Which tasks should a small business hand to AI agents first?

When owners first hear “AI agents,” they picture something that runs the business while they sleep. Then they try to build that, it goes sideways, and they decide agents are hype. The ones that actually pay off are much less exciting. They sort the inbox before you sit down, tell you who's due a follow-up, and write the report you'd otherwise put together on Friday afternoon.

I spent more than 25 years in operations, most of it as a COO, and the rule I used for delegating to people works for delegating to AI too: hand over the work that's frequent, clear and checkable first. Judgment calls come later, once trust is earned.

What is an AI agent, in plain terms?

An AI agent is an AI assistant set up to do one job the same way every time. You don't need code. In Claude or ChatGPT, every agent is built from the same four parts:

• A project, which is one home for one job. Think of it as the agent's desk.
• Instructions it reads before every request. That's its job description.
• Files: your price list, FAQ, templates and a few past examples. That's its training binder.
• A schedule, for recurring work. A saved task that runs daily or weekly and leaves the result for you to review.

Connect the apps it needs (email, calendar, a shared drive) and you have an agent. The schedule is where most of the savings come from, because you stop having to remember to ask.

How do you know if a task is a good job for an AI agent?

Run every candidate task through four questions. The best first agents get four yeses.

1. Does it come back on a schedule? Daily, weekly or every time a new client signs. One-off projects don't need an agent; a good prompt will do.
2. Could you write the rules down? If you can explain how to do it to a new assistant in a page, an agent can follow the same page.
3. Does it take at least 20 minutes? An agent that saves 30 minutes but takes 25 to check isn't worth building.
4. Is it low-risk, or reviewed before it goes out? Summaries, reports, sorting and drafts are safe. Anything that reaches a customer, a bank account or an employee stays a draft that you approve.

Most owners are surprised by the answer. The task they hate most is rarely the one that saves the most hours.

Which AI agents should a small business build first?

These four pass the test in almost every business I've looked at. The savings below assume a typical business of 3 to 20 people, price each hour at what the role that usually does the job costs (U.S. median pay plus benefits), and already subtract review time.

• Inbox triage. Every weekday morning it reads the last day's email, sorts it into “needs you today,” “can wait” and “ignore,” and drafts replies to the routine ones. Forty-five minutes a day becomes fifteen. About 10 hours and $530 a month.
• Follow-up. Each morning it checks your pipeline spreadsheet or CRM, finds every lead, proposal and past client who is due a nudge, and drafts a short message that adds something useful instead of “just checking in.” About 6 hours and $318 a month, plus the deals that stop slipping through the cracks.
• Meeting notes. Paste a transcript or your rough notes and get decisions, owners, due dates and a follow-up email draft in a consistent format. About 6 hours and $318 a month.
• Weekly operations report. Every Friday it pulls the numbers and notes you point it to and writes a one-page summary of what moved, what's stuck and what needs a decision. About 10 hours and $700 a month, and your team stops waiting on you for status.

Which three you start with depends on where your week goes. If email and meetings swallow your day, start with inbox triage, meeting notes and follow-up (about $1,170 a month). If you sell services, start with prospect research, follow-up and a proposal builder (about $1,010 a month). If you run a team, start with the weekly report, an SOP writer and a team Q&A agent that answers “how do we do this?” from your own documents (about $1,360 a month).

What does a working agent actually look like?

Here are instructions for a simple inbox triage agent. Paste them into a new project in Claude or ChatGPT, fill in the brackets, connect your email, and test it on a real morning's inbox before you put it on a schedule.

Notice the last line. Every good agent has a “never” list. It's the difference between an assistant you trust and one you're nervous about.

How much money can AI agents save a small business?

Here's the formula I use, with nothing hidden:

Hours saved per month = (minutes the task takes by hand − minutes to review the agent's work) × times per month ÷ 60

Monthly savings = hours saved × the hourly cost of whoever would normally do it

Net savings = monthly savings − what the AI tool costs you

Inbox triage, for example: 45 minutes a day by hand, 15 minutes to review the agent's summary and drafts, so 30 minutes saved × 20 workdays = 10 hours a month. At $53 an hour, roughly what an executive assistant costs with benefits, that's $530 a month. One $20-a-month plan can run several agents for one person.

Two things keep these numbers honest. First, review time is always subtracted. Second, revenue gains (getting paid sooner, winning one more proposal) are real but harder to promise, so leave them out of the math and treat them as a bonus. Then time the task by hand once before you build, and time your review after two weeks. That's the number to quote to your partner, your boss or your client.

What shouldn't you hand to an AI agent?

• Anything that sends, pays or deletes on its own. Agents can prepare an invoice reminder, a refund reply or a payroll summary. A person sends, pays and approves.
• Relationship moments. A tough conversation with a client, a referral ask to a friend, a reply to an angry review. Let AI draft a starting point if it helps, then make it yours.
• Private data you haven't cleared. Check your plan's data settings, and don't upload client records, health information or anything covered by a confidentiality agreement until you know where it goes.
• Work with no rules yet. If nobody in the business can explain how a task should be done, write that down first. An agent will only automate the confusion.

A useful way to think about it: agents earn independence in stages. Every agent starts draft-only. After a month of drafts you rarely change, move to light review. Only internal, low-risk work (summaries, reports, sorting) ever gets to spot-checks. Anything customer-facing or money-related stays reviewed. That isn't a limit of the technology. It's good management.

Is an AI agent the same as automation like Zapier?

No, and you'll probably want both. Traditional automation moves data along fixed rules: when a form comes in, add a row and send a confirmation. It's perfect for high-volume, predictable steps. An AI agent reads and writes the way a person would, so it handles work that varies, like deciding which emails matter, summarizing a messy meeting or drafting a follow-up that fits the conversation. And plenty of work is best left as a person using AI as a helper, with no agent at all. Matching each workflow to the right level is most of the work in an AI rollout for a company.

Three mistakes that sink a first AI agent

• Building ten agents in a weekend. You'll maintain none of them. Build three properly, one per week, and add a rule to the instructions every time you fix the same thing twice.
• Vague instructions and no files. “Help with email” fails. “Sort into three groups, flag these senders, draft replies under 80 words using this FAQ” works.
• Trusting a green checkmark. A scheduled task that “completed” didn't necessarily do the job well. Read every output for the first few weeks.

Frequently asked questions

Rolling AI out across a whole team rather than just your own week? I help companies pick the right workflows, the right tools and the right builders, and get people to actually use them. Book a free discovery call or read more about AI implementation for companies.