Most explanations of AI agents stay abstract. This page goes department by department through the kinds of tasks agents are being given at work, and puts the human role and the level of review next to each one. The examples are illustrative, to show the pattern; they are not a survey of how widely each is used.
The short answer: AI agents at work typically take on multi-step, rules-based tasks such as handling routine requests, matching documents, researching, drafting and updating records. In every useful example a person still sets the goal and the limits, approves high-impact actions and answers for the outcome. The more an agent can change, the more review it needs.
Examples by department
| Department | What the agent does | What the person does | Review level |
|---|---|---|---|
| Customer service | Reads a refund request, checks the order and drafts or issues the refund within set limits | Sets refund limits, handles exceptions and complaints, reviews a sample | Medium: approve above a set amount |
| Finance | Matches invoices to purchase orders and flags mismatches | Approves payments, investigates flagged items | High: a person approves every payment |
| HR | Answers policy questions and schedules interviews | Owns hiring decisions, handles sensitive cases | High for anything about hiring or performance |
| Sales | Researches prospects and drafts follow-up emails | Decides who to contact and what to offer, sends or approves messages | Medium: review before sending |
| IT support | Resets passwords, routes tickets and suggests fixes from past cases | Handles access changes, outages and security issues | Medium to high: limit what it can change |
| Operations | Tracks shipments and updates customers on delays | Decides on rerouting, refunds and supplier issues | Low to medium |
| Marketing | Produces variations of ads and posts for testing | Sets strategy, checks claims, approves what is published | Medium: nothing published unreviewed |
| Software teams | Writes code changes and tests for a described task | Reviews and merges changes, owns the system | High: human review before merge |
| Research and analysis | Gathers sources and drafts a summary or first analysis | Checks sources and numbers, draws conclusions | Medium: verify every source cited |
An agent here means software that takes actions toward a goal through connected tools, not only a chatbot that answers questions. See AI agent vs chatbot for the difference.
What the good examples have in common
- A narrow, clear task with a defined start and finish.
- Limited permissions: the agent can only touch what the task needs. The security project OWASP recommends limiting AI tools to the minimum functionality and permissions necessary (OWASP LLM06:2025).
- Approval for high-impact steps such as payments, deletions, external messages and anything about a person's job. OWASP recommends human-in-the-loop control to require a human to approve high-impact actions.
- Logs so someone can see what the agent did and why.
- A named owner who answers for the agent's work. See human in the loop and accountability.
Where to be careful
- Decisions about people: in the EU, AI used for recruitment, promotion, task allocation or performance monitoring is listed as high-risk under the AI Act. See AI at work and workers' rights.
- Money and access: payments, refunds and account changes are where an agent's mistake costs most. See AI agent security.
- Personal data: agents often read emails, files and records full of it. See AI agents and personal data.
- Wrong but confident output: NIST calls this confabulation (NIST AI 600-1). An agent that acts on a wrong assumption can repeat it across many records before anyone notices.
How to start
The safest first agents take on a task that is frequent, rules-based and easy to check, with a low cost if something goes wrong: routing tickets, matching documents, drafting replies for review. Expand what an agent may do only after its results have been measured. Our step-by-step guide to introducing AI agents at work covers naming roles, writing a job for each agent and matching review to risk, and measuring AI agent performance covers how to tell whether it is working.
What it means for the people involved
In each example, the person's job moves from doing every step to briefing, checking and handling the exceptions. That is the core of digital collar work, and the skills it needs are covered in digital collar skills and AI agents as coworkers. Some organizations give this supervision its own role; see the AI agent manager job description.
Questions and answers
What are examples of AI agents at work?
Handling routine refund requests, matching invoices to orders, scheduling interviews, resetting passwords, researching prospects, drafting code changes and tracking shipments, each with a person setting limits and approving high-impact steps.
Which tasks are AI agents best suited to?
Frequent, rules-based, multi-step tasks that are easy to check and cheap to correct, such as routing, matching documents and drafting replies for review.
What should an AI agent not do without approval?
High-impact actions such as payments, deletions, external messages, access changes and decisions about people. OWASP recommends requiring human approval for high-impact actions.
Who is responsible for what an AI agent does?
The organization using it, through a named person who sets the agent's goal and limits and reviews its work.
How do you start using AI agents at work?
Pick a narrow, low-risk task, limit the agent's permissions, keep logs, require approval for high-impact steps and measure results before widening its role.