Guides / AI agents

Guide · Basics

What Is an AI Agent — and Does My Business Need One?

01From answering to doing

Everything in our AI basics guide describes a tool that talks: you ask, it answers, and then you go and do the thing. Ask ChatGPT to handle an appointment request and it will draft you a lovely reply — but it won't check your calendar, send the email, read the response, or book the slot. You're still the one doing the doing.

An agent closes that gap. It's the same kind of AI model, but connected to tools and given a goal rather than a question. "When a booking enquiry comes in: check the calendar, offer two free slots, and when the customer picks one, book it and send a confirmation" — that's agent territory. The AI isn't just writing about the task; it's carrying it out.

02How an agent actually works

Under the bonnet, an agent is three things: a language model, a set of tools it's allowed to use, and a loop. The loop is the magic ingredient. The model looks at the situation, decides on an action, takes it, looks at what happened, and decides again — repeating until the job is done or it needs help.

Here's a real-world example — an invoice-chasing agent:

1. Each morning, look up invoices more than 14 days overdue in the accounts system.
2. For each one, check the email history — has the customer already replied or paid?
3. If not, draft a polite reminder matched to how overdue it is (gentle at 14 days, firmer at 45).
4. Send reminders below £500 automatically; queue anything larger for the owner to approve.
5. Log everything it did, and flag anything odd — a disputed invoice, an angry reply — for a human.

Notice the shape of it: clear rules, small steps, a money threshold above which a human signs off, and a paper trail. That's not an accident — that's what a well-built agent looks like.

03What agents can safely do today

The agents earning their keep in small businesses right now share a family resemblance: the task is repetitive, rule-describable, and checkable. For example:

• Inbox triage — reading incoming enquiries, categorising them, drafting replies for approval, escalating the urgent ones.
• Chasing and follow-ups — overdue invoices, unanswered quotes, appointment reminders.
• Data shuttling — moving information between systems that don't talk to each other: order forms into the spreadsheet, spreadsheet into the accounts package.
• Report assembly — pulling the week's numbers from a few places into one summary every Monday morning.
• Scheduling — the back-and-forth of finding a time that works, which nobody on your team will miss.

The common thread: if the agent gets one wrong, the cost is small and the fix is easy. Low stakes, high volume — that's the sweet spot.

04Where the hype is

You'll also hear much bigger claims — agents that run entire departments, autonomous businesses, "digital employees". Treat these the way you'd treat any salesman whose product has no off switch.

The honest technical picture: agents inherit every limitation of the models underneath them, including making things up — and in an agent, errors compound. A wrong guess at step two becomes a wrong action at step three. The longer and more open-ended the task, the more chances there are to wander off course. That's why the agents that work are short-leashed, and the ones in the press demos so often aren't doing what the caption says.

Our rule of thumb: trust an agent the way you'd trust a capable temp on their first week. Short tasks, written instructions, no signing authority, and someone checks the work. Expand the trust as it earns it — which, pleasingly, is exactly how you'd manage the human version.

05Does your business need one?

A quick test. An agent is a strong candidate when you can answer yes to all three:

• Is there a repetitive, multi-step digital process eating hours every week?
• Could you write the rules down for a new starter in a page or less?
• Is a mistake cheap to catch and fix — or easy to gate behind an approval?

If the process relies heavily on judgement, happens rarely, or a mistake is expensive and irreversible — don't start there. And whatever you automate, mind what the agent can touch: an agent with access to customer data needs the same care we describe in Is AI Safe to Use With My Business Data?

One more honest note: plenty of businesses don't need a true agent at all — a simpler automation (the same trigger-and-action plumbing, minus the AI decision-making) or just using AI tools well by hand gets 80% of the benefit with less to maintain. The goal is hours back, not technology for its own sake.

06What it costs and how it gets built

Agents for small businesses are typically assembled from existing parts — workflow platforms like n8n or Make providing the plumbing, an AI model providing the judgement, and your existing systems (email, calendar, accounts) as the tools. That makes them a one-off build with a small monthly running cost, not an enterprise software project. A focused agent like the invoice-chaser above is days of work, not months — and it pays rent in saved hours from week one.