Who decides what an AI employee may do?
The server. An AI can ask to do anything; the tool layer decides whether it is allowed. Instructions given to a model are never treated as a security control.
Files, vaults, the CRM, dashboards, search and the AI tool layer all call the same check, can(member, action, resource), for people and AI alike. Tasks stay in your task tool, Tecfy Tasks first, which applies its own rules to the access OrgKeep syncs into it, and each AI employee has its own account there. Access follows project membership: an AI employee added to a project can use exactly that project's files, repositories and credentials, and nothing else.
How can an AI use a password it cannot see?
Credentials live in named vaults. A member with Use access can use a credential without seeing its value:
- The AI employee decides to call a system, such as the online store's API, and passes a reference to the vault entry.
- The credential broker on the server inserts the secret into the outgoing request.
- The result comes back to the AI. The secret never does.
Give each AI employee its own API key with only the permissions it needs. It can be revoked on its own, and the outside system's logs show exactly what the AI did. Secret values are never indexed by search.
Which steps always need a person?
Payments, contracts and production deploys always need a human approval, whatever an AI employee's autonomy settings say.
- Workflows can mark any step as human-only.
- A rejected task goes back to the previous assignee with the reviewer's comments.
- When an AI employee has a question, the task shows Waiting on you for the person asked, and work resumes when they answer.
- An AI employee that manages people can't approve a human-only step for them. That step goes to a person.
Who can an AI employee email?
Only the people an admin allows. Each AI employee's reply policy says whether it may write to people outside the organization, and what it does on a thread with both. The server applies the policy to every outgoing email, so a prompt can't get around it. See the reply policy options.
What gets logged?
Every action, under the member who took it, human or AI, in one audit log. Each AI employee also has a monthly budget, and its cost per task is visible to its manager.
Why only reviewed skills?
Public skill marketplaces for AI agents have carried malware. OrgKeep employees use only skills from your organization's reviewed library, with versions, and a manager approves any new skill an AI employee proposes.
Where does your data go?
Nowhere you don't send it. OrgKeep runs on your servers, telemetry is off by default, and with local models no prompt leaves your network.