Human in the loop: using AI in a small business without losing control

Why not just let it run?

AI is good at drafting, summarising, sorting and spotting things that look odd. It's also confidently wrong some of the time, and it doesn't know your customers the way you do. In a big company there are layers of people to catch a mistake. In a small one, a wrong invoice or a sharp reply goes straight to a customer you've known for years, and you're the one who answers for it.

That's why most owners who switch automation on end up scaling it back. Not because the AI is useless, but because they couldn't choose where it acted alone and where it asked first.

Three settings are enough

• Off: a person does the job, as now.
• Draft it for me to check: the AI prepares the work and a person approves, edits or turns it down. Nothing is sent or changed until they do.
• Do it and tell me: the AI does the job and reports what it did, ideally with a way to undo it for a while afterwards.

The point is to set these per job, not once for the whole business. The same owner might happily let the AI chase overdue paperwork but want to see every message to a customer.

Which jobs to hand over first

Start where a mistake is cheap and easy to spot, and the work is repetitive:

• Summaries you read yourself: the week's figures, a customer's history before a call, what's overdue.
• Flags: invoices that look low, timesheets that don't add up, stock that's about to expire.
• First drafts you'd write anyway: reminders, follow-ups, notes tidied from someone's scribbles.
• Rewording: "make this friendlier" or "make it shorter" on something you wrote. This is the AI habit most people actually keep.

What should always wait for a person

• Money decisions: prices, discounts, refunds, writing off a debt.
• Anything about whether someone may work, or is competent to.
• Complaints and disputes.
• Judgements that need professional training: clinical, legal, safety.
• Deleting anything.

Good software makes these impossible to automate, rather than relying on you to remember not to.

Let it earn more freedom

1. Start every new job at "draft it for me to check".
2. Watch how often you approve without changing anything. If you've sent the last few dozen reminders exactly as drafted, the AI has earned that job.
3. Move that one job to "do it and tell me", keep an undo window, and look at what it did each day for a while.
4. If it goes wrong, move it back. Nothing about the setting should be permanent.

Ask three questions of any AI tool

• Can I see everything it did, who approved it, and what it cost?
• Can I choose the setting for each job myself, without asking the supplier?
• What does it send to the AI provider, and does the provider use my data to train its models?

If a tool can't answer the third question clearly, ask before you put customer or staff details through it. You're responsible for the personal data you share, and the business needs a lawful basis to use it this way.