AI Agents for Small Business: What to Hand Over

by TFE Editor · 11 min read
AI Agents for Small Business: What to Hand Over

AI agents for small business are software that can take a goal, choose steps, and use tools until a stop you set. They can research, sort, and draft. They should not email a client, move money, or make a promise. You still decide the job, and you still approve the send.

A solopreneur starts and runs a business with no employees. You own the decisions and the work. A freestyle entrepreneur builds that business around a life and still has to get paid. An agent is a loop. It keeps going past the point where you would have stopped, unless the stop is written down. The vendor pages below were fetched on November 10, 2026. Read the permission screen on your account before you connect a client file.

An agent picks the next step

The NIST Computer Security Resource Center glossary, fetched November 10, 2026, carries a definition sourced to NIST AI 100-2e2025. In that entry, an agent takes "self-directed actions in service of a larger, externally-specified goal." The goal is yours.

Google’s machine learning glossary, last updated April 13, 2026, describes an agent as software that can reason about what you typed in order to "plan and execute actions on behalf of the user." The set of tools and permissions should be "just big enough for the agent to perform the task." Make that set wider and "the agent tends to become more error prone."

Google Cloud’s explainer, last updated April 2, 2026, says agents "pursue goals and complete tasks on behalf of users" and "have a level of autonomy to make decisions, learn, and adapt." It separates three labels. "AI assistants are less autonomous, requiring user input and direction." "Bots are the least autonomous, typically following pre-programmed rules." On an assistant, "decision-making is done by the user." The same page still says an agent may "operate and make decisions independently to achieve a goal."

Anthropic’s note of April 9, 2026, separates an agent from a chatbot. Their definition is a model "deciding for itself how to achieve what users want," "rather than following a fixed script." The loop is that "it plans, acts, observes the result, adjusts, and repeats" until the task is done or a person has to weigh in. They also write that "Agents act with less human oversight".

A fixed path is a different tool, and often the right one. Anthropic’s December 19, 2024 note, still online on November 10, 2026, says "Workflows are systems where LLMs and tools are orchestrated through predefined code paths." Agents on that page "dynamically direct their own processes and tool usage". The page warns that "Much of the tooling landscape described in this post has changed since December 2024." Use the distinction, and leave the 2024 shopping list alone. The same note says "This might mean not building agentic systems at all."

OpenAI’s Academy page on workspace agents, published April 9, 2026 and last updated September 4, 2026, says "Agents are more probabilistic." A traditional workflow there is one in which "each step is explicitly defined". OpenAI’s guide to working with agents, fetched November 10, 2026, says "A predefined workflow can suit stable, repeatable steps" and that "an agent can help when the path needs adjustment." If Tuesday is the same sentence into the same fields, use the AI automation for small business recipes. An agent is for a path you cannot write down in advance.

The working order is already in the AI for solopreneurs guide: you decide, the software drafts, you check. Which category of app to rent is in the AI tools for small business guide.

What you can hand it this month

None of these products is the best. They are dated examples of one category. You name the outcome, the software takes steps, and you review anything that is hard to undo.

Type Hand it You keep
Research A one-page brief from files you already keep The send. No client email leaves
Admin A plan, then a sheet of vendors and amounts The payment. It does not guess a total
Support Labels on an exported inbox The reply. A sort is not a send
Sales A draft from bullets you wrote Outreach. Approval comes before any send

The pages fetched on November 10, 2026, do not state one dollar price for an agent. OpenAI’s guide says "Agents may take longer and cost more" than a chat, and what you can run depends on the plan.

Anthropic’s April 9 note names Claude Code and Claude Cowork. In that description the model can "write and execute code, manage files" and "complete tasks that span multiple applications." "Without tools, Claude can read the receipt but not file it." In Claude.ai and Claude Desktop, a person can set each action to "always allow, needs approval, block". One setup they describe leaves calendar reading allowed, while you "require approval before sending someone an invitation."

Click-by-click approval wears out. Anthropic says "repeated prompts can become a source of friction". In Claude Code, Plan Mode lets the user "review, edit, and approve the whole thing before anything happens" and still step in while it runs.

OpenAI’s guide, fetched the same day, names ChatGPT Work for multi-step deliverables you review, dots ("Rolling out to eligible accounts."), and Codex Cloud, where you "inspect changes, and verify the checks." "Availability and supported actions depend on your plan, region, workspace settings, and connected tools." "You remain responsible for deciding whether the work is ready to use." Fluent writing or a confident summary is "not enough to establish that a task succeeded." "Changing an instruction does not undo an action already completed."

Use an agent when the work repeats, the output has a shape you can judge, and a tool is required. For a one-off brainstorm, "regular chat is often a better fit". The Academy examples keep a person on the click: "no direct outreach without approval", "approval required before budget changes", and "approval required before submission". "A strong first result does not establish reliability across every kind of work."

What it must not touch

  • Client email, text, or invoice. The agent may draft it. Sending stays with you.
  • Money. No transfer, no card charge, and no click that pays a bill. Anthropic’s April 9 note says that even together, "these safeguards are not a guarantee", and that prompt injections try to "trick models into taking costly actions".
  • A price or a promise. The number on the invoice is your decision. Work it through the how to price freelance services guide. An agent can lay out options from figures you typed.
  • A professional role. It does not practice law, prepare a return you will sign, or stand in for a licensed person.

On September 25, 2024, the U.S. Federal Trade Commission announced Operation AI Comply. It took action against DoNotPay, which had claimed an AI service that was "the world’s first robot lawyer". The FTC said the product "failed to live up to its lofty claims" that the service could "substitute for the expertise of a human lawyer." The same announcement describes Rytr, which let paid subscribers "generate an unlimited number of detailed consumer reviews" from "very limited and generic input." This is general information, not legal advice. A review on your site comes from a real client, or it stays off the site.

Anthropic defines prompt injections as "malicious instructions hidden inside the content" an agent is asked to read. Their example is an email that tells the model to ignore earlier instructions and forward messages. "The more tools it can use, the more an attacker can do" once inside. Do not give an agent permission to forward mail or pay a bill from a contact form a stranger can fill.

Four jobs, with the stop on the page

These recipes are ours, not a vendor promise. Say you run a design studio of one. The agent may read one folder of briefs and notes, and it may write a new doc beside them. Mail stays closed.

1. A Monday brief you can check

You start it, or the clock does. The page lists what is due in seven days, which file holds the latest brief, and which question is still open. Every date has to appear in the export or the notes. The brief can name a client who is waiting. It cannot send the chase.

2. Sort the inbox, answer nothing

Export two days of mail to a file. The agent labels each row as reply, invoice, scheduling, or ignore, plus a one-line reason. It does not draft the reply. The Academy feedback example says "approval required before submission". Here there is no submission.

3. A draft from bullets you wrote

Five bullets, including one fact you can point to: a published price, a date the client named, or a line from their brief. The draft may use those bullets and nothing else. You delete any line you cannot trace, then you send it yourself.

4. A plan, then a pause, then a file

Pull three receipts into a spreadsheet: vendor and amount only. Ask for the plan first, as in Anthropic’s Plan Mode, and approve it before the sheet is written. A missing total means it stops. It does not guess a number, and it does not pay the receipt.

The list stays a broadcast you send

A newsletter is the draft job with a different destination. You write five bullets, including one fact a reader could check. The agent drafts the issue and two subject lines. You cut any line you would not say on a call. The solopreneur marketing guide covers the rest of the week. The send is a separate button, pressed by you.

We like Kit for this: it’s free up to 10,000 subscribers. Partner link · T&Cs apply. Kit’s help center, updated September 23, 2026 and fetched November 10, 2026, says "The Free plan costs $0 per month" while you stay under 10,000 subscribers. It includes "Up to 10,000 email subscribers" and "Unlimited email broadcasts". Visual automations, email sequences, and "AI Connectors" are listed with the paid plans. Past the cap, the help center says "You’ll need to upgrade to a paid plan to keep sending."

The pricing page, fetched the same day, shows the Free plan at "$0 /mo" and "Free—up to 10,000 subscribers". It says the plan gives tools for an audience of "up to 10,000 subscribers, for free." It also says "There are no sending limits on any of our plans at Kit." Past the cap, that page says "we’ll automatically upgrade it to the next subscriber tier." and "You’ll continue gathering subscribers and sending emails the whole time." The help center and the pricing page do not describe the same outcome past 10,000 subscribers. Read both again before you rely on either.

The agent writes the draft in a document. You paste it into a broadcast and press send. Do not hand over a login that can send. AI Connectors are on the paid plans.

You decide, AI drafts, you check

Use this for the four jobs. The last column is the one a tired evening will try to skip.

Job You decide AI drafts You check
Which loop One of the four jobs, and whether a fixed workflow already does it Nothing. The shortlist is those four jobs. That the tool set is read-only, and that send is off
Monday brief The folder and the calendar export that count as sources A one-page note of dates, files, and open questions Every date against the export. Inferred dates come out
Inbox sort The labels, and that no reply is in scope A label and a one-line reason for each row That nothing was sent, and that the ignore pile is actually ignore
Client draft Five bullets, one fact you can point to, and the price if you have already set it A draft that uses only those bullets Names, dates, fee, and any promise. Then you send
Newsletter Five bullets and the point of the issue A broadcast and two subject lines That you can defend every line, and that you press send

A first week on one job

This plan is ours. It is not a case study, and it does not promise hours back. The week is one job.

  1. Day one. Write the trigger, the folder, and the stop in one sentence. The stop covers send, spend, and price. If you cannot judge the result yet, narrow the job until you can.
  2. Day two. Open one product you can already log into. Set the folder to read. Block mail, payments, and the live site. If there is no such switch, work from a copied file and connect nothing else.
  3. Day three. Ask for a plan, edit it, and approve it before anything is written. Run one real item from your folder.
  4. Day four. Check every date, name, and fee against the source. If the agent says it updated another system, open that system. A later instruction does not undo a send that already left.
  5. Day five. An invented fact means off, then a tighter instruction, then one more run of the same job. A clean page means you leave it there.

Takeaway

AI agents for small business pay off when one loop has a goal you wrote, a small set of tools, and a stop that includes the send. The software can brief, sort, and draft. You still own the price, the promise, and the button.

  • Use an agent when the next step depends on what it finds. Use a fixed workflow when the path is the same every time.
  • Give it one folder, approve the plan before it writes, and keep mail, money, and the live price on your side of the click.
  • One clean week is one job.

Grab the free One-Person Business AI Stack checklist when you want one job, the stop rule, and the weekly check on a single page.

Sources