What is an AI agent, and does your business actually need one?

01What is an AI agent, in plain words?

Every piece of software you pay for is now selling you an agent. Your CRM has them, Microsoft has them, your phone system probably has one. Yet Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Both are true for the same reason: the word now covers everything from a chatbot to a system that runs a job on its own.

So here’s the plain version. An AI agent is AI with a goal, some tools, and permission to decide the next step. It reads what’s in front of it, picks a step (look something up, update a record, draft a reply), checks the result, and carries on until the job is done or it hands over to a person.

What separates it from the rest of the AI you’ve met is who decides what happens next.

A chat answers your question and stops. An assistant helps inside your tools when you ask. A workflow follows steps set in advance, with AI at some of them. Only an agent chooses its own steps. That’s how the people who build these systems define it, from Anthropic and OpenAI to Gartner.

02What would one do in a business like yours?

Five examples, one per kind of business. Each shows what runs on its own and where a person still decides, and each links to our console, where you can watch it run for a demo business.

• Agency or consultancy. A prospect books a call. The agent researches the company, works out what they’re likely to need, and drafts the proposal from your similar past work. You decide: the scope and the price. See it run ›
• Home services. A call comes in after hours. It’s sorted into emergency or tomorrow, booked onto the right technician’s day, and the customer gets a text. A person decides: anything flagged as an emergency. See it run ›
• Online shop. “Where’s my order?” is answered from Shopify and the carrier; returns are logged and the label sent. You decide: any refund over your limit. See it run ›
• Logistics or wholesale. Purchase orders arrive as emails and PDFs. They’re read, checked against stock and price lists, and entered. A person decides: anything that doesn’t match. See it run ›
• Content studio. One recorded talk becomes a month of posts, a newsletter section and a blog draft, each checked against the author’s voice. An editor decides: what goes out. See it run ›

Look closely and only some of those steps need an agent. Researching a prospect does; entering a purchase order mostly doesn’t. Which raises the question for your own processes.

03Does your process need one?

Answer five questions about one process. You’ll see whether it needs an agent, a workflow with AI, plain automation or nothing yet, and roughly what it’s worth.

Whatever it said, the same few worries come next. Here they are, with the answers first.

04How do you stay in control of it?

By deciding, before it runs, what it may do on its own and what always comes to a person. Control is designed in from the start, not bolted on after something goes wrong:

• What it can do by itself, and what always goes to a person: refunds over a limit, anything a customer will see, anything that touches your books.
• What it can read and what it can change, set system by system.
• A log of every step, so you can see what it did and why.
• A person on the exceptions, whose decisions become the rules it follows next time.

Those limits matter because it will sometimes be wrong. The AI underneath is very good, not perfect, and an agent makes many small decisions in a row. Andrej Karpathy, a founding member of OpenAI, calls getting agents from usually right to reliably right a long grind. With limits like these, a mistake reaches a person instead of a customer.

Who’s responsible when it’s wrong? You are, as with any employee or supplier, which is why the limits come first. A supplier who can’t show you them hasn’t finished the job.

05Is your data safe with it?

Yes, when two things are true.

• It runs on a business account. Business accounts of the main AI services are set up not to train on your data by default; free and personal accounts often work differently. Keep company data off personal accounts.
• It can only reach what you allow. We set it out per system as reads · writes · never: what it can look at, what it can change, and what it never touches. Access is by invitation, never by sharing a password, and you can remove it at any time.

06What happens to your team?

They lose the repetitive part of the job, not the job. The receptionist stops retyping call notes; the office manager stops chasing the same invoices; the account manager stops starting every proposal from a blank page. What’s left is the part that needed a person anyway: the judgement calls, the customers, the exceptions.

It goes better when the people who do the work help shape it, because they know where the process really bends. We build with them, train them, and fix what they find in the first 30 days at no charge.

07Doesn’t the software I use already have AI?

Most of it does now, and some of it is genuinely useful:

• HubSpot has Breeze agents for support, prospecting and CRM data, built on HubSpot’s own records.
• Microsoft 365 has Copilot across Word, Excel, Outlook and Teams.
• Shopify has Sidekick in your admin, which proposes changes to your store for you to approve.
• ServiceTitan has a voice agent that answers calls and books into your ServiceTitan schedule.
• QuickBooks and Xero have AI that categorises, reconciles and drafts invoices.

Whether that’s enough depends on which of three situations you’re in:

• The job lives inside one tool, and the tool fits. Use its AI. It may be all you need, and we’ll say so.
• The job crosses tools. The enquiry starts in HubSpot, becomes a job in ServiceTitan, an invoice in QuickBooks and a review request after the visit. Each built-in AI sees its own part; nobody sees the whole job. That path between them is what we build, with one view of it and rules you can change yourself.
• The tool itself is the problem. A CRM you pay for per seat and fight every day; five subscriptions, each doing a slice. Then software you own, built around how you work, often costs less and does more of what you actually need than the tool it replaces.

Use the AI in your tools where it fits. Where it doesn’t, build what does.

08How do you tell a real agent from a relabelled chatbot?

Because agent sells, it gets stuck on things that aren’t agents. Gartner calls it agent washing: chatbots and older automation relabelled. In 2025 it estimated that of the thousands of vendors claiming agentic AI, only about 130 were building the real thing. Five questions sort it out:

• Does it decide its own next step, or follow a script?
• Can it act in your systems, or only talk?
• Does it check its own result, and try again when something’s wrong?
• Does it know when to stop and hand over to a person?
• Can you see what it did, step by step, and what each step cost?

A no to the first two means it’s a chatbot or an assistant. That may be exactly what you need; it just shouldn’t be priced like an agent. Then ask how it’s priced (per user, per conversation, per resolution or per build), and what happens to the bill when you get busier.

What agents cost, from published prices, and what it takes to make one pay: How much does an AI agent cost? ›

09What’s actually under the hood?

• Is calling an AI through its API the same as an agent? No. An API is how one piece of software asks another for something. When software sends text to an AI model through its API, it gets text back: one exchange. An agent is a program that makes many of those calls in a loop, uses tools in between, and decides each next step from what came back.
• Can you train an agent on your business? Not in the way people usually mean. The AI model is trained by the lab that makes it and doesn’t change as you use it. The agent learns your business from what you give it: instructions, your documents, notes from past cases, and the tools it may use.
• Is a chat in Claude an agent? A plain chat isn’t. The same AI acts as an agent when it’s given tools and a goal: searching, working through files, or writing and running code until a task is done.

The technical comparison, one customer request handled four ways: AI agent vs chatbot vs workflow vs custom code ›

10Why should you take our word for it?

You shouldn’t have to. We build these systems for small businesses, and we’d rather show you than tell you:

• You can click through a working example now. Our demo console runs common processes for demo businesses end to end: the agents, the rules, the exceptions and what each step cost. Yours would be your own free prototype, built around your process and your tools.
• Our prices are published, and so is what moves them. First builds from $5,000, and on average a build pays for itself in about 12 weeks on your own numbers. Work out yours in our calculator ›
• We start from the process, not the tool. Sometimes the process changes and the software gets smaller.
• You own what we build: the code, the data and the accounts are yours.
• We’ll tell you when you don’t need us. The check above says no AI needed or not yet when that’s the answer. So do we.

11Where should you start?

With one job: the one that eats the most hours, or the one where a mistake costs most. Write down how it actually runs today, step by step, including where it waits. That alone often shows a quick win.

Then try it before you commit. We’ll build a free, clickable prototype of that one job around your own tools, with the price and the payback worked out on your numbers. If it needs an agent, you’ll see where. If it doesn’t, we’ll say so.