What is the difference between an AI agent and a chatbot?

Three things: a trigger, tools, and a finished output. A chatbot waits for a person and returns text. An agent is started by an event, can read and write in the systems it has been given access to, and its output is a completed task such as a scheduled appointment, an updated record or a sent document.

How much does it cost to run an AI agent?

The model usage behind a single well-scoped agent for a small business is typically a few dollars to a few tens of dollars per month, on top of whatever the surrounding tools cost. The real cost is the build: defining the task precisely, connecting it to your systems, and testing it against real work before it runs unattended.

Article · 8 minute read

What is an AI agent, really?

Q: What can an AI agent actually do in a small business?

Realistically: intake and qualification of new enquiries, drafting and scheduling content, follow-up sequences, meeting notes turned into tasks and summaries, invoice and document chasing, weekly reporting assembled from systems you already use, and first-line customer questions answered from your own documentation. The common thread is repeatable work with a clear finished state.

Q: Where do AI agents still fail?

Agents fail on tasks with no defined finished state, on judgment calls the business has never written down, and when given broad credentials without a human checkpoint. They also fail quietly: an agent that produces plausible but wrong output every time is worse than one that stops, which is why every deployment needs a defined scope, narrow credentials and a review step.

By Chaka Serrant August 13, 2026 Fundamentals

An AI agent is software that has been given a job, access to the tools needed to do that job, and permission to act on its own when something triggers it. A chatbot answers when spoken to. An agent starts work because a form was submitted, a date arrived or a file landed, and it finishes the job without being prompted at each step.

What is the difference between an agent and a chatbot?

Three things, and all three have to be present before the word agent means anything.

A trigger. Something other than a human typing starts it. A booking is made, an invoice hits fourteen days overdue, Monday morning arrives.
Tools. It can read and write in specific systems you have given it access to. Your calendar, your CRM, a folder, an inbox. Without tools it can only produce text about the work.
A finished output. Not a suggestion. An appointment on the calendar, a record updated, a document drafted and filed, a summary sitting in your inbox.

Take away the trigger and you have a very good assistant that still needs you to remember to ask. Take away the tools and you have a writer. Take away the finished output and you have a meeting about automation.

The test is simple: if it only produces words about the work, it is not an agent. If the work is done when it stops, it is.

What can an AI agent actually do in a small business?

Setting aside the demos, here is the work that agents genuinely run well today in businesses under fifty people. The common thread is repeatable work with a clear finished state.

Intake and qualification. A new enquiry arrives, gets a reply within minutes, is asked the right qualifying questions, and lands in your CRM with the answers attached and a summary written.
Content production. One recorded idea becomes the week of posts, cut per platform, scheduled, with the caption rules each platform needs already applied.
Follow-up. The sequences that stop when someone replies and resume when they go quiet, which is the part humans reliably fail at once they are busy.
Meetings into motion. A recording becomes notes, decisions, and tasks assigned in the tool your team actually uses, before anyone has left the call.
Chasing. Documents, signatures, overdue invoices. Politely, on schedule, without you having to feel like the person who nags.
Reporting. The weekly numbers assembled from the systems that already hold them, written up the same way every week.
First-line answers. Common customer questions answered from your own documentation, with anything unfamiliar handed to a person rather than guessed at.

Each of those is one job with a defined start and a defined end. That is not a limitation to work around. It is the design, and it is why they work.

The shape of one agent

A new enquiry agent, described the way it has to be built

Trigger: a form submission on the website.
Inputs: the form fields, plus your service list and pricing rules.
Tools: read and write on one CRM pipeline, send from one mailbox, read a calendar.
Steps: create the contact, send the acknowledgment with the booking link, request the two intake documents, write a three line summary, tag the deal by service type.
Escalation: anything that does not match a known service goes to a human, unanswered.
Finished state: contact exists, reply sent, documents requested, summary written.

Notice how much of that is specification rather than technology. Vagueness at this stage is the single most common reason an agent underperforms.

Where do AI agents still fail?

Worth being blunt about, because the failure modes are predictable and mostly avoidable.

No defined finished state. “Handle marketing” is not a job an agent can complete. “Draft next week's posts by Friday at 9am” is.
Undocumented judgment. If the rule lives only in your head, the agent will invent one. It will be reasonable, consistent, and not yours.
Too much access. An agent given broad credentials and no checkpoint can do a lot of damage quickly and politely. Scope credentials to exactly the systems the job needs, nothing more.
Quiet wrongness. This is the one that actually costs money. An agent that fails loudly gets fixed on day one. An agent that produces plausible, wrong output every week gets discovered in month four. Build the review step before you build the automation.

Every one of those is a design decision rather than a technology limit, which is the useful part: they are all fixable in advance by somebody who has seen them before.

How much does an AI agent cost to run?

Two separate numbers, and conflating them is why the question gets confusing answers.

Running cost is small. The model usage behind a single well-scoped agent in a small business is typically a few dollars to a few tens of dollars a month, on top of whatever the surrounding tools already cost you. Frontier capability has been getting cheaper per unit of work every quarter, which means automation you priced out as too expensive a year ago is worth re-pricing now.

Build cost is where the real money sits: defining the task precisely, connecting it to your systems, and testing it against real work before it runs unattended. That is a one-time cost per agent, and it is the reason the order you build them in matters so much.

How many agents does a small business need?

Fewer than the marketing implies. Most owners under fifty people get the majority of the available benefit from three to five agents covering intake, follow-up, content and reporting. Building twelve is a way of spending money on the ones that were never going to matter.

Build one, run it for a month, measure it, then build the next. The sequence is the strategy. A business with one agent that reliably saves six hours a week is materially better off than a business with nine that each save twenty minutes and all need supervision.

The ten jobs worth starting with

The free 10 AI Agents to Replace Your $30K/yr VA guide names ten of these agents by function, says what each one replaces and what it costs to run. No card required.

If you would rather know which of them applies to your specific business and in what order, that is what the Handover Assessment produces. And if the question you are actually asking is how much of the business still depends on you personally, start with what to automate first.

Get the 10 AI Agents guide All free resources