico.app.
-
Your workspace
Step 2: Your workspace prepares it
Your firm’s own application adds the conversation context and any relevant file excerpts, then sends it through our AI gateway, which applies your budget and stores no prompts or responses.
-
US region
Step 3: A US provider region answers
The request goes only to US processing regions of Anthropic (Claude), OpenAI (GPT) or AWS (Claude fallback), through business APIs that do not train on your data.
-
Your database
Step 4: The answer is saved with you
The response comes back to your workspace and is stored in your firm’s own database, for as long as your retention setting allows.
US-only processing
Requests are sent only to US processing regions of Anthropic, OpenAI and AWS.
No training on your data
We use the providers’ business APIs, under terms that do not allow them to train models on your prompts, files or answers. We don’t train models either.
Controls
Access, encryption and backups.
Who can sign in, how traffic is protected, and how long anything is kept.
-
Rolling out
Microsoft 365 sign-in
Rolling out to workspaces now. Your IT admin can connect Microsoft 365 so staff sign in with the accounts they already have. Your workspace only accepts accounts from your own Entra tenant, so nobody from another organization can sign in, even with a valid Microsoft account.
-
Two-factor authentication
Required for firm owners and admins. Offered to every user, and workspaces are invite-only with no public sign-up.
-
Encryption in transit
All traffic uses TLS: from your browser to Cloudflare, and from Cloudflare to our servers with origin certificates.
-
Nightly encrypted backups
Each workspace is backed up every night. Backups are encrypted before they leave the server and are stored off the server.
-
Logs without your content
Our operational logs record events like sign-ins, errors and request counts. They never contain prompts, AI responses or uploaded files.
-
Retention you control
Your admins choose how long conversations are kept, or have them deleted automatically after a set period.
AI data use
What each provider retains.
Retention periods and training terms for every model in your workspace, and what our own gateway keeps.
Full AI data use pageLast updated: October 14, 2026
This page explains what happens to your firm's information when someone in your Zrico workspace uses an AI model. It is written for lawyers, accountants and brokers who need to answer client questions about confidentiality.
The short version
- Your data goes to an AI provider only when someone in your firm uses a model.
- None of the providers we use train their models on data sent through our API accounts.
- Each provider keeps data for a limited time to detect abuse, up to 30 days. AWS Bedrock keeps nothing for the models we use.
- All AI processing happens in the United States.
- Our own AI gateway does not store your prompts or the responses.
What we send to AI providers
When a user sends a message, the workspace sends the provider:
- The prompt the user typed.
- Conversation context: earlier messages in the same conversation, so the model can follow it.
- Relevant excerpts of uploaded files. When a user asks about an uploaded document, the workspace finds the most relevant passages and sends those, not necessarily the whole file.
When a user uploads a file for search, the workspace also sends its text to OpenAI's embeddings model (text-embedding-3-small). This turns the text into numbers that let the workspace find relevant passages later. The embeddings are stored in your firm's own workspace database, not by OpenAI.
We send only what is needed to answer the request. We do not send your firm's other conversations or files.
No training on your data
Each provider's API terms say that data sent through the API is not used to train their models by default. We have not opted in to any data-sharing program with any provider.
Provider by provider
| Provider | Models in Zrico | How long the provider keeps it | Used for training? | Source |
|---|---|---|---|---|
| Anthropic | Claude (Sonnet, Opus, Haiku) | Up to 30 days. Content flagged for a usage policy violation: up to 2 years. | No | Anthropic data retention |
| AWS Bedrock | Claude (backup route when Anthropic's API is unavailable) | Not stored for the Claude models we use. | No | Bedrock data retention, Bedrock abuse detection |
| OpenAI | GPT, and text embeddings for file search | Up to 30 days for abuse monitoring. | No | OpenAI: your data |
A few details behind the table:
- Anthropic may also keep automated safety scores (not the content itself) for longer, up to 7 years.
- Human review: each provider may have authorized staff look at content that its automated systems flag as a possible policy violation.
- Law and legal holds can extend any of these periods. A provider may keep data longer if a court order, legal hold or law requires it.
Models we do not use
Some AI providers offer models that require them to keep every request for a set period and allow safety staff to review it. We do not route your firm's requests to any model with terms like that. If a provider changes its terms so that a model we use would require longer retention, we will stop routing to it or tell you before anything changes.
Where processing happens
All AI requests are processed in the United States. On AWS Bedrock we use US-only routing. Your workspace itself is hosted in a US data center.
What our gateway keeps
All AI requests pass through our AI gateway, which enforces your firm's usage budget. The gateway does not store prompts, responses or file contents. It records only usage metadata: the user's email address, the model used, the number of tokens and the cost. We keep that metadata for 400 days for billing and support.
Our system logs never contain prompts, responses, file contents or secrets.
Zero data retention
Some providers offer "zero data retention" agreements, under which they keep nothing at all. We do not have these agreements in place today. They are on our roadmap. We will update this page if that changes.
What your firm can do
- Set a chat retention period. In your firm's settings, choose to keep chats, or delete them automatically after 30, 90 or 365 days. This controls what is kept in your workspace. It does not change what providers keep (described above).
- Leave out client identifiers you do not need. Often the model does not need a client's name, account number or other identifiers to help. Use initials or a matter number instead.
- Choose the model with care. If a client has concerns about a specific provider, ask your users to pick a different model for that work.
- Do not enter information that does not belong in Zrico, such as patient health information. See our Acceptable Use Policy.
- Check your professional rules. Your duties of confidentiality and any client instructions still apply.
More information
Questions: support@zrico.com.
Limits
What we don’t claim.
Being clear about what Zrico is not for matters as much as what it does.
- Certifications We do not have SOC 2, HIPAA or any other security certification today.
- Health data Zrico is not for patient health information. We don’t offer a HIPAA business associate agreement.
- Regulated comms Zrico is not for broker-dealer or registered investment adviser regulated communications. It is not a compliant archive.
- AI output AI output can be wrong. It must be reviewed by a qualified professional before anyone relies on it.
If something goes wrong
Incident response
If a security incident affects your data, we notify affected customers without undue delay, explain what happened and what we are doing about it.
Our incident response policyFor researchers
Responsible disclosure
Found a vulnerability? Email security@zrico.com. We’ll acknowledge your report and keep you updated. Please don’t access other customers’ data or disrupt the service while testing.
security.txtGive your firm a private place to use AI.
$39 per seat per month after a 30-day free trial, 5-seat minimum. Your workspace is ready minutes after you sign up.
Your card is charged when the trial ends. Cancel any time before then and you pay nothing.