AI agents for small business: 15 jobs you can hand over

Ask how many small businesses use AI and you get two very different answers. The U.S. Chamber of Commerce's 2025 report says 58% of small businesses use generative AI, up from 23% in 2023. The U.S. Census Bureau, asking whether a business used AI in the past two weeks, found only 17% to 20% of firms did between December 2025 and May 2026. Part of the gap is what counts: chatting with an assistant now and then, or AI doing real work inside the business. This guide is about the second kind. Here are 15 jobs where AI agents for small business can take work off your plate, what you should still approve for each, and the risk that comes with it.

First, what makes an AI agent different

A chatbot answers. An AI agent acts: you give it a goal, and it decides the steps, uses tools (your email, a browser, a spreadsheet, a database) and keeps going until the job is done or it needs you. If the difference is new to you, start with what an AI agent is and AI agent vs chatbot.

That ability to act is the whole point, and it's also the risk. The OWASP project, which publishes the best-known list of security risks for AI applications, calls it excessive agency: damage done because an agent had more functions, more permissions or more autonomy than its job needed. Its fixes are plain: give the agent only the tools it needs, the narrowest access that works, and a person to approve high-impact actions.

So for every job below, think in four levels of trust:

Most jobs should start one level lower than feels necessary and move up once you've seen a few weeks of good work. More on that in human-in-the-loop AI. And any agent that reads outside content (emails, web pages) can meet instructions planted for it, so read how prompt injection works before you hand over your inbox.

AI agents for small business: 15 jobs to hand over

Where crew bot does a job today, we say so and how. Where we don't mention it, the job is still worth doing with whichever tool fits.

1. Answer your website's visitors

What the agent does: answers questions on your site at any hour, from what your site and your team actually say, and takes the visitor's details when a person is needed.

What you approve: nothing per message, but you decide what it knows and when it hands over. Refunds, complaints and account changes go to a person.

The risk: a wrong answer is your answer. In 2024 a Canadian tribunal ordered Air Canada to pay after its website chatbot told a grieving customer he could claim a bereavement discount retroactively, after buying his tickets, which the airline's policy didn't allow. The airline argued the chatbot was responsible for its own actions; the tribunal said it was "still just a part of Air Canada's website".

In crew bot: yes. One line on your site adds a chat answered by a bot you pick for that site. It reads the site's pages, answers in the visitor's language, and says plainly when it doesn't know instead of inventing prices or promises. It hands over to your team when a visitor asks for a person or the matter needs one. The support bot has no tools, so nothing a visitor types can make it change anything. Full method in how to add an AI chat to your website.

2. Triage the inbox and draft replies

What the agent does: reads incoming email, sorts what needs you from what doesn't, drafts replies, files receipts and statements, and clears out newsletters you never read.

What you approve: every reply before it goes out, at least at first.

The risk: a confident reply to the wrong person, or an important email filed as noise. Keep "needs you" generous.

In crew bot: yes. It reads and sorts your Gmail, drafts the replies that need you, files the receipts and unsubscribes you from the noise. Nothing is sent until you say so.

3. Follow up with leads

What the agent does: replies to new enquiries quickly, asks the qualifying questions you'd ask, and schedules a follow-up when a lead goes quiet.

What you approve: the questions, the follow-up cadence, and any quote or discount.

The risk: pestering people, or promising what sales can't deliver. Cap the number of follow-ups and keep pricing out of its hands.

4. Send appointment reminders

What the agent does: reminds customers of bookings, handles "can I move it to Thursday?", and fills cancelled slots from a waiting list.

What you approve: the rules: how far ahead, how many reminders, who gets offered a freed slot.

The risk: double bookings when the agent and your calendar disagree. Let one system own the calendar.

5. Chase overdue invoices

What the agent does: checks which invoices are past due, drafts a polite reminder for each, and escalates the tone on a schedule you set.

What you approve: the wording, and which clients it may chase on its own.

The risk: chasing a client who paid yesterday, or nudging your biggest account in the wrong tone. Have it check payments received before every run.

In crew bot: yes, as a routine: an overdue-invoice follow-up every Monday at 09:00, for example. It asks before it sends an email unless you've told it to go ahead.

6. Do the bookkeeping admin

What the agent does: collects receipts and statements, exports invoice lists, and moves figures between the tools you already use.

What you approve: anything that goes into the books as final, and anything sent to your accountant.

The risk: a misread amount that nobody notices until year end. Ask for a short list of what it changed, every time.

In crew bot: yes. It files receipts and statements from your inbox, and on your Mac it can work desktop apps like Excel and your accounting software with mouse and keyboard, while you watch.

7. Handle new orders and other events

What the agent does: when something happens in another system (a new order, a form filled in, a payment received), it picks up the work: confirms with the customer, updates a record, tells the right person.

What you approve: anything involving money, like marking an order paid or issuing a refund.

The risk: acting twice on the same event, or on a fake one. Ask your tool how it avoids duplicates.

In crew bot: yes. Webhooks let your other systems hand it work: a new order from your online shop, say, that the crew then confirms with the customer.

8. Write you a morning brief

What the agent does: before you're at your desk, sums up what came in overnight, what's due today and what needs a decision.

What you approve: nothing; it only reads.

The risk: low. The main failure is a brief so long you stop reading it.

In crew bot: yes. Routines run on a schedule (a morning brief at 07:00, say) and the results come to your phone or Telegram.

9. Answer questions from your own data

What the agent does: you ask "how many repeat customers did we have in September?" in plain words, and it queries your database or spreadsheet and gives you the number.

What you approve: the access. Read-only is enough for questions.

The risk: a wrong query gives a confident wrong number. Ask to see how it got the figure, at least for numbers you'll act on.

In crew bot: yes. It connects to databases like Postgres and MySQL and to web apps you sign it in to, and answers from your own data. Every query it runs stays on record, so you can check how it got the number.

10. Research suppliers, competitors and prices

What the agent does: searches, reads and compares, then hands you a short summary with links.

What you approve: nothing to read; any decision you take from it.

The risk: made-up facts. In 2023 a U.S. judge fined a law firm $5,000 after its lawyers filed a brief citing court cases that ChatGPT had invented. Always ask for sources and open the ones that matter.

In crew bot: yes. The crew has a research specialist and every plan includes web search; a skeptic attacks every finding before it reaches you.

11. Draft social media posts

What the agent does: turns your news, offers and photos into post drafts for each channel, and suggests when to publish.

What you approve: every post. Your public voice is not a job to fully automate.

The risk: off-brand posts, wrong claims, and worse. The U.S. FTC's 2024 rule bans fake reviews and testimonials, naming AI-generated fake reviews. Never let an agent write reviews of your business.

12. Fill in portals and old desktop apps

What the agent does: the copy-and-paste work between systems that don't talk to each other: supplier portals, government sites, the old app with no API.

What you approve: submissions, payments and anything signed in your name.

The risk: a slip in the wrong field, done at machine speed. Watch the first few runs.

In crew bot: yes. It uses a computer the way you do: mouse, keyboard, any app on your Mac including the old ones, and a real browser signed in to your portals. You watch every move live and can take over at any time. You type passwords into sites yourself.

13. Watch your website

What the agent does: checks that your site is up, its forms work, its security certificate isn't about to expire, and pages aren't broken, then tells you before a customer does.

What you approve: nothing to watch; any fix.

The risk: alert fatigue. Agree on what's worth a message at 2 a.m. and what can wait for the morning brief.

14. Check your servers and network

What the agent does: signs in to servers and routers, reads logs and settings, finds what's wrong and proposes the fix.

What you approve: every change. Reading is cheap; changing is where outages come from.

The risk: a "fix" that takes everything down, or a leaked password. See AI agents for DevOps for the safe way to do this.

In crew bot: yes. It signs in to your servers, routers and databases, reads what's wrong and tells you the fix, or makes it once you approve. Logins stay in a vault no model ever sees, and SSH uses five-minute certificates.

15. Put website and app changes live

What the agent does: builds the new version, starts it, checks it works, and switches visitors over.

What you approve: the release.

The risk: a broken release in front of customers. A good process checks the new version answers before switching and keeps the old one ready.

In crew bot: yes. It deploys to your own server: builds, starts, checks the app answers, then switches over with HTTPS, and keeps the release before so you can roll back.

The 15 jobs at a glance

A worked example: chasing overdue invoices

Here's how job 5 plays out for a small agency with a dozen active clients.

1. Monday, 09:00. A scheduled routine starts. The agent opens the invoice list and finds four invoices more than a week late.
2. It checks before it chases. One of the four was paid on Friday; the accounting tool hadn't caught up. It drops that one.
3. It drafts. Three reminders, each with the invoice number, amount and a link to pay. One client is 45 days late, so that draft is firmer, as your rules say.
4. It asks. You get one message on your phone: three drafts, Approve or Review. You change one line on the firmest one and approve.
5. It sends and records. The emails go out, and the job keeps a record of what it read, what it drafted and what you approved.
6. It remembers. You tell it "never chase the hospital group without asking me first". Next Monday, that client's reminder waits for you while the others go.

Everything before step 4 only reads. After a month of approving the same drafts unchanged, you might let the polite first reminder go on its own and keep approval for the firm ones.

How to choose your first job

Pick the job that scores yes on most of these:

• It happens at least weekly
• It follows rules you could write on one page
• You can check the result in under a minute
• A mistake is embarrassing, not expensive, and easy to undo
• The information it needs is somewhere an agent can reach
• You'd be glad never to do it again

Inbox triage, the morning brief and website chat usually win. Servers and deploys pay off too, but deserve the most care, so they're rarely the first job.

What AI agents cost

Prices come in three shapes:

• A subscription, per user or per workspace, often with usage limits.
• Pay for usage: the AI work itself, measured in what the job consumed.
• Pay per outcome: for example, Intercom's Fin support agent charges $0.99 per outcome, such as a resolved conversation.

The hidden cost is your time: setting it up, writing the rules, and reviewing drafts in the early weeks. Budget for that, then compare against the hours the job takes you now.

For crew bot: Free to start with no card, then $19, $49 or $149 a month per workspace, with every plan getting the whole crew. Every job has a budget, fifty US cents by default, and the crew stops and asks if it needs more. Details on the pricing page.

FAQ

What is the best AI agent for a small business?

The one that does your first job well and lets you control it. A support chat tool is enough if website questions are the problem; a general crew helps when the jobs span email, apps and servers. We compare options in the best AI agents for small business.

How much does an AI agent cost for a small business?

Plans range from free tiers to monthly subscriptions, usage-based billing and per-outcome pricing. Add your own setup and review time. Start with one job and compare its cost with the hours it saves.

Will AI agents replace my staff?

They replace tasks, not judgement. The jobs above still need someone to set the rules, approve what matters and handle the cases the agent hands back. Most small teams use agents to clear the backlog nobody had time for.

Do I need technical skills to use AI agents?

Not for most jobs on this list. You describe the work in plain words, and good tools ask before they act. Servers, databases and deploys go faster if someone on the team understands them, because they'll be the one approving changes.

Is it safe to give an AI agent my email and accounts?

It can be, with narrow access, approval before anything is sent or changed, encrypted logins and a record of every step. Avoid any tool that wants your passwords pasted into a chat or acts without asking from day one.