AI Agent Examples: What Agents Really Do at Work, and Where They Fail

Read enough lists of AI agent examples and you’ll notice something odd: many of them are just chatbots with a new label. Gartner even has a name for it, “agent washing”, and estimates that only about 130 of the thousands of vendors selling agentic AI are the real thing.

So here’s a simple test to keep in mind as you read. Did it only say something, or did it actually do something?

01AI agent vs chatbot: the actual difference

Ask a chatbot where your order is and it might tell you how to track an order. Ask an agent and it looks up the order, sees that it’s stuck at the depot, raises a query with the courier, and tells you what it did.

That last row is why agents need more care. The line between the two is also blurry, and that’s fine. A support bot that can issue a refund has started to become an agent, whatever its maker calls it.

02Agents you can look at today

Fin, by Intercom, handles customer support. It answers from a company’s knowledge base and can take actions such as processing refunds and changing account details. Anthropic’s case study gives an average resolution rate of 51% out of the box and up to 86% in the best cases. Look at the gap between those two numbers. Much of the result comes from the setup work, and that’s worth remembering when you see a headline figure.

GitHub’s Copilot coding agent writes code. You give it a task, and it researches the repository, plans the change, edits files and runs tests in the background. Then a developer reviews the work. GitHub’s own documentation says logs don’t replace your own review and testing, and that’s a healthy attitude for any agent.

Swiggy lets assistants place orders. Since January 2026, people can order food, shop on Instamart and book tables from inside AI assistants such as ChatGPT and Claude. Swiggy didn’t build an agent here. It built the connection that lets other people’s assistants act for a customer, using an MCP server. Expect more companies to do the same.

03Everyday jobs an agent can take on

The famous examples can make this feel out of reach, so let’s bring it closer to home. The businesses below are made up, but the jobs aren’t.

The enquiry that gets researched before anyone calls

A property dealer in Mohali gets 40 website enquiries on a Sunday. By Monday morning, an agent has checked each one against the CRM, spotted the repeat enquiries, noted the budget and area, and sorted the list so the serious buyers sit at the top. The salespeople still make every call, but they start the day knowing who to ring first.

The documents nobody wants to chase

An immigration consultant’s staff spend hours reminding clients about missing passports and bank statements. An agent reads each file, works out what’s missing against the checklist, and drafts a WhatsApp message to each client. Someone glances through the drafts and approves them in one go. When a document arrives, the agent files it and updates the checklist.

The invoice that doesn’t match

A distributor receives supplier invoices as PDFs. An agent reads each one and compares it with the purchase order. If everything matches, the entry goes into the accounting system, ready for approval. If a supplier has charged ₹4,200 instead of ₹2,400, it goes to a person with both documents side by side. Most of the value is in catching that second kind.

The Monday report

Every week, someone at a small online store opens four tools and pastes numbers into a sheet. An agent can pull sales, ad spend and support tickets, write three lines on what changed, and send it before the 10 o’clock meeting. If something goes wrong, the worst case is a bad report, and that’s what makes it a good first project.

The cancelled appointment

A dental clinic has a cancellation at 11. An agent checks the waiting list, messages the next patient who wanted an earlier slot, and rebooks the calendar when they say yes. No one had to spend the morning on the phone, and the chair doesn’t sit empty.

The support queue that sorts itself

A software company wakes up to 200 tickets. An agent reads them, tags the bug reports, merges the duplicates, answers the password resets, and flags the customer threatening to cancel for a senior person. Knowing what to leave alone is a big part of this job.

04Two public lessons in how this goes wrong

Klarna announced in 2024 that its AI assistant was handling two-thirds of customer service chats, the work of 700 people. By 2025, its CEO was saying that focusing too much on cost had lowered quality, and the company began hiring people for support again. The assistant didn’t vanish. The idea of removing the humans did.

Then there’s Air Canada, whose website chatbot told a grieving customer that a bereavement fare could be claimed after travelling. That was wrong. The airline argued in a Canadian tribunal that it wasn’t responsible for what the chatbot said, and it lost. That was a chatbot, not an agent, but the message carries over: customers and courts are likely to see what your AI does as something you did.

05When an agent is the wrong tool

Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, because of rising costs, unclear value or weak risk controls. Plenty of those were probably never a good fit.

If a process runs the same way every time, ordinary automation will do it faster, cheaper and with no surprises. “When a form is submitted, add a row and send an email” needs no AI at all. Agents earn their cost when the input is messy and some judgement is needed, such as reading a free-form email and working out what the person wants. They also cost more to run than people expect, as our guide to AI development cost explains.

06How to pick your first agent

Look for a job with these four qualities: it happens often, it crosses more than one tool, someone can easily check the result, and a mistake is cheap to undo. The Monday report passes all four. Anything that moves money on its own doesn’t, at least not yet.

Give it one job, the fewest permissions that job needs, and a person who approves anything that sends, pays or deletes. Then watch it for a few weeks before you widen what it’s allowed to do.

Think about your own week. Which task do you dread because it means opening four tabs and copying things between them? Tell us about it using the form below and we’ll let you know whether an agent is a good fit and what it might cost, or read more about our AI development services.