AI Agents for Small Businesses: Ten Jobs They Can Take On

An AI agent is useful to a small business when it takes a job that is repetitive, happens often and can be checked quickly: sorting the inbox, drafting quotes and replies, preparing receipts for the bookkeeper, proposing appointment times, drafting social posts and review replies, flagging low stock, writing the weekly summary and chasing unpaid invoices. In every one of those, the agent does the gathering and the first draft, and a person approves anything that goes to a customer, moves money or cannot be undone. Start with one job, run it in draft-only mode for a couple of weeks, and widen it only when the drafts need few corrections.

If you want a way to sort your own list of jobs into “automate”, “agent drafts, person decides” and “keep with a person”, that method is in which tasks can AI agents automate. This page is the shortlist: ten jobs that fit most small businesses, whatever they sell.

How much setting up: three levels

• Light: an afternoon. Turn on an AI feature in a tool you already use, such as your email, accounting or booking software, or give a chat assistant a written procedure and paste work into it.
• Medium: a day or two. Connect two tools with an automation tool that has an AI step, so that something arriving in one produces a draft in the other. Building an agent in n8n shows what that involves.
• Heavy: a week or more, usually with help. A custom agent with its own tools, access to your systems and a way to log what it did.

Most of the ten below work at the light or medium level. None needs the heavy one to start.

The ten jobs

1. Inbox triage

• The agent: labels each new email as an enquiry, an order, a complaint, a supplier, an invoice or noise, and writes a one-line summary of the ones that need you.
• A person: reads the “needs you” list and spot-checks what was filed as noise, daily at first.
• Setup: light to medium. Keep it to labelling; an agent that reads email should not also be able to send it without a person, because an email can carry instructions aimed at the agent, as indirect prompt injection explains.

2. Quotes and estimates

• The agent: turns an enquiry and your price list into a draft quote in your usual format, and lists anything it had to assume.
• A person: checks every figure and every assumption before it is sent. A wrong quote is a promise.
• Setup: light if you paste the enquiry and the price list in; medium if the draft is built straight from the enquiry email.

3. Bookkeeping preparation

• The agent: reads receipts and supplier invoices, pulls out the date, supplier, amount and tax, suggests a category, and lists the ones that do not match a bank line.
• A person: you or your bookkeeper approves the categories and resolves the mismatches. Nothing is posted without that.
• Setup: light where your accounting software already reads receipts; medium otherwise.

4. Customer replies

• The agent: drafts answers to the questions you get every week (opening hours, delivery times, how to return something) from a short document of approved answers.
• A person: sends each draft, editing where needed. Once the routine answers need no edits for a few weeks, you may let those go out on their own and keep everything else in draft.
• Setup: light. Writing the approved-answers document is most of the work, and it is useful to staff as well.

5. Scheduling and bookings

• The agent: reads a request, checks the calendar, proposes two or three times, and drafts the confirmation and the reminder.
• A person: confirms anything unusual, such as a long job, a new customer or a change to someone else’s diary.
• Setup: light if your booking tool has this built in; medium if it has to read email and a calendar together.

6. Social media drafts

• The agent: drafts a week of posts from what happened (finished jobs, new stock, an event) in the tone you give it.
• A person: approves each post and checks that any photo or claim is yours to use.
• Setup: light. Keep a short style note with examples of posts you liked.

7. Stock checks

• The agent: compares stock levels with your minimums each morning, lists what is low, and drafts a reorder for each supplier.
• A person: places the order. The agent’s list is advice; the purchase is yours.
• Setup: medium. It needs read access to wherever stock is counted, which is easy with a stock system and fiddly with a spreadsheet nobody keeps up to date.

8. Review replies

• The agent: drafts a reply to each new review, thanking the good ones by detail and answering the bad ones calmly with a way to get in touch.
• A person: approves every reply to a negative review. A public argument is hard to take back.
• Setup: light.

9. The weekly summary

• The agent: every Friday, writes one page from your sales, bookings and jobs lists: what came in, what went out, what is stuck.
• A person: reads it, and checks one or two numbers against the source each week so that you would notice if it drifted.
• Setup: light to medium, depending on how many places the numbers live.

10. Chasing unpaid invoices

• The agent: lists invoices past their terms and drafts a polite reminder for each, firmer for the second one.
• A person: sends them, and holds back any customer you are already talking to. The agent does not know about the phone call last Tuesday.
• Setup: light where your accounting software can list overdue invoices; medium otherwise.

Data protection basics

Most of these jobs involve customers’ names, emails and sometimes more. The rules do not change because an agent is doing the work. In the UK, GOV.UK’s guide to data protection↗ (opens in a new tab) says you must tell people who you are and how you will use their information, keep it secure, accurate and up to date, and answer when someone asks what you hold about them.

• Know where the data goes. Read the AI tool’s terms for whether your inputs are stored or used for training, and pick a business plan or setting that says they are not.
• Give the agent only what the job needs. The review-reply agent does not need your customer list.
• Keep a person on decisions about people. The ICO’s guidance on AI and data protection↗ (opens in a new tab) covers the extra rules for decisions made solely by automated means that significantly affect someone; it also says the guidance is under review after the Data (Use and Access) Act.
• In the US, the SBA’s advice on AI for small business↗ (opens in a new tab) says to avoid feeding sensitive data into AI tools and to make sure a person assesses the messages AI generates.
• Secure the accounts the agent uses. The NCSC’s small organisations guide to cyber security↗ (opens in a new tab), where the old Small Business Guide address now leads, covers backups, device protection, account security, scams and malware; an agent’s login is one more account to protect.

Starting small

1. Pick one job from the ten, the one that annoys you most often. Frequent beats important for a first try.
2. Write down the procedure as you would for a new member of staff: what comes in, what good looks like, what to never do.
3. Run it in draft-only mode for two weeks. The agent prepares; a person sends, posts or orders.
4. Count corrections. If most drafts go out unchanged, let the routine cases run on their own and keep the rest in draft. If they do not, fix the procedure before blaming the tool.
5. Only then add the second job.

Written down, that rule is a small version of the governance larger companies do. AI agent governance for small teams covers the rest: who owns each agent, what it may touch, and how you would turn it off.

Keeping the jobs in one place

Once an agent is drafting, something has to hold the list of what is waiting for a person. On a fenbs board that is a task in To Do or Next Up: “Reorder from two suppliers”, with the list in the note, or “Three review replies to approve”. The assistant connects over MCP with a token you issue under Settings, with a name, only the scopes it needs (read, write, comment) and an optional expiry, and each task it adds or comments on is recorded under its name. Moving a task between lanes is its own permission, separate from adding and editing, so the assistant can be given a role that files and comments while staff and you move the work along. When you want the assistant to take a job without asking first, you pre-approve that task with a plan and limits, and it picks it up from there.

fenbs is the jobs list, not the tools around it: it does not send email, keep a calendar, hold stock or do invoicing, and it has no due dates. A task’s note can link to the email, the booking or the invoice. Why a small team is better off with that little is in a simple project management tool for small teams.