Security and privacy
Privacy and your data
What IA4AI keeps about people and calls, for how long, how content mode works, and how to export or erase it.
IA4AI is built for the GDPR first, because watching how people use work tools is regulated in the EU and expected to be transparent everywhere. You are the controller of your workspace's data; IA4AI processes it for you.
#What is kept
| From | Kept | Never kept |
|---|---|---|
| The gateway | Per call: model, provider, tokens, cost, latency, tool names, the agent, and fingerprints (hashes) used to group similar jobs | Prompts, answers, files, tool arguments and results |
| OpenTelemetry and direct ingest | The usage fields of each span; tool arguments and results redacted on arrival | Prompt and completion text: reduced to hashes |
| The browser extension | Per message: the tool, the model, the time, the message's length in characters, answer time, feature code words, and the person's work email (or a pseudonym) | What anyone typed or read, conversation ids (only salted hashes), any other website |
| Vendor admin data | Per licensed person: usage counts from the vendor's report | Anything the vendor's report does not contain |
| Your workspace | The people you invite (their email), your contact address, your stored keys (encrypted) | Passwords: people sign in with Google |
Prompts and answers are kept only under content mode, which is off unless your workspace's admin turns it on.
#Pseudonymous mode
A workspace can store people under a keyed pseudonym instead of their email, for browser usage and vendor admin data alike. Reports then show the pseudonym. We recommend it for every pilot, and we set it when we create your workspace if you ask.
#How long
| What | Kept for |
|---|---|
| Usage records (calls, traces, cases) | 30 days by default. It can be set longer, a year of cost history for example. |
| Captured content | 30 days by default, even when usage records are kept longer. |
| The evidence and audit logs | Kept: they are the record of what was stored and decided, and hold no prompt content. |
#Content mode
Content mode lets IA4AI keep redacted prompts and answers, so that a cheaper model can be verified on your real inputs. Only your workspace's own admin can turn it on, and an IA4AI operator cannot turn it on for you.
- Turning it on records who approved, for how long (1 to 90 days), which kinds (prompts and answers, reasoning, tool arguments), which agents (all, or some), the purpose, and that the terms were accepted. It closes by itself.
- Turning it off stops capture at once.
- Export returns everything captured for the workspace, decrypted, as JSON. It needs your approval code or a signed-in workspace admin: the workspace key alone is not enough. Every read is recorded.
- Erase removes every copy of the captured content. You confirm by typing the workspace id. The screen also turns content mode off by default, so capture does not start again; if you keep it on, capture goes on under a new key.
- Support access. IA4AI support cannot read captured content unless your admin grants access, for 24 hours to 7 days. Every read under it is listed for your admin, and it can be revoked at any moment.
Captured content is redacted before it is stored and encrypted with a key unique to your workspace. The redaction is pattern-based (API keys, tokens, emails, phone and card numbers, IP addresses): names and free-form personal details are not recognised, so choose the agents in scope with that in mind.
#A person's access and erasure requests
One request finds every record of a person in every table of the workspace, and exports it or erases it. The audit log records that the request happened without storing the person's email. Send requests to samstern@odux.io.
#Telling your staff
If you deploy the browser extension, tell people first. In the EU, monitoring work tools requires telling the people monitored, and sometimes consulting a works council. IA4AI provides a one-page staff notice, and the extension's popup shows each person what is recorded.
#Sub-processors
| Who | For | Data | Where |
|---|---|---|---|
| Hetzner Online GmbH | Hosting the service and its databases | All service data | Falkenstein, Germany |
| Signing in | Email address | Google's sign-in service | |
| DeepSeek | Writing investigation conclusions, only for a workspace that asked for the model | The findings' sentences (numbers and short labels) | China |
Your own model providers are not our sub-processors: the gateway calls them on your behalf, with your keys.
#The legal pack
A pilot comes with a beta agreement, a data processing agreement on the EU Standard Contractual Clauses, the sub-processor list, a privacy policy and a DPIA helper. They are drafts under legal review; we send the current versions with the pilot.

