What an "AI Agent" Actually Is, Without the Hype

Every software company now has an “agent.”

Your booking tool has one. Your email app has one. The person who messaged you on LinkedIn last week is selling one that will “run your business while you sleep.” A year ago the same products were called assistants, copilots or chatbots. The software mostly didn’t change. The label did.

If you run a small business, this is confusing in a very practical way. You’re being asked to pay for something, and you can’t tell what it is.

So let’s take the hype out. This post explains, in plain English, what separates a chatbot, an automation and an agent. Why most things called agents are something simpler, and why simpler is often what you want. Where real agents go wrong. And the questions to ask before you pay for one.

Three things that get called “AI”

Almost every AI product you’ll be sold falls into one of three buckets. The difference is not how smart the model is. It’s who decides what happens next.

1. A chatbot: it talks

A chatbot answers when you ask. You type a question, it types an answer. ChatGPT in a browser tab is the classic example.

A chatbot can be very useful. It can draft a reply to a guest, explain a tax term or turn your notes into a tidy email. But on its own it doesn’t do anything in the world. It doesn’t send the email. It doesn’t update your calendar. You read its answer and then you act.

You decide every step. The chatbot just helps you think.

2. A workflow: it follows your steps

A workflow (people also say “automation”) is a fixed list of steps that runs on its own when something happens. Tools like Zapier, Make and n8n are built for this.

Here’s an example. When a new booking arrives, add a row to my spreadsheet, then send the cleaner a WhatsApp message with the date.

Nothing here is “intelligent.” It’s a recipe. The same trigger always leads to the same steps, in the same order. That’s the point. You designed it once, and now it runs the same way every time.

Many workflows now have an AI model in the middle of them. When a guest message arrives, ask the model to sort it into “urgent”, “question” or “thanks”, then put the urgent ones at the top of my list. The model does one small, clearly defined job inside a path that you drew.

You decide the steps. The AI fills in one or two of them.

3. An agent: it picks its own steps

An agent is different in one important way. You give it a goal, not a recipe.

A real agent has four things:

A goal. “Find a replacement cleaner for Saturday.”
Tools. Things it’s allowed to use: your contact list, your calendar, WhatsApp, your booking system.
It decides the steps itself. Nobody wrote “first check the calendar, then message Lan, then message Hoa.” The model looks at the situation and chooses.
It loops until it’s done. It tries something, looks at the result, and decides what to do next. Lan doesn’t reply? Try Hoa. Hoa says she can only come at 3 p.m.? Check if that works with the next check-in. Keep going until the goal is met, or until it gives up.

That loop is the heart of it. A workflow runs once, top to bottom. An agent thinks, acts, checks, and thinks again, possibly many times, and you don’t know in advance how many steps it will take.

Anthropic, one of the big AI model companies, draws the same line in its own guide for builders: in a workflow, the model and tools follow paths written in advance by people, while in an agent, the model directs its own process and decides how to use its tools. That’s a good definition to keep in your head.

A ladder, not a switch

It helps to think of this as a ladder. Each step up gives the software more freedom to decide. More freedom means it can handle messier situations. It also means more cost and less predictability.

Each step up the ladder hands more decisions to the software, and with them more cost and less predictability.

Two things about this ladder matter for a business owner.

Higher is not “better.” It’s just different. A hammer isn’t worse than a robot. For hanging one picture, it’s much better.

Most of your work belongs on the middle steps. The jobs that eat a small business owner’s day are usually repeated and predictable: confirm the booking, tell the cleaner, answer the parking question, chase the invoice. A predictable job wants a predictable tool.

Most “agents” are workflows with a model inside

Here’s the honest truth about the market right now: a lot of products called agents are really on the third step of the ladder. A fixed workflow, with an AI model doing one part of it.

That isn’t just my opinion. In June 2025, the research firm Gartner gave this a name, “agent washing”: rebranding existing chatbots, assistants and automation tools as “agentic” without much real agent capability. Gartner estimated that only about 130 of the thousands of vendors selling “agentic AI” were the real thing. It also predicted that over 40% of agentic AI projects could be cancelled by the end of 2027, because of rising costs, unclear business value or weak risk controls.

You might expect me to say this is bad. Mostly, it isn’t.

For a small business, a workflow with a model inside is often exactly the right product. It’s:

Predictable. It does the same thing every time, so you can trust it after a week of watching it.
Cheaper. One or two calls to an AI model per task, not dozens.
Easy to fix. When it goes wrong, you can see which step broke.
Easy to explain. You can tell your cleaner or your partner exactly what it does.

The problem isn’t that these products are workflows. The problem is when the label makes you expect something they don’t do, or pay a price that only makes sense for something they aren’t.

So don’t be disappointed when you find out the “agent” is a workflow. Be disappointed if nobody told you.

Where real agents go wrong

Real agents, the ones that pick their own steps, are impressive when they work. But the same freedom that makes them flexible makes them risky. Here are the three failure modes I’d want any owner to understand.

1. Wrong actions, done confidently

An agent with tools can do things: send messages, change bookings, issue refunds, delete rows. If it misunderstands the goal, it doesn’t just give you a wrong answer you can ignore. It takes a wrong action you have to clean up.

And small mistakes compound. In a loop of ten steps, an early misunderstanding gets built on by every step after it. Anthropic’s guide warns about exactly this: the potential for compounding errors, and the need for testing with guardrails.

Here’s an example. Imagine an agent told to “reply to guests who asked about early check-in.” It reads a message that says “Can we check in early? If not, no worries, we’ll cancel and find somewhere closer to the airport.” It reads the word “cancel,” decides the guest wants to cancel, and starts the cancellation. That’s a made-up case, but it’s the kind of mistake language models really do make, and an agent is the kind of system that can act on it before anyone looks.

2. Cost you can’t predict

Every time an agent “thinks,” it calls an AI model, and each call costs money. A workflow might use one or two calls per task. An agent might use five, or fifty, depending on how hard the task turns out to be and how many times it tries.

That means the cost per task isn’t fixed. Agents trade speed and money for flexibility. If you’re paying per use, ask what happens on a busy week. If you’re paying a flat fee, ask how the company makes that work, because a price that loses money on every heavy user tends to change.

I go deeper into this in the hidden costs of AI tools.

3. Different results on different days

Give an agent the same goal twice and it may take two different paths. Most days that’s fine. But when a customer complains, you need to know what happened and why. If the tool can’t show you the steps it took, you’re left guessing.

Predictability is underrated. For a small business, a tool that’s 90% as clever but does the same thing every day is often worth more than a brilliant tool that surprises you.

Why we chose to keep the host in charge

When we designed Santara, we had this exact choice in front of us.

Santara reads a host’s bookings and guest messages from Airbnb, Booking.com and WhatsApp, and sends one morning briefing: who’s arriving, who’s leaving, which units need cleaning, and which messages still need a reply. Each item links back to its source. When something is unclear, it says “I’m not sure” instead of guessing.

What it doesn’t do is act on its own. It doesn’t reply to guests for you, change bookings or message your cleaner without you. The host makes the call.

That was a deliberate choice, not a missing feature. Guest messages are full of exactly the kind of ambiguity that trips agents up, and a wrong reply to a guest costs real money and real reviews. A host who checks a short list and decides is slower than a fully autonomous agent on paper. In practice, a host who trusts the list is the whole point. I wrote more about that thinking in why Santara sends one briefing, not a chatbot.

You don’t have to agree with where we drew the line. But every AI product you buy has drawn one somewhere, and you should know where.

Questions to ask a vendor

When someone sells you an “AI agent,” you don’t need to understand the technology. You need answers to these questions. A good vendor will answer them simply. A vague answer is an answer too.

“Does it follow fixed steps, or does it decide its own steps?” Either answer can be fine. You just want to know which one you’re buying.
“What can it do without asking me?” Get the actual list. Send messages? Spend money? Change or delete records? Anything on that list can go wrong without you seeing it first.
“Can I make it ask before it acts?” The best tools let you start with “draft and wait for approval” and loosen up later.
“Can I see what it did and why?” You want a history of every action, with the message or record it was based on.
“What happens when it’s not sure?” Does it stop and ask, or does it guess?
“Can I undo it?” A sent message can’t be unsent. A deleted booking might not come back.
“How is it priced when I’m busy?” Per task, per message, per seat, flat? What would my busiest month cost?
“Can I try it on my own data for a week?” Not their demo. Your messy inbox, on a normal Tuesday. (More on this in how to tell if an AI product is real or just a demo.)

If you only ask one, ask number 2. What it can do without you is what you’re really trusting it with.

When you should want an agent, and when you shouldn’t

So, should a small business want a real agent at all? Sometimes. Here’s a simple way to decide.

You probably want a workflow (maybe with AI inside) when:

The job is the same every time. New booking, new invoice, new lead. Same steps, every time.
Mistakes are expensive or public. Anything that touches customers, money or reviews.
You can describe the steps. If you can write the recipe on a napkin, a workflow can run it.
You want a fixed monthly cost.

That covers most of the busywork in a small service business. It’s also where you’ll get the fastest, safest win. If you’re not sure where to start, your spreadsheet plus one well-built automation goes a surprisingly long way.

An agent can make sense when:

The steps really change each time. Research, comparing many options, digging through scattered information.
A wrong move is cheap and easy to undo. It drafts, searches, sorts or suggests, rather than sends, pays or deletes.
A person checks the result before it matters.
The time saved is worth the variable cost.

A good example is research. “Find three local suppliers for bed linen, compare their prices and delivery times, and give me a short table.” The steps are different every time, nothing is sent or bought, and you read the table before you decide anything. That’s a fine job for an agent.

And a useful rule in between

Even when you use an agent, keep the final action with a person, at least at first. Let it gather, draft and propose. You approve. After a month of it being right, you can decide whether to let it do more. Trust should be earned, by software as much as by a new employee.

The takeaway

An AI agent isn’t magic and it isn’t a scam. It’s a specific thing: software that’s given a goal and tools, decides its own steps, and loops until it’s done.

Most products with “agent” on the label are something simpler: a workflow with a model inside. For a small business, that’s often the better choice. It’s predictable, cheaper and easier to fix.

Before you buy, find out where the product sits on the ladder, what it can do without asking you, and whether you can see what it did. And remember what you’re actually trying to buy. Not an agent. Fewer tasks on your plate.

Sources: Anthropic, “Building effective agents” (anthropic.com/engineering/building-effective-agents), checked October 2026; Gartner press release of 25 June 2025, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” as reported by W.Media (w.media), checked October 2026.