The short version: an AI agent is software that can take a goal, break it into steps, use tools (email, your calendar, your CRM, a website) to complete those steps, and check its own work, mostly without you sitting there clicking buttons. It is not the same as a chatbot, and it is not the same as a simple automation, though most of what gets sold to small businesses under the name “AI agent” is one of those two things wearing a better suit. Used well, an agent can answer your calls, qualify your leads, chase your invoices or draft your first-pass content, and it can cut hours off a week. Used badly, it is an expensive way to annoy your customers.
What an AI agent is (and what it isn’t)
I get asked this question at almost every talk I do now, usually by someone who’s been told by a salesperson that they “need an AI agent” without anyone explaining what that means. So let’s sort the definition out first, because the word gets stretched to cover things that have nothing to do with agency at all.
A basic automation follows a fixed path. If a form is submitted, send an email. If a payment fails, send a reminder. Zapier and Make have been doing this for years and it’s brilliant, but it’s not intelligent, it’s just plumbing. There’s no decision-making in it.
A chatbot answers questions from a script or a knowledge base. It can sound clever, but it can’t do anything, it can only talk. Ask it to rebook your delivery slot and most of them fall over and hand you to a human.
An AI agent sits above both of those. It’s given a goal, not a script, “get this customer’s refund processed” rather than “answer refund FAQ”, and it works out the steps itself: check the order in the system, confirm it meets the refund policy, issue the refund, email the customer, log the interaction. It uses tools to do each of those steps, and if step three fails, it can try a different approach rather than just giving up. That’s the bit that makes it an agent rather than a script with a friendly voice.
I wrote a more technical breakdown of the mechanics, memory, tools, planning, and how to build a simple one yourself, in this piece on what an AI agent is and how to build one, so I won’t repeat all of that here. This post is about the business case, not the plumbing.
Where agents help a small business
I’ve now tested this twice, not as a thought exercise, with my own phone number and my own money.
The first time, I let an AI answer my business calls for six weeks. The second time, I went further and replaced my business phone with a voice agent for 30 days and tracked what it cost me down to the pound. Between the two experiments, I got a fairly honest picture of where this stuff works and where it doesn’t, and it’s not evenly split.
Where it worked well: taking messages, booking calls into my calendar, answering “what do you charge” and “do you work with X industry” style questions, and filtering out the tyre-kickers who were never going to become clients anyway. The agent didn’t get tired, didn’t get short with anyone at 5pm on a Friday, and didn’t lose a single lead to a missed call, which used to happen more than I’d like to admit.
Where it struggled: anything that needed judgement about a relationship. A returning client with a slightly odd request, someone upset about a delayed project, a prospect who wanted to negotiate. The agent either handled it too rigidly or escalated it to me anyway, which meant the “saving” was partly an illusion for those calls. I still had to deal with them, just after a slightly longer delay.
That’s the uncomfortable bit nobody selling you an agent wants to say out loud: most AI agents on the market today are brilliant at the boring 70% of a job and mediocre at the 30% that requires judgement, and that 30% is usually the part your customers remember. If you deploy an agent expecting it to replace a person, you’ll be disappointed. If you deploy it expecting it to remove the repetitive part of a person’s job so they can spend their time on the 30% that matters, it earns its keep fast.
The numbers, honestly
During the 30-day voice agent test, the cost worked out to roughly £180 for the month, including the platform fee, the number, and the per-minute charges for calls, against a part-time receptionist cost that would have run into the thousands for the same coverage. That’s a real gap, and it’s the number that gets quoted everywhere.
What doesn’t get quoted as often is the setup time. It took me the better part of two working days to get the call flow, the tone, and the escalation rules right before it was fit to answer a real client call, and even then I was still tweaking it in week three. If someone tells you an agent is “plug and play,” they haven’t built one that had to deal with an actual awkward customer yet.
For context on cost more broadly, if you’re trying to work out what fair pricing looks like for getting proper AI help set up in your business, whether that’s an agent, a workflow, or a wider strategy, I’ve laid out real UK and US pricing bands in a separate piece on what an AI consultant costs, because the range is huge and most of it depends on scope, not skill.
Five jobs an AI agent can take off your plate
- Call and message handling, answering, qualifying, and booking, as above, at a fraction of the cost of a person doing the same hours.
- Invoice and payment chasing, checking what’s overdue, sending a graded sequence of reminders, and flagging stuck accounts to you rather than nagging politely forever.
- Lead qualification, working through inbound enquiries, asking the three or four questions that determine fit, and only handing you the ones worth your time.
- First-draft content and research, pulling together a competitor scan, a first draft of a proposal, or a summary of a long document before a human tidies it up.
- Scheduling and admin, finding times, sending calendar invites, chasing confirmations, the stuff that eats twenty minutes a day and adds up to a working week a year.
None of these need a huge budget or a technical team. Most small businesses I work with can get a first working agent live within a week using off-the-shelf platforms, the trick is picking one job, not five, to start.
A step-by-step way to deploy your first one
- Pick one repetitive task, not a whole department. Call answering, invoice chasing, or lead qualification are the three I’d start with for most small businesses.
- Write down the actual steps a good human does right now for that task, in order, including the awkward exceptions. This document is the single biggest predictor of whether the agent works, more than the platform you choose.
- Choose a tool that matches the job, not the hype. A voice platform for calls, a CRM-native agent (Salesforce’s Agentforce, HubSpot’s Breeze, or similar) for lead handling, a workflow tool with an AI step for admin.
- Build it with real edge cases, not just the happy path. Feed it the awkward customer, the wrong postcode, the client who shouts. Watch what it does.
- Run it alongside a human for two to three weeks before you trust it alone. This is the step everyone skips and the step that saves you a bad review.
- Review the transcripts weekly for the first month. Not monthly. Weekly. Things drift fast in the first few weeks.
If you’d rather have someone sense-check the build before it goes anywhere near a customer, that’s exactly the sort of hands-on work I do through AI implementation coaching, because the difference between an agent that helps and one that embarrasses you is usually in the detail of that edge-case testing, not the tool itself.
The uncomfortable truth about “autonomous” agents
Here’s the bit that gets glossed over in most of the AI agent articles you’ll read this year: the word “autonomous” is doing a lot of marketing work it hasn’t earned yet. Fully autonomous, no-human-in-the-loop agents making decisions and acting on them without any check exist, but they’re rare in small business right now, and the ones that do exist without a human check tend to be the ones that eventually make an expensive mistake, refunding the wrong customer, quoting the wrong price, or sending an email that reads exactly as robotic as it is.
I’ve seen this first hand. Early in my testing, an agent I’d built for lead qualification quietly started telling prospects a price band that was six months out of date because nobody had told it the prices had changed, and it had no reason to think to check. It ran for eleven days before a client mentioned it to me on a call. Eleven days of slightly wrong quotes going out, politely, confidently, and completely wrong. That’s the risk nobody puts on the sales page.
The businesses getting real value from agents right now are almost all running a “human in the loop” model, the agent does the work, a person spot-checks a sample of it weekly, and there’s a clear, easy way for a customer to say “I want a person” and get one immediately. That’s not a failure of the technology. It’s just where the technology is in 2026, whatever the demo videos suggest.
What this means for how you plan your business around it
If you’re rebuilding or reshaping how your business runs this year, and a fair few of you reading this are, because I get emails from people doing exactly that every week, the sensible move isn’t to bolt an agent onto everything at once. It’s to treat it the way you’d treat hiring: start with one role, get it right, watch it for a month, then expand.
Barbara Corcoran talks about this same instinct in a different context, hire slow, fire fast, know exactly what job you’re hiring for before you go looking, and I think about that a lot when businesses come to me wanting to “add AI” without being able to tell me which specific job they want it to do. If you haven’t read it, there’s a good rundown of her thinking in this piece on business lessons from Barbara Corcoran, and the parallel to hiring an agent is closer than people expect. You wouldn’t hire a person without a job description. Don’t deploy an agent without one either.
I also keep a running weekly note on what’s changing in this space, because it moves fast enough that anything older than a couple of months can be out of date. The 26 July roundup and the 19 July roundup both cover agent platform updates worth knowing about if you’re weighing up tools right now, rather than relying on a blog post written last year that’s already describing tools that have changed twice since.
So, is it worth it for your business?
If your business has a repetitive, rules-based task that a person currently does badly or reluctantly, calls that go to voicemail, invoices that don’t get chased, leads that sit in an inbox for three days, an agent will probably pay for itself within a couple of months. The cost is modest, usually somewhere between £50 and £300 a month depending on volume and platform, plus the setup time I mentioned earlier.
If what you need is judgement, relationship management, or someone who can read a room, an agent will disappoint you, and no amount of clever prompting fixes that yet. Know which one you’re buying before you buy it.
Related reading: How much do you charge?.
Frequently asked questions
What’s the difference between an AI agent and a chatbot?
A chatbot answers questions from a script or knowledge base and can only talk back to you. An AI agent is given a goal, decides the steps needed to achieve it, and uses tools like a calendar, CRM, or email to complete tasks, checking its own results along the way, rather than just producing a reply.
How much does it cost to set up an AI agent for a small business?
Running costs typically sit between £50 and £300 a month depending on the platform and call or task volume, but the real cost is setup time. Building and testing a working agent, including edge cases, usually takes one to three working days before it’s fit to face real customers.
Can an AI agent replace a member of staff?
It can replace the repetitive part of most roles, call handling, invoice chasing, scheduling, lead qualification, but it struggles with judgement calls, upset customers, and relationship nuance. Most businesses seeing real value keep a human checking a sample of the agent’s work weekly rather than removing people entirely.
What’s the safest way to start using AI agents in my business?
Pick one repetitive task, not several. Document exactly how a good human currently does it, including the awkward exceptions, build the agent against that document, then run it alongside a person for two to three weeks before letting it operate alone, checking transcripts weekly during that period.
Useful references
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.
