What is an AI agent in KAI?
An AI agent in KAI is a predefined role with its own behaviour, its own knowledge areas and a fixed, explicit list of permitted actions. These roles are defined centrally – a tenant can enable or disable an agent or restrict it further through its own policies, but can never give an agent more permissions than it has by design.
Specialised agents and roles
Each agent deliberately stays within its domain – for questions outside it, it points to the right assistant instead of guessing.
| Agent | Domain | Visibility |
|---|---|---|
| KAI Assistant | General assistant – sees exactly what the requesting person already has access to. The default for every request without a selected agent. | Visible by default |
| IT Operations | Infrastructure, systems, backups and monitoring. The only agent with action permissions (see below). | Visible by default |
| Developer | Writing, explaining, reviewing and debugging code. | Visible by default |
| Security Analyst | Security posture, network security and security-related diagnostics. | Only when assigned |
| HR Assistant | HR policies and general company knowledge – without access to infrastructure, code or finance. | Only when assigned |
| Finance Assistant | Finance, accounting and general company knowledge – with the same separation as HR. | Only when assigned |
Tools, capabilities and permissions
Most agents have no action permissions at all – they answer questions from knowledge without executing anything. Only the IT Operations agent currently has a narrowly defined list of permitted actions: read-only diagnostics (system status, service status, recent service logs, disk usage, Docker container status) and two changing actions (restart service, reload service). Each of these actions additionally passes tenant, role, time-window and target-system checks before it can even be proposed – a permission can only ever narrow this chain, never bypass it.
Human-in-the-loop and approvals
No agent executes changing actions directly. It can only create a request – another, separately authorised person must review and approve that request before anything is executed. Self-approval by the same person is not possible.
Diagnostics vs. active changes
KAI structurally distinguishes between read-only diagnostics (e.g. status or log queries) and actions that actively change something. Diagnostics are the rule and run without approval; changing actions are the deliberate exception and always go through the full approval process.
Auditable execution
Every agent activity – proposed, approved, rejected or executed – is logged and traceable in an administrative overview. AI actions therefore remain reviewable instead of being a black box.
AI agents and company knowledge
An agent never gets more knowledge access than a normal, access-controlled request: just like RAG at KAI, it only searches the documents and sources the requesting person is already allowed to access, and it is additionally limited to its own domain knowledge area.
Agents are the acting part of a larger control structure.
RAG governs knowledge access, Private AI governs where processing happens and which provider is used, and our governance page describes which rules, approvals and logs apply overall.
AI agents at KAI in detail.
Do KAI agents work fully autonomously?
No. Every agent has a fixed role, fixed knowledge areas and an explicit list of permitted actions. Most agents have no action permissions at all. No agent can grant itself additional permissions.
Can agents make changes to systems?
Only the IT Operations agent has any action permissions at all, and only two narrowly defined actions among them change anything. Both require approval by another, separately authorised person – self-approval is not possible.
How are risky actions controlled?
Through several independent checks that must all pass: assignment to the agent, tenant and role/department permissions, parameter and target-system checks, the online status of the installation and its security posture. Only then can an approval be requested.
Are agent activities logged?
Yes. Proposed, approved, rejected and executed actions are traceable in an administrative activity overview.
Can agents use internal company knowledge?
Yes, but only within the knowledge areas and access rights assigned to the respective agent and the requesting person – just like a normal, access-controlled RAG request.
Talk to us about controlled AI agents for your business.
We will show you which agent roles and approval processes fit your organisation.