What Is an AI Agent and How Does It Work in Practice

The short version: an AI agent is software that can look at a goal, break it into steps, use tools or data on its own to get there, and check its own work before it hands anything back to a human. It's different from a chatbot because a chatbot answers you and stops, while an agent keeps going until the job is done, or until it hits something it can't handle and passes it up the chain. In practice, most "agents" running in small businesses today are far narrower and more supervised than the marketing suggests, and that's the bit worth understanding before you spend a penny.

In plain terms, what an AI agent does differently

A chatbot is reactive. You ask, it answers, the conversation ends. An AI agent is given a job, not just a question, and it works through that job across multiple steps without you typing a new prompt at every stage.

Say the job is "qualify this inbound lead and book a call if they're a fit." A chatbot can answer questions about your pricing. An agent can read the enquiry, pull up your calendar, check the lead against your criteria, send a reply, book the slot, and log the whole thing in your CRM, all without a person touching it in the middle.

The three things that make something an agent rather than a script:

• It can decide which tool to use next based on what it finds, not a fixed order someone coded in advance.
• It holds context across several steps, so step four remembers what happened in step one.
• It can check its own output against a goal and try again, or stop and ask a human, rather than just outputting whatever comes first.

That last point is the one people get wrong most often. Genuine agent behaviour includes a feedback loop. If it doesn't check itself against something, it's a workflow with a language model bolted on, not an agent.

How it works in practice, step by step

Underneath the label, most working AI agents follow the same rough pattern. I've watched this built for small businesses more than a dozen times now, and it almost always looks like this:

• 1. Trigger. Something happens: an email lands, a form is submitted, a calendar slot opens up, a certain time of day arrives.
• 2. Retrieve context. The agent pulls in whatever it needs to understand the situation: past emails from that contact, your pricing sheet, your booking rules, your tone-of-voice guide.
• 3. Plan. The underlying model (often GPT-4o, Claude, or a fine-tuned version of one of those) works out what steps are needed to reach the goal it's been given.
• 4. Act. It calls tools: sends an email through your email API, checks a calendar via Google Calendar's API, writes a row into a spreadsheet or CRM, searches a knowledge base.
• 5. Check. It compares the result against the goal. Did the email send? Does the booked slot match availability? Is the answer it's about to give backed up by something in the knowledge base, or is it guessing?
• 6. Escalate or loop. If something doesn't check out, or the situation falls outside its rules, it stops and hands the case to a person, ideally with a note explaining why.

That escalation step is the one businesses skip when they build these things in a rush, and it's the one that causes the most damage. An agent with no escalation path will confidently send the wrong refund amount, book two people into the same slot, or promise a delivery date nobody agreed to, and it will do it politely and fluently, which makes it worse, not better.

A real example: what this looked like for one small agency

Last year I worked with a seven-person lettings agency in Kent that was getting around 40 property enquiries a week through their website, WhatsApp, and Rightmove. Two people were spending most of their mornings just triaging: reading each enquiry, checking if the person's budget matched the property, replying, and trying to book viewings around a calendar that never seemed to have a free slot everyone agreed on.

We built an agent, sitting on top of their existing booking system, that did the first pass. When an enquiry came in, it read the message, pulled the property details and the enquirer's stated budget and move-in date, checked those against the listing's actual terms, and if everything matched, it replied with three available viewing slots pulled straight from the live calendar and booked whichever one the person picked.

Where it got interesting was the exceptions. Around 15% of enquiries didn't fit the rules cleanly: someone asking about a pet policy that wasn't documented, a budget that was close but not quite there, a message in a mix of English and another language. For those, the agent didn't guess. It flagged the enquiry, wrote a one-line summary of why it wasn't sure, and put it in a queue for a human to handle within the hour.

The result after six weeks: the two staff members got back roughly six hours a week each, and the average time to first response dropped from around four hours to under three minutes for the enquiries the agent could handle cleanly. The 15% that got escalated still took a human to sort, and that's exactly the point, because the agent wasn't pretending to be smarter than it was. This is close to the same setup I've described when talking about how real estate agents can use AI for appointment scheduling, and the pattern holds for other service businesses too.

The uncomfortable bit nobody puts in the sales deck

Here's what gets left out of most explainer articles on this topic: the word "agent" is doing a lot of marketing work right now, and a large share of what's sold as agentic AI is a well-organised workflow with a chatbot wired into the middle of it. That's not a criticism of the technology, it's a criticism of how it's being sold.

A agentic system that plans its own steps and adapts on the fly is powerful, but it's also harder to control, and control is what most small businesses need more than raw capability. The lettings agency example above worked because we deliberately narrowed what the agent was allowed to decide. It could pick a time slot from a fixed calendar. It could not decide to offer a discount, waive a deposit, or override a policy, even if a clever enough prompt from a customer tried to talk it into it. Every time I've seen an agent cause real damage, it's because someone gave it more decision-making room than the business could afford to have unsupervised.

The other thing worth saying plainly: an agent connected to real tools, real emails, real bookings, real money, is a different risk category to a chatbot that just talks. A chatbot that hallucinates gives you a wrong answer you can correct. An agent that hallucinates sends a wrong email, moves real money, or cancels a real booking, and by the time a human notices, the action has already happened. If you're building or buying one of these, the review step and the escalation rules matter more than how clever the underlying model is.

Where agents earn their keep for small businesses right now

Not every job suits an agent, and it's worth being blunt about which ones do. The pattern I see working repeatedly:

• Lead triage and booking, like the lettings example above, where the rules are clear and the exceptions are a manageable minority.
• Reporting that pulls from several sources, checking numbers against each other and flagging anomalies before a human reads the summary. This is the exact shape of what I've written about for AI reporting and analytics in dental practices, where an agent checks appointment data, no-show rates, and revenue figures against each other overnight so the practice manager gets a clean summary rather than three separate spreadsheets.
• First-draft content that goes through a human gate, which fits into the same thinking I've laid out around what generative AI does for marketing results, where the agent drafts and checks, and a person still signs off before anything goes public.
• Approval routing, where an agent drafts a social post, checks it against brand guidelines, and pushes it into the right person's queue rather than posting it blind, which is the same logic behind a good AI social media approval workflow.

Notice the pattern: in every useful example, a human still checks the output before anything final happens, or the agent's own decision space is deliberately small. That's not a limitation of the technology so much as the correct way to use it while it's still this new and this fast-moving.

What it costs to set one up

Prices vary a lot depending on how much custom building is involved. A basic agent built on off-the-shelf tools, something like a Zapier or Make agent wired to your existing calendar and email, can be set up for a few hundred pounds if you're doing it yourself with some patience, or somewhere between £1,500 and £5,000 if you're paying someone to build and test it. A more bespoke agent, custom-coded and connected to several internal systems, tends to run from £8,000 upward, and that's before ongoing model costs, which for a small business running a moderate volume of enquiries usually land somewhere between £50 and £400 a month depending on how much the agent is doing and which model it's calling.

The bit that catches people out isn't the build cost, it's the testing time. A workflow you can plug in and trust in a day, an agent needs weeks of watching it handle real enquiries with a human checking every decision before you let it run unsupervised. Rushing that stage is where most of the horror stories come from. If you're weighing up whether to build this in-house or bring someone in, it's worth reading through what a proper realistic AI implementation timeline looks like before you commit to a launch date, because the testing phase is almost always longer than the sales pitch implies. And if you'd rather not learn all this the hard way on your own business, working with an AI implementation coach to scope the first agent tends to pay for itself in avoided mistakes alone.

A quick way to check if something calling itself an agent is one

Before you buy or build, ask the vendor or the person building it these four questions:

• What decisions is it allowed to make on its own, and what's the exact list of things it's not allowed to do?
• What happens when it's unsure, does it guess, or does it stop and hand off?
• Can you see a log of every action it took and why, after the fact?
• How was it tested before it touched real customers, and for how long?

If the answers are vague, what you're being sold is probably a chatbot with extra steps, not an agent, and the gap between those two things is exactly where most disappointment with this technology comes from.

Frequently asked questions

Is an AI agent the same thing as a chatbot?

No. A chatbot responds to a message and stops there. An AI agent is given a goal, works through several steps on its own using tools like a calendar or email account, checks its own output, and only stops to ask a person when it hits something outside its rules.

Can an AI agent run without any human oversight at all?

Technically yes, but in practice it's a bad idea for most small businesses. Every working example I've seen that avoided real damage had a narrow decision space and a clear escalation path to a human, rather than full autonomy.

What's the first process a small business should turn into an AI agent?

Pick something with clear rules and a manageable exception rate, like lead triage, appointment booking, or first-pass reporting. Avoid anything touching money, legal commitments, or customer promises until you've watched a simpler agent run cleanly for a few weeks.

How long does it take to get an AI agent working reliably?

The build itself can take days to a few weeks depending on complexity, but reliable running usually needs another two to six weeks of a human checking every decision before you trust it unsupervised. Businesses that skip that testing window are the ones most likely to have a public mistake.

Primary sources

• Anthropic's own documentation
• the NIST AI Risk Management Framework

Related reading: How to Build Your First AI Agent for Marketing: A Non-Technical Founder's Guide (2026) and The AI for Marketing Era Is Quietly Ending. AI Agents Are Taking Over.

For the bigger picture, see my full guide to AI marketing.