AI agents that do real work, within clear limits
An AI agent does not just answer. It plans steps, uses tools such as your ticketing, directory or ERP systems, and checks its own results. That makes it useful, and it also makes limits essential: what it may read, what it may change, and what a person must approve.
An IT service desk agent, step by step
The agent fixes a VPN problem on its own, is blocked when it tries something outside its permissions, and stops to ask a manager before granting access to financial data. You decide whether the manager approves.
Chatbot, assistant or agent?
| What it does | Example | Risk if it goes wrong | |
|---|---|---|---|
| Chatbot | Answers from what it knows | General questions | Wrong answer |
| Assistant with RAG | Answers from your documents, with sources | Policy and product questions | Wrong answer, or a leak if permissions are ignored |
| Agent with read-only tools | Looks things up in your systems to help | Checks order status, reads logs | Wrong conclusion |
| Agent that can act | Changes things: creates tickets, resets accounts, updates records | Renews a certificate, raises a purchase request | A wrong change in a real system |
Most organisations should move down this table one step at a time, proving each level before adding the next.
Guides and examples
MCP architecture
How the Model Context Protocol connects agents to your systems, and how to design the servers.
Read the guide →Guardrails and approvals
Least privilege, human approval, prompt injection and audit.
Read the guide →Running agents
Tracing, evaluation, cost and how to roll out safely.
Read the guide →Use case: IT service desk agent
An illustrative scenario: from read-only helper to approved actions, in three phases.
Read the use case →Agent questions, answered
Autonomy, safety, cost and where agents work well today.
Read the FAQ →Related: RAG for answering from documents, and enterprise LLM for the model behind the agent. Full reference: agentic AI whitepaper.
Good first candidates
- IT and HR service desks: triage, diagnosis, and routine fixes such as unlocks and certificate renewals.
- Document-heavy checks: comparing invoices with purchase orders, or tender documents with requirements.
- Operations support: gathering logs and metrics during an incident and drafting the summary for engineers.
- Internal requests: collecting details, checking policy and raising a correctly filled request for approval.
Good candidates are frequent, rule-based, involve several systems, and have a clear point where a person can approve.
Thinking about AI agents?
Tell us the process you would like to automate and the systems it touches. We will come back with a plain view of what an agent could safely do, what should stay with people, and how to start small.