How to use AI agents: a practical guide for small business owners

To use an AI agent, give it one recurring job you already understand, connect it to the tools where that job happens, describe the result you want, and decide which actions it takes on its own and which wait for your OK, then widen its job as its work proves out.

Key Takeaways

  1. To use an AI agent, hand it a responsibility the way you'd hand one to a new hire: one recurring job, the tools that job lives in, a clear description of a good result, and a rule for which actions wait for your OK.
  2. The best first jobs are ones you already do on a schedule and can check at a glance, like a Monday summary of your numbers, chasing overdue invoices, or a morning briefing from your email and calendar.
  3. Draw the approval line at actions that are hard to undo, such as sending an email, publishing a page, or spending money. Let the agent handle the reading, research, and drafting around them.
  4. In the first accounts from owners, most early trouble came from setup rather than the agent's intelligence: the wrong app connected, a missing permission, or a goal framed too narrowly.
  5. Many agents built for owners have a free plan with usage limits or start around $20 a month. In the first month, set aside time to read the agent's work, which shrinks as you hand over more.
On this page

How to use AI agents, in short

You use an AI agent by giving it a job, not a prompt. Pick one piece of recurring work, connect the agent to the apps where that work happens, tell it what a good result looks like, and decide which actions it may take without asking you. Then read its first results closely and give it more room as its work proves out.

That's what separates an agent from a chatbot. A chatbot answers the message in front of it and stops. An agent keeps working between conversations: it notices what needs doing next, uses your connected tools to do it, and comes back with results or with the decisions it needs from you. OpenAI's documentation for its dots puts it plainly: the agent "tracks progress, works out what needs to happen next, and follows through as things change" (OpenAI).

This became much more practical for small businesses in 2026. Anthropic released Claude for Small Business in May, which Jeff Barnes, who tested it on a six-person firm, describes as a switch you turn on inside Claude Cowork that connects QuickBooks, PayPal, HubSpot, Google Workspace, and similar tools (Digital Evolution Marketing Group). On September 29, OpenAI launched dots and Meta launched Muse for Small Business, two agents that stay on in the background and work across your apps (SLIDEFACTORY, Meta). You no longer need to build anything to put an agent to work.

The early accounts agree on where the effort goes. Mark Nguyen, who runs a Portland agency, spent his first day with a dot on real client work and summed it up this way: "The hardest parts of my day were integration, permissions, and knowing what had actually finished. The model's intelligence wasn't the main obstacle" (SLIDEFACTORY). This guide is mostly about those practical parts: choosing the job, giving the agent what it needs, deciding where it asks first, and reviewing its work.

Should you use a ready-made agent or build your own?

Most owners should start with a ready-made agent, because the hard part is choosing and supervising the work, not assembling the software. Ready-made agents such as Kite, OpenAI's dots, Meta's Muse, Claude Cowork, and Lindy come with the reasoning, memory, scheduling, and approval controls already in place. You connect your apps and start describing work.

Building your own makes sense when you have a technical bent and a job no product covers. Tools like n8n and Zapier let you assemble an agent from a model, written instructions, and the apps it may use (n8n, Zapier). Yousef Mathis, who runs a one-person software business, built his support agent in an afternoon, and he's clear that "building it was the cheap part" (Yousef Mathis on Medium). The ongoing work of deciding what it handles and checking what it does is the same either way.

It's also worth knowing when you don't need an agent at all. If a job runs the same steps every single time, such as copying each new form entry into your CRM, a fixed automation is simpler and more predictable. Anthropic's engineering guidance recommends "finding the simplest solution possible," and saves agents for work where the steps depend on what turns up along the way (Anthropic). For a broader look at which agents fit which small-business jobs, see The best AI agents for small business in 2026, sorted by the job.

Which job should you give an AI agent first?

Give your first agent a recurring job you already do yourself, where you know what a good result looks like and can check it quickly. OpenAI's own advice for new dot owners is to "start with work you regularly need to check or update," then review the first result and correct it before expanding the responsibility (OpenAI).

Ali Abouelatta, who runs Lazyweb and has tried many of these agents, uses a two-part test before he delegates anything: "how clearly do I know what I want, and how much would I care which of the plausible outcomes I get?" He'd hand off reordering eggs, but he'd keep shopping for watches and flights (First 1000). Jobs where you know exactly what you want, or where any reasonable outcome is fine, are the easy ones to hand over.

Jeff Barnes's three-week test on a professional services firm with about $1.2 million in revenue shows what a good first job looks like in practice. He found that four of Claude for Small Business's 15 workflows changed the owner's week, and he recommends starting with two of them (Digital Evolution Marketing Group):

  • Invoice Chaser. A $6,800 invoice was 47 days overdue after three unanswered follow-ups. The agent pulled it from PayPal and queued three escalation emails. Barnes reviewed them in four minutes, sent them, and the invoice was paid within 72 hours.
  • Monday Business Pulse. The agent reads QuickBooks and HubSpot on a schedule and writes a plain-English summary of cash, sales, pipeline, and the week's commitments. It replaced about forty minutes of clicking between tools every Monday.

His advice on pacing is just as useful: run those two for a month before adding more, because "you'll end up reviewing AI outputs instead of running your business" if you switch on everything at once.

Other strong first jobs share the same shape. Each one runs on a rhythm you already know, and each produces something you can check in a minute or two:

  • A morning briefing that pulls your calendar, the emails you promised to answer, and the files waiting on you into one list.
  • A weekly report on your website traffic, search rankings, or ad results.
  • Research you'd otherwise do by hand, such as a competitor's new pricing or a list of prospects that match your best customers.
  • Drafts that you'll read before anyone else does, such as a landing page, an email sequence, or next week's social posts.

Mathis adds one more filter from his own experience. His support agent read his inbox every ten minutes, drafted replies from his sent folder, and sent them with nothing holding back a reply that made a promise. It told six customers in eight days that a fix would be "out shortly" while the fix took ten. His rule since then is that "an agent's first job should be one where the only person who ever sees it fail is me" (Yousef Mathis on Medium). Starting with work that reaches you before it reaches a customer gives you a month of evidence before anything leaves the building.

How do you give an agent a goal, context, and access to your tools?

Describe the result you want and the limits around it, then connect the agent to the apps where the work lives. The agent works out the steps. Kite's documentation gives the same advice to its own users: "Describe the outcome you want, not the steps" (Kite).

Frame the goal carefully, because the agent will pursue exactly what you say. The owner of Wibci, an Australian AI consulting firm, gave Claude Code $1,500 and control of a Meta ads account for his small newsletter. The brief asked for the lowest cost per lead over a 30-day experiment. The agent tested about 50 ad variants, found a whiteboard-style ad that hit $1.29 a lead, and raised its own budget under rules he'd set. Because it knew the test ended on day 30, it doubled down on what was already working instead of trying bolder ideas, and nothing in the brief told it what a good lead looked like. It didn't check who was actually signing up until day 16. His conclusion: "'Minimize CPL over 30 days' produces very different decisions than 'build a sustainable acquisition engine'" (Technically). If quality matters to you, put it in the goal.

Give it the context only you have. A ready-made agent starts with whatever it can read. Tim Moylan found that his new dot had already read his past ChatGPT conversations and opened by naming his projects and offering specific help (AI: Beyond the Buzz). When you sign up, Kite reads your website, researches your competitors, checks how you show up in search and AI answers, and records your positioning and brand voice in a knowledge base it works from. What no agent can read is what's in your head: who your best customers are, which competitor worries you, what you're launching next. Kite's documentation suggests replying to its first message with exactly those things, because correcting its picture of your business in the first week saves effort on everything that builds on it (Kite).

Connect only what the job needs, and connect it read-only where you can. Most agents let you set what they may read separately from what they may change. OpenAI's documentation notes that you can let a dot read your email without letting it send any (OpenAI). Moylan suggests a simple test before you start: list everything you've connected, and disconnect any folder you wouldn't hand a new assistant on their first day (AI: Beyond the Buzz).

Budget setup time for anything unusual. Common apps tend to connect quickly. Barnes connected QuickBooks, PayPal, and HubSpot in about 40 minutes, and his existing permissions carried over, so the agent could only see what he could see (Digital Evolution Marketing Group). Nguyen's afternoon went differently because his setup was less standard. His dot assumed he used Gmail when his business runs on Zoho, and connecting his agency's own monitoring tool took most of the afternoon. Once the connections worked, the same dot reached his clients' Google Analytics and Search Console data and crawled both client sites (SLIDEFACTORY).

How does an agent's work start?

An agent's work starts in one of four ways: you ask, a schedule fires, an event happens in a connected app, or the agent notices something on its own. Knowing which one you're relying on tells you what to expect each morning.

  • You ask. You message the agent in its app, in Slack, or by email, and it takes the job from there.
  • A schedule. You set a time, such as every weekday at 9 a.m. for the next four weeks, and the agent runs the job and reports. ChatGPT's scheduled tasks, Claude Cowork's scheduled tasks, and Zapier's agents all work this way (OpenAI, Anthropic, Zapier).
  • An event. A new email, a new row in a spreadsheet, or a new message in a Slack channel starts the job. Zapier lets you pick an app event as an agent's trigger, and ChatGPT can run a task when a new Gmail message arrives.
  • The agent notices. Some agents watch for work themselves. OpenAI says a dot "can decide when to pause and wake up to continue work" (OpenAI). Kite scans the Slack channels it's in every half hour, matches new messages against your active marketing goals, and posts a digest when it finds a move worth making. It also watches your site analytics, search rankings, and AI visibility, and tells you when something changes (Kite).

The fourth kind is where an agent starts to feel like a colleague. When Moylan asked his dot to get him ready for the next day, it went through his calendar, his email, and his shared files, then turned the result into a daily 8 a.m. briefing without being asked. It also wrote its own limits into the task: "Only read and summarize; do not send messages to others or change my calendar" (AI: Beyond the Buzz).

Whichever way work starts, tell the agent which results deserve your attention and where to send them. Connecting an agent to a channel isn't the same as telling it what to watch: OpenAI notes that adding a dot to a Slack channel "alone doesn't start monitoring" (OpenAI). And put the agent where you'll see it. Kite's documentation recommends a channel people check every day, because the agent "moves at the speed of your replies" (Kite).

How much should an agent do without asking you?

Decide action by action, and put the approval line at the things that are hard to undo. Reading, researching, analyzing, and drafting can run freely in almost every setup, because a wrong draft costs you a rewrite. Sending an email, publishing a page, and spending money usually wait for your OK at first, because you can't take them back.

The agents built for owners already work this way by default:

  • Kite proposes and waits before anything that changes public messaging, spends money, or affects legal standing, revenue, or customer retention. Routine, reversible work it does and reports. You approve by replying "Publish it" or "send it" in the conversation, in the Kite web app or in Slack (Kite).
  • Meta's Muse asks by default before every action that changes something in a connected app and before important reads, and Meta's payments help says it will always ask before completing a purchase (Meta, Creative AI News).
  • OpenAI's dots run an automatic review before any action that touches your accounts, and you can add rules such as "ask before taking action" for messages to customers (OpenAI).
  • Lindy waits for approval on anything with outside impact, such as sending an email or publishing a document, while lookups from approved sources run freely (Lindy).

Two details make this work in practice. First, permission to draft isn't permission to send. OpenAI's documentation says so directly: "Asking your dot to draft replies doesn't give it permission to send them" (OpenAI). Second, you can make the line sharper than any product default. After his support agent's "out shortly" replies, Mathis added a hold: a draft now waits for him if it names a date, says "I will" or "we will," or mentions money. About one reply in seven waits, and the rest go out on their own (Yousef Mathis on Medium).

Be deliberate about shortcuts that switch approvals off. Muse offers an "Always allow" button whenever it asks, and Creative AI News points out that one tap stops the prompts for that action on that app from then on, including email and Klaviyo (Creative AI News). That's a useful button once you trust the agent with that action. It's worth pressing on purpose rather than on a busy afternoon.

How do you review an agent's work and give it more room?

Read the first results closely, judge them by whether the job actually got done, and widen the agent's job one action at a time as its record builds. That's how you'd bring on a new hire: you check their work for the first stretch, and once you're confident, you let them run.

Check results, not status messages. OpenAI warns that "a completed run doesn't by itself confirm that the requested result was achieved or delivered" (OpenAI). One of Nguyen's hardest problems on his first day was "knowing what had actually finished." Open the page, read the email, look at the report. Mathis spends the first hour of most mornings reading what his agents finished overnight, and he calls that hour the real cost of running them.

Correct with reasons. When a draft misses, say why. Kite's documentation notes that "Too salesy for this audience" teaches the agent more than a plain no (Kite). The agent keeps what it learns in its notes or knowledge base, so the same correction rarely has to be made twice.

Tell it what to do when something looks wrong. John Roman runs an online store whose listening agent launched on June 4 able to reply in one Slack channel and message him, and nothing else. After four weeks of watching it call requests correctly while people did the work, his team let it build one kind of product listing on its own. On its first build, it read a store field that reported zero sales channels, decided it lacked permission, and quietly skipped attaching the product to the store's 24 channels. The field was unreliable, and the instructions hadn't said whether to stop and report or carry on when a number looked strange. Now every new action his team hands over spells that out (Online Queso). His readiness test is a good one to borrow: if you can't describe the trigger, the exact steps, and what the agent should do when something looks off, the job isn't ready to hand over yet.

Widen the job as the record builds. Enrich Labs, which sells an AI marketing agent called Helena, gave her a Shopify store, a Meta ads account, a supplier, and a Slack channel. The team supervised closely for two weeks, then pulled back to approvals in chat and a short morning report. Every morning at 7:00 she walked the buying path on every page with ad spend behind it, which caught a dead checkout and broken images before they wasted a day of spend. In August, she spotted a trending product on her own, found a supplier, built the page, prepared the ads, and queued the launch. The team first heard about it as "a finished launch waiting for a yes" (Enrich Labs). They said yes, and it became the store's best product. That's what a widened job looks like: the agent does the work and the judgment, and you make the call on the money.

Most products let you formalize that step when you're ready. Kite has an "act first" setting, which its documentation describes as being "for teams who have seen Kite's work and want it to keep going" (Kite). OpenAI's custom rules let you mark a specific action "take action without asking." Muse's help page says the agent "learns over time which decisions need your sign off" (Meta).

One thing doesn't move as the agent earns more room: you're accountable for what it does. Meta's help page states it plainly: "Because Muse acts on your behalf, you're responsible for guiding it carefully and approving its actions." The agent can own the work and much of the judgment. The responsibility for the outcome stays with the person who put it to work.

A worked example: putting an agent to work on your marketing

Marketing is one of the best places for a small business to use an agent, because the results are measurable. A page either brings in signups or it doesn't, and either way the result tells the agent what to try next. Here's what the steps in this guide look like with Kite, an AI marketer you can work with in your browser or in Slack.

Starting. You sign up and choose whether to work with Kite in the Kite web app or in Slack. Kite reads your website, researches your competitors and how you show up in search and AI answers, asks about your goals, and suggests work it could start on. Soon after, it delivers a growth report on your company. In Kite's own walkthrough, for a three-person startup that sells scheduling software to clinics, it offers a comparison page against the competitor the site mentions most, and a rewrite of the pricing page headline (Kite).

Giving it the job. The founder replies "Do the comparison page." Kite researches the competitor, writes the page, and builds it on the site as a draft. If the team wanted something else, they'd just say so in the thread, such as asking for a list of 50 multi-location clinics in Texas with a contact at each.

Reviewing. Kite posts the draft with a Preview button. The designer replies, "Lead with the time saved at the front desk, not the feature list," and Kite revises. When the founder replies "Publish it," Kite publishes the page and posts the live link.

Letting it find work. From there, Kite brings work to the team as well as taking requests. It posts digests when something in the team's conversations lines up with an active goal, and it proposes larger initiatives, such as winning the comparison searches against a rival, each with an Approve button. Recurring jobs, like a weekly conversion check or a monthly competitor sweep, run on a schedule (Kite).

Approving what matters. Under the default setting, Kite proposes and waits before anything that changes public messaging or spends money, and it handles routine, reversible work on its own. It saves a version of your site after every change, so any edit can be restored (Kite).

This example also shows why it pays to give a marketing agent the goal and not just the tasks. Joseph Darnell's agency set up a Claude Cowork agent to sell a $149 report to estate and elder-law firms, with the target customer, product, and price locked as his decisions and the tactics spelled out. After four weeks, it had made no sales. The agent's daily logs had flagged early on that the people actually reading the agency's blog were agencies and operators, not lawyers. Darnell's own assessment is that he "over-specified the tactics ... so the agent had no room to find a smarter path," and that the strategy underneath "was mine, and it was wrong" (Automaton). An agent built for marketing can weigh in on who to sell to and what to say, not only carry out the plan, and you get the most from it when you let it.

What do AI agents cost?

Most agents built for owners have a free plan with usage limits or a paid plan starting around $20 a month, and the always-on agents meant for heavier use start around $100 a month. The table below shows published prices as of October 8, 2026.

Agent

Free option

Paid plans

Kite

Free plan with access to the complete product and $50 in one-time signup credits

Growth is $100 a month and includes $100 in usage credits; Business is custom-priced

ChatGPT and OpenAI's dots

ChatGPT Free has no scheduled tasks or dots

Go at $8 a month and Plus at $20 a month include scheduled tasks; dots come with Pro, which starts at $100 a month

Claude Cowork

Not included on the free plan

Pro is $20 a month, or $17 a month billed yearly, and includes handing off and scheduling tasks

Meta's Muse

Free with a weekly usage limit that Meta doesn't publish

Power is $20 a month; Maximum is $100 a month

Zapier Agents

Free for up to 400 agent activities a month

Pro is $400 billed yearly, about $33 a month, for 1,500 activities

Lindy

$50 in credits that last seven days

Team starts at $29.99 a month for each active user

Sources: Kite, Kite credits, OpenAI, Anthropic, Meta, Zapier, Lindy.

Pricing models differ in ways that matter for budgeting. Kite, Lindy, and Zapier charge by usage, so bigger jobs draw more from your balance. Muse measures usage in tokens a week, and Creative AI News notes there's no published guide to what a typical task costs (Creative AI News). Kite's pricing is per workspace, so one plan covers everyone on your team.

The cost that doesn't show up on a pricing page is your attention in the first month. Mathis puts it bluntly: "People search for what an AI agent costs, and they mean the subscription. The subscription is the smallest number involved." Plan on reading the agent's work daily at first. The point of choosing a checkable first job is that the reading goes quickly, and it shrinks as you hand over more on trust.

Where to start this week

You can put an agent to work this week with three decisions.

  1. Pick one recurring job you already do and can check in a minute, such as a Monday summary, a morning briefing, or a weekly marketing report.
  2. Connect only the apps that job needs, read-only where you can, and describe the result you want, including what a good result looks like.
  3. Set the approval line at sending, publishing, and spending, read the first few weeks of work closely, and hand over the next action once the agent has earned it.

If you're still choosing an agent, The best AI agents for small business in 2026, sorted by the job compares the options by the work they do. If marketing is the job you want to hand over, What is an AI marketing agent? explains what a marketing agent can take on today.

Common questions

Are AI agents free to use?

Many are, within limits. Meta's Muse is free with a weekly usage limit, Zapier Agents is free for 400 activities a month, and Kite's free plan includes the complete product with $50 in signup credits. Agents that stay on in the background and handle larger jobs usually need a paid plan. OpenAI's dots, for example, come with ChatGPT Pro, which starts at $100 a month.

Do I need to know how to code to use an AI agent?

No. Ready-made agents are set up by connecting your apps and describing the work in plain language. With Kite, you sign up and reply to its messages in your browser or in Slack. Coding only comes in if you want to build your own agent from parts, and even then tools like n8n and Zapier offer visual builders.

Is ChatGPT an AI agent?

The ordinary ChatGPT chat is an assistant: it responds to each message and leaves the decisions to you. Google Cloud draws the line this way: an assistant "responds to requests or prompts from the user," while an agent works toward a goal on its own initiative (Google Cloud). OpenAI's dots, available on its Pro and Business Premium plans, are the agent version. They keep working between conversations and act in your connected apps.

How are small businesses using AI agents?

The documented uses cluster around recurring work the owner used to do by hand. Barnes's test firm uses an agent to chase overdue invoices, summarize the week's numbers, and prepare the month-end close for its accountant. A grocery owner quoted by Meta photographs employees' time cards and has Muse add them up, which he says cuts at least 15 minutes from payroll (Meta). Marketing is another common job: researching competitors, drafting and publishing pages, building prospect lists, and reporting on what's working.

What happens if an agent makes a mistake?

You fix it the way you'd fix a colleague's mistake, and some actions are easier to reverse than others. A draft can be rewritten and a website change can be rolled back. Kite saves a version of your site after every change for exactly that reason. A sent email can't be recalled, as Meta's help page points out, which is why sending is the most common action to keep behind an approval while you build trust. Stopping an agent also doesn't reverse what it already finished, so review the work itself, not just the agent's report.

Sources

Questions this page answers

  • how to use ai agents

About the author

Devon Wells leads product marketing at Kite, the AI marketer. He spent a decade marketing industrial AI, from self-driving trucks to computer vision, but now Kite handles most of the work itself.

Tell me your goal and I’ll get to work.

Get started for free