AI agent examples: 12 real jobs, and 5 to keep for yourself

An AI agent is software that is given a goal and works out the steps itself, using your tools (email, a CRM, a calendar, your accounting package) to get there. Good examples for a business are researching a new lead before you reply, chasing overdue invoices, matching supplier bills to orders and sorting maintenance requests. The ones that work share a pattern: one clear job, a short list of tools, and a person who approves anything that matters. Below are twelve real jobs agents do well today, and five they still should not be trusted with.

What is an AI agent, in plain English?

Anthropic, the company behind Claude, draws the line neatly. A workflow is a system where AI and tools follow "predefined code paths". An agent is one where the AI decides its own process and which tools to use, "maintaining control over how they accomplish tasks".

In practice, the difference looks like this:

If a job is the same every time, you do not need an agent. A simple automation is cheaper and more predictable, and Anthropic's own advice is to use the simplest solution that works. Agents earn their place when every case is a little different.

12 AI agent examples, by job

Each example lists the business it suits, what the agent does, and the checkpoint where a person stays in charge. The tools named are examples; the same job can be built on whatever you already use.

Customers and sales

1. The lead researcher (agencies, consultants, B2B services). When an enquiry arrives, it reads the company's website, checks your CRM for past contact, scores the fit against your own criteria and drafts a first reply that mentions something specific about them. Checkpoint: you read the draft and press send. More on this in our AI lead generation service.

2. The support triager (online shops). It reads every new ticket, looks up the order, answers "where is my order?" with live tracking, drafts replies to the rest and routes refunds and complaints to a person. Checkpoint: refunds and complaints always reach a human. See AI customer service.

3. The review responder (hospitality, trades, clinics). It notices new reviews, drafts a reply in your voice that mentions the details the customer gave, and flags anything negative straight away. Checkpoint: you approve every reply, and write the difficult ones yourself.

Money

4. The invoice chaser (any business that invoices). Each morning it checks your accounting software for overdue invoices, sends the right reminder for each stage, and logs every message against the invoice. Checkpoint: anything disputed, or more than 60 days late, comes to you.

5. The supplier bill matcher (trades, restaurants, retailers). It reads supplier invoices from your inbox, matches each one to the purchase order and delivery note, flags price differences and prepares the bill for approval. Checkpoint: you approve payment. The agent never pays anything. This is the heart of AI document processing.

6. The receipt sorter (sole traders, freelancers, individuals). It gathers receipts from email and phone photos, reads the amount and supplier, categorises each one and files it ready for your tax return. Checkpoint: you or your accountant review the month before it is filed.

Operations

7. The maintenance coordinator (letting agents, property managers). It reads a tenant's repair request, asks for photos, checks who is responsible under the tenancy, books an approved contractor and keeps the tenant updated. Checkpoint: emergencies, and any job above your spending limit, go to a person.

8. The scheduler (clinics, consultants, busy individuals). It finds times that work across calendars, offers slots, handles rescheduling and sends reminders. Checkpoint: anything outside your rules, such as a weekend booking, waits for you.

9. The stock watcher (shops, cafés). It checks stock levels each day against how fast items are selling, and drafts a reorder to each supplier before you run out. Checkpoint: you approve each order.

10. The Friday report writer (owners and managers). It pulls figures from your sales, bookings and accounting tools, and writes a short plain-English summary of what changed and why. Checkpoint: none needed to start, because it only reads. Check the first few against your own numbers.

Growth and your own time

11. The tender finder (contractors, consultants). It searches public procurement sites every day for contracts that match your work, summarises the requirements and deadlines, and drafts a bid or no-bid note. Checkpoint: you decide whether to bid.

12. The inbox assistant (individuals, freelancers, founders). It sorts your email, drafts replies in your voice, pulls tasks and deadlines into your to-do list and proposes meeting times. Checkpoint: drafts only. Nothing is sent without you.

How much should an agent do on its own?

The checkpoints above are not an afterthought. They are the design. We think about autonomy as a ladder, and every new job starts on a low rung.

Diagram

How much to let an agent do on its own

1. Rung 1

   Suggests

   It reads and recommends. You do the work.

   For example: Ranks this week’s enquiries by how likely they are to buy.

2. Rung 2

   Drafts

   It prepares the work. Nothing leaves until you approve it.

   For example: Writes the reply to a complaint. You press send.

3. Rung 3

   Acts within limits

   It acts alone below a limit you set, and asks you above it.

   For example: Books a contractor for repairs under $300.

4. Rung 4

   Acts alone

   Only for low-risk jobs that are easy to undo.

   For example: Files supplier invoices into the right folder.

Anthropic learned this the expensive way. In 2025 it let a Claude model, nicknamed Claudius, run a small shop in its office: choosing stock, setting prices and talking to customers. Claudius found unusual suppliers well, but handed out discount codes freely, sold tungsten cubes below cost and, for a while, told customers to pay into an account it had made up. The shop lost money.

What fixed it in the second phase was not mainly a smarter model. Anthropic gave the agent a CRM and inventory records showing what it had paid for each item, made it follow procedures before quoting a price, and still "didn't give it access to a payment interface", so a human approved every purchase. Losing weeks largely stopped. Anthropic's summary was that "bureaucracy matters", which is a good motto for anyone building an agent.

5 jobs to keep for yourself, for now

1. Spending money without a limit. The shop experiment above is the clearest example. An agent can prepare a purchase or a payment; a person should release it, or the agent should work within a small, fixed limit.

2. Anything you cannot undo. In July 2025, during a public test by SaaStr founder Jason Lemkin, Replit's coding agent ran a command that wiped a live database while a code freeze was in place. Replit's chief executive called it "unacceptable and should never be possible", and the company promised automatic separation of test and live databases. The lesson for any business: agents work on copies or drafts, and deleting, cancelling or sending in bulk needs a person.

3. Decisions about people. Hiring, firing, disciplinary matters and credit decisions need human judgement, and in the UK and EU data protection law gives people rights when decisions with significant effects on them, such as rejecting a job application, are made by automated means alone. An agent can collect and summarise. A person decides.

4. Upset or vulnerable customers. Complaints, bereavement, safety worries and anything emotional need a person who can listen. Agents are trained to be agreeable, which Anthropic found made its shop agent easy for colleagues to talk into bad deals. Route these conversations to a human, with the agent's summary attached.

5. Big, vague goals. "Run my marketing" or "grow sales" has no finish line, so an agent cannot tell whether it is done or doing well. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls, and warned that many products are "agent washing": ordinary chatbots or automation tools relabelled. One job with a clear outcome beats a grand plan.

Are businesses really using AI agents yet?

Some are, and most are still trying them out. McKinsey's State of AI survey, published in November 2025, found that 23% of respondents' organisations were scaling an AI agent somewhere in the business and another 39% were experimenting with agents. Even among the leaders, use was narrow: in any single business function, no more than 10% of respondents said they were scaling agents.

Small firms look earlier still. A 2025 survey for Intuit QuickBooks put early adopters of agentic AI at about one in ten small business owners, with marketing, customer service and admin as the top jobs owners use AI for in general. If you have not started, you are in the majority. The businesses getting value now are the ones that picked one dull, frequent job and did it carefully.

How to start: a five-step checklist

1. Pick one job that happens every week, crosses two or three tools and has a clear finish line. Our guide on which tasks to automate first has a scoring method.
2. Write its job description: the goal, the steps a good employee would take, what "done" looks like and what must never happen.
3. Choose its rung on the ladder above, and set the limits in writing: amounts, recipients, hours.
4. Rehearse on last month's real cases before it goes live, and count how often it got each one right.
5. Go live with a log you can read. Review it weekly, and move the agent up a rung only when the log gives you a reason to.

If the job involves answering questions from your own policies, manuals or files, the agent will need a way to look things up reliably. That is what a RAG system does; we explain it in what RAG is, in plain English.

What we would do

For most businesses we would start with example 4 or 5, the invoice chaser or the supplier bill matcher. Both are frequent, both use data you already have, both save hours a month, and both are easy to keep on rung 2 or 3 while you build trust. After a month of reading the agent's log, you will know exactly what it gets right, where it hesitates, and which job to give it next.