Key points
- An agent is an AI system that pursues a goal by planning and using tools, not just replying with text.
- Agents are most valuable for multi-step, repetitive work with clear rules and easy-to-check results.
- Safe agents use least-privilege access, approval steps, limits, and logging — every time.
Chatbot vs. agent
A chatbot takes your message and returns an answer. An agent takes a goal — "triage this week's volunteer sign-ups and draft confirmation emails" — then decides what steps to take, uses tools (email, calendar, spreadsheets, databases, web search), looks at the results, and keeps going until the job is done or it needs help.
| Chatbot | Agent | |
|---|---|---|
| Input | A question or request | A goal |
| Output | Text | Completed actions plus a report |
| Tools | Usually none | Email, files, apps, APIs, browser |
| Main risk | Wrong answer | Wrong action |
How an agent works
Under the hood, most agents run a simple loop:
goal: "Follow up on unpaid invoices over 30 days"
loop:
1. think → "I need the list of overdue invoices"
2. act → accounting.search(status="overdue", days=30)
3. observe → "7 invoices found"
4. act → email.draft(…) # × 7
5. pause → "Drafts ready — approve before sending?" # human in the loop
until: goal met or needs a humanThe model provides the judgment ("what should I do next?"), while tools provide the ability to act. Standards like the Model Context Protocol make it much easier to connect agents to tools.
Real-world examples
- Inbox triage: sort incoming requests, draft replies, and route urgent items to the right person.
- Research & reporting: gather information from several sources and assemble a weekly brief.
- Scheduling: coordinate volunteer rotations or client appointments across calendars.
- Data cleanup: find duplicate contacts, fill missing fields, and flag inconsistencies for review.
- IT & operations: answer common help-desk questions and open tickets with the right details.
- Software development: coding agents that write, test, and fix code under developer review.
Choose the right level of autonomy
Not every agent should be fully autonomous. We help clients pick the lowest level of autonomy that still delivers the value.
| Level | What the agent does | Good for |
|---|---|---|
| 1 · Suggest | Recommends actions; a person does them | Anything sensitive or new |
| 2 · Draft & approve | Prepares the action; a person approves each one | Emails, record updates, payments |
| 3 · Act within limits | Acts on its own inside strict boundaries | Low-risk, high-volume, reversible tasks |
| 4 · Autonomous | Acts freely and reports afterward | Rarely appropriate outside sandboxes |
What can go wrong
- Wrong actions: a misunderstanding becomes a sent email or a deleted record, not just a bad answer.
- Prompt injection: instructions hidden in an email or web page trick the agent into doing something else. See AI security essentials.
- Permission creep: an agent given broad access can reach far more than its task requires.
- Runaway costs or loops: an agent that keeps retrying can burn through usage budgets.
- Unclear accountability: when something goes wrong, who owns it?
Guardrails we build into every agent
- Least privilege: separate, scoped credentials — read-only wherever possible.
- Human approval for anything external, financial, irreversible, or affecting a person.
- Limits: caps on actions, spending, and run time, plus an off switch.
- Logging: a full audit trail of what the agent saw, decided, and did.
- Testing: evaluation against real scenarios, including adversarial ones, before launch.
- A named owner responsible for each agent, recorded in your AI use register.
How to get started
Start with one well-defined, repetitive process that's easy to check — something your team does weekly and would happily hand off. Run it at "draft & approve" for a few weeks, measure time saved and error rates, and only then consider more autonomy.
Need help choosing the first use case? A WHU AI automation pilot takes you from idea to measured results in a few weeks.