AI Agents for Small Business Operations: Which Jobs They Can Take Today
- An AI agent can take narrow, checkable jobs today: sorting an inbox, reading supplier documents into drafts, searching for suppliers. Keep a person on anything that pays money, promises a customer something or changes your records.
- The best agent in a 2025 benchmark of office tasks finished 30.3% of them unaided, and admin and finance tasks scored among the lowest. Hand over the parts of a job you can check in a minute, and keep the rest.
- A Canadian tribunal held an airline liable for what its website chatbot told a customer. Check an agent's answers against your policies before you let it reply to customers on its own.
- Give an agent the fewest permissions the job needs. An agent that only drafts can't send the wrong thing to a customer.

In a 2025 Carnegie Mellon benchmark of AI agents doing office work, the best one finished 30.3% of its tasks without help. That number answers the question of what AI agents for small business operations can take on today: small, well-defined jobs whose output a person can check before anything leaves the building.
Your software vendors have started selling "agents". Below are six claims you'll hear in those pitches, each tested against a named source: a research benchmark, a shop run by an agent for a month, an analyst forecast, a tribunal ruling and a security standard. The last section lists the jobs that survive the test.
Myth 1: An AI Agent Can Run a Whole Job on Its Own
Not yet, going by the best evidence available. Researchers at Carnegie Mellon built TheAgentCompany, a fake small software company with its own chat, documents, project tracker and simulated colleagues, and set agents 175 tasks drawn from real job roles.
In the September 2025 version of the paper, the best model, Gemini 2.5 Pro, completed 30.3% of the tasks on its own. With partial credit for finished steps, it scored 39.3%.
Read that the other way round. Given a task an office worker would get on a normal Tuesday, the strongest agent tested left roughly seven in ten unfinished.
The failures were often mundane. On tasks inside the shared file system, a closable welcome popup stopped one agent cold, because it couldn't click the "x" to close it.
Myth 2: Admin and Finance Work Is the Easy Part to Hand Over
The benchmark found the opposite. Data science, admin and finance tasks had the lowest success rates of any department in TheAgentCompany, below software engineering.
The authors' explanation fits a small business exactly. Admin and finance tasks involve making spreadsheets, collecting information from several people and reading images scanned by employees. One finance task could only be finished by asking the finance director to clear up ambiguous questions. Most models also struggled in the chat tool where they had to ask colleagues for information, and did worse still in the shared office documents.
That is your accounts desk. The work looks simple because a person who knows the business fills the gaps without noticing. An agent has to be told every gap, and it tends to guess.
Myth 3: Anything Sold as an AI Agent Is One
Gartner doesn't think so. In a June 2025 forecast, it called the relabelling of chatbots, assistants and robotic process automation "agent washing" and estimated that only about 130 of the thousands of agentic AI vendors are real. It also predicted that over 40% of agentic AI projects will be canceled by the end of 2027.
The more useful line for an owner is this one: "Many use cases positioned as agentic today don't require agentic implementations."
If a job follows the same steps every time, such as copying a paid invoice's status from the bank feed into the job sheet, a plain automation rule does it more cheaply and fails in ways you can predict. Gartner's own advice is to use agents where decisions are needed and ordinary automation for routine workflows.
So ask the vendor one question: what does this tool decide that a rule couldn't?
Myth 4: An Agent Can Be Trusted With Money and Customers
Anthropic and Andon Labs tested this directly. For about a month in 2025, an instance of Claude Sonnet 3.7 ran a small office shop: a fridge, some baskets and a self-checkout iPad. It chose stock, set prices, ordered from a wholesaler and answered customers over Slack.
It was good at one thing an operations person would recognise. It found suppliers fast: asked to stock a Dutch chocolate milk, it quickly turned up two sellers of Dutch products.

The money side went badly. Offered $100 for a six-pack of Irn-Bru that costs $15 online, it said it would keep the request in mind. For a time it told customers to pay into a Venmo account it had hallucinated. It priced metal cubes below cost, handed out discount codes when asked nicely and gave some items away. When an employee pointed out that a 25% staff discount made little sense when "99% of your customers are Anthropic employees", it agreed, announced it would drop the codes, then offered them again within days.
The researchers' verdict was that the shop "did not succeed at making money". They also think better tools and instructions could fix much of it, and that's fair.
Which of your jobs could an agent prepare?
- INBOX SORTING??
- INVOICE ENTRY??
- SUPPLIER SEARCH??
- CUSTOMER REPLIES??
- REFUNDS??
Send us the jobs that eat your week. We'll mark which ones an agent can draft for a person to approve, and which need plain software first.
MARK_MY_JOBS→Myth 5: A Business Isn't Liable for What Its Bot Says
A Canadian tribunal tested this in 2024. In Moffatt v. Air Canada, the airline's website chatbot told a grieving customer they could apply for a bereavement fare after travelling. The airline's actual policy didn't allow that.
Air Canada argued it couldn't be held liable for information from the chatbot. The tribunal member called that "a remarkable submission" and wrote: "It should be obvious to Air Canada that it is responsible for all the information on its website."
The airline was ordered to pay C$812.02, made up of C$650.88 in damages plus interest and tribunal fees. A customer who relies on your agent's answer is relying on you.
Before an agent replies to customers unsupervised, read a few weeks of its draft answers against your actual policies. Anything about refunds, prices or deadlines should go out under a person's name.
Myth 6: An Agent Works Better With Access to Everything
More access makes an agent more capable and more dangerous at the same rate. The OWASP security project lists excessive agency among the top risks for AI applications, and names its causes as too much functionality, too many permissions and too much autonomy.
OWASP's own example is a small one: a tool that summarises email only needs to read the mailbox, so it shouldn't be able to send from it. For anything high-impact, it recommends a person approve the action before it happens.
Applied to a small business, that means separate logins for the agent, read-only by default and write access only to the one place its job needs. Our post on vibe coding internal tools covers what happened when a coding agent could reach a company's live database.
Which Operations Jobs AI Agents Can Take Today
The jobs that survive these six tests share one feature. The agent prepares the work, and a person commits it in under a minute.
Sorting what comes in. Tagging a shared inbox by customer, urgency and type, then routing each message to the right person. A wrong tag costs a few seconds to fix.
Reading documents into drafts. Pulling supplier name, amounts and due dates from bills and delivery notes into a draft entry that someone approves.
Drafting replies. A first answer to a customer question, written from your own policy pages and sent by a person.
Searching. Finding suppliers and comparing their prices on a list of parts. This was the shop agent's best skill.
What stays with a person, for now: paying anyone, setting a price, promising a customer something, and changing a record other people depend on.
Our view: agents work on top of a business that already has its rules written down. If the job lives in one person's head or in a spreadsheet only they understand (the first sign in our post on outgrown spreadsheets), an agent will guess where that person would know. Put the rules in a spec and build them into software first, then give the agent a narrow job inside it.
Which job on your list could an agent draft tomorrow, and who would approve it?
FAQs
Q.What is the difference between an AI agent and a chatbot?
A chatbot answers questions in a conversation. An AI agent can also take actions in other software, such as filing a ticket, updating a record or placing an order, which is why its permissions matter more.
Q.Can an AI agent use the software my business already has?
It can if that software has an API the agent can call, or if the agent can operate its screens. Working through screens is the weaker route: in TheAgentCompany, browsing web pages was often the biggest obstacle.
Q.Will AI agents replace my office staff?
Gartner predicts at least 15% of day-to-day work decisions will be made autonomously by AI agents by 2028, up from 0% in 2024. That leaves the large majority of decisions with people, so plan for agents that take tasks off a role.
Q.How much does it cost to run an AI agent on an office task?
In TheAgentCompany, the best model cost about $4.20 in model fees per task attempt, averaging 27 steps. That leaves out setup, connecting the agent to your software and the time a person spends checking its work.
References
- Frank F. Xu et al., TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, arXiv 2412.14161, version 3, 10 September 2025
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", press release, 25 June 2025
- Anthropic, "Project Vend: Can Claude run a small shop? (And why does that matter?)", 27 June 2025
- Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency


