Agents for a One-Person Business

Most writing about agents assumes an integration team, a security review and a governance programme. Working alone you have none of those, and you also have advantages: your systems are modern, your data is small, and you can decide something in an afternoon.

Where the advantage is

Your tools have interfaces. A small business running current software may have less integration work than an enterprise with decades of accumulated systems. See integration is the hard part.

For a vendor-side perspective on being proactive at work, see this resource from Monitask.

Your data is small and you know what it means. The thing enterprises struggle with — nobody understands the field names — does not apply.

You can change your mind immediately.

And you are the reviewer, so the oversight question has an obvious answer.

Where the disadvantage is

No one catches your mistakes. The review that an organisation gets from having several people is absent.

Cost surprises hurt proportionally more. See below.

You have no security review, and the same risks apply. An agent connected to your accounts is connected to your business.

And there is nobody to tell you the thing is not working. Sunk cost is harder to see alone.

What actually works at this scale

Patterns that hold up, rather than a tool list.

Drafting from your own material. Proposals, replies, summaries, first versions. You review everything, which is the correct oversight level.

Research with sources. Gathering and organising, with links you can check. Verify the sources — plausible citations that do not exist are a known failure.

Structured extraction. Pulling fields out of documents, receipts, emails into a consistent form. Repetitive, checkable, and errors are obvious.

Code and scripts, if you write any. The area where agents are most mature.

Triage. Sorting incoming things into categories for you to handle, without acting on them.

And preparation rather than execution. The pattern that works best solo: the agent prepares, you release.

What to never connect

Anything financial. Payment methods, banking, purchasing, invoicing that can send. No exceptions, and this is where the incidents happen.

Anything that sends without review. Email, messages, social accounts, client communication.

Client data you are contractually obliged to protect. Check what your contracts say before connecting anything to it.

Production systems you cannot restore.

Your primary credentials. Use separate accounts you can revoke without disrupting yourself. See permissions.

Keeping the bill sane

Set a spending limit at the provider, before the first long task. Most allow one.

Watch the first week closely. Your real usage will differ from your expectation.

Be careful with anything unattended. A scheduled agent that loops on a failing call is the scenario that produces the unexpected bill.

Failed tasks cost more than successful ones, because struggling means more iterations. See what agents cost to run.

And check what per-seat tools actually include. Usage caps are frequently buried.

The security you should actually do

Proportionate, and not nothing.

Separate accounts, revocable.

Read-only where the task allows it.

Review anything reaching a client.

Treat everything the agent reads as untrusted, including documents people send you. An instruction planted in a document is the realistic attack, and it applies at any scale. See prompt injection.

Check what a connected integration is actually granted. The consent screen usually grants far more than the task needs, and it is the place to notice. For broader independent background, see NIST AI Risk Management Framework.

And know how to turn it off. One action, tested.

Deciding whether it is worth it

Give it two weeks. If it is not saving time by then, it will not.

Count the review time. An agent that produces drafts you spend as long correcting has saved nothing.

Count the fixing. Including the ones you nearly sent.

Compare against the simplest alternative. A template, a saved search, a keyboard shortcut, a checklist. Frequently these win. See when an agent is the wrong answer.

And notice if you have stopped reading the output. That is the point to narrow what it does, not to trust it more.

The thing that catches people out

Delegating the work removes the understanding along with it.

For anything you are accountable for — client deliverables, advice, decisions — you need to understand the output, not just approve it. An agent-produced answer you cannot defend is a liability with your name on it.

Which makes the assistant pattern — you drive, it helps — the right one for most solo work, rather than autonomy. See agent, assistant, chatbot, automation.

The short version

Your advantage is modern tools, small data and fast decisions. Your disadvantage is that nobody catches your mistakes.

Prepare, do not execute. The agent drafts, you release — that is the right oversight level solo.

Never connect anything financial, and use separate revocable accounts for everything else.

Set a spending limit before the first long-running task, and be careful with anything unattended.

And if you stop reading the output, narrow what it does rather than trusting it more.