What Is an AI Agent? A Plain-English Guide for Business Owners

An AI agent is software that pursues a goal by taking actions (opening websites, reading and writing files, sending messages, updating records) and checking the results as it goes, rather than just answering a question. A business could use one to reply to routine enquiries, reconcile bank transactions or research suppliers, with a person approving anything that matters.

The difference from a chatbot is the difference between advice and action. Ask a chatbot how to compare three suppliers and it tells you. Ask an agent and it visits the three websites, pulls the prices into a spreadsheet and comes back with the file. That's what makes agents useful, and it's also what makes them riskier: an agent that can act can act wrongly, and do it several times before anyone notices.

Follow me on Instagram@sagnikteaches

For a small business, agents arrive in two ways. Some are general-purpose, such as ChatGPT Work or Claude in Chrome, where you describe a task and the agent works out the steps. Most, though, turn up inside software you already use: an accounting package that reconciles transactions on its own, a CRM that answers customer questions, a messaging app that replies to enquiries. You may be closer to using one than you think.

Connect on LinkedInSagnik Bhattacharya

What makes software an agent rather than a chatbot

A useful way to picture an agent is a capable temp on their first day, given a login and a task list. Four things make that temp effective, and the same four make an agent:

Subscribe on YouTube@codingliquids

1. A goal. "Build a spreadsheet comparing these suppliers' prices for our 20 most-ordered items", not "help with suppliers". The clearer the finish line, the better the agent does.
2. A brain. A large language model that reads the goal, decides the next step, and reads what happened after each one. This is the same kind of model that powers a chatbot; the difference is what it's allowed to do.
3. Tools. The things it can use: a web browser, your files, your email, your CRM, a spreadsheet. An agent with no tools is just a chatbot. Every tool you add is a new way for it to help and a new way for it to go wrong.
4. A leash. The limits: which sites and apps it may touch, which actions need your approval, when it must stop and ask. A good agent product makes these limits visible and adjustable.

The engine that ties them together is a loop: plan a step, take it, look at the result, decide the next step, and repeat until the goal is met or it gets stuck. A chatbot answers once and waits for you. An agent keeps going.

One task, followed step by step

Here's what that loop looks like in practice. A physiotherapy clinic was choosing new booking software and asked ChatGPT Work to do the legwork:

Compare three clinic booking systems: [vendor A], [vendor B] and
[vendor C]. From their own websites, get the monthly price for a
clinic with 5 practitioners, whether they send SMS reminders, and
whether they integrate with Xero. Put it in a spreadsheet with a
source link for each figure. Don't sign up for anything.

An illustrative log of what the agent did:

Step 1  Plan: visit each vendor's pricing page, then feature pages.
Step 2  Vendor A pricing page: per-practitioner price found.
Step 3  Vendor B pricing page: prices hidden behind "Book a demo".
        Searched help pages; found a published price list PDF.
Step 4  Vendor C: pricing shown in another currency. Noted, did not
        convert; flagged for you.
Step 5  Feature pages: SMS reminders (A yes, B yes, C add-on).
        Xero integration (A yes, B yes, C not found - marked
        "not found", not "no").
Step 6  Built spreadsheet with 3 rows, 6 columns, source links.
Step 7  Stopped. Did not start any free trials.

Two things in that log show a well-behaved agent. At step 4 it flagged the currency problem instead of guessing an exchange rate, and at step 5 it recorded "not found" rather than "no", which are different claims. What the clinic still had to do was check the three prices against the vendors' pages (one had changed since the agent visited) and ask vendor B directly, since a price list buried in a PDF may not be current. The agent saved about two hours of clicking; it didn't make the decision.

Agents a small business can already switch on

Two details in that table matter more than the prices. First, "per resolution" pricing depends on how the vendor defines a resolution: Intercom counts an "assumed resolution" when the customer goes quiet for 24 hours after Fin's last answer, and silence isn't the same as satisfaction. Second, several of these agents are gated to higher plans, so the real cost can include an upgrade. Ten AI agent use cases for small businesses works through examples with before-and-after numbers.

How the leash works in two real products

Claude in Chrome shows what good limits look like. Anthropic offers three permission modes: "Manually approve", where Claude asks before each action; "Automatically approve", where it keeps working but reviews each action for safety and pauses when needed; and "Skip all approvals", which Anthropic says to use only when you completely trust everything involved. Some actions, such as making purchases or creating accounts, are blocked regardless, and Anthropic advises against using it to manage financial accounts.

OpenAI takes a similar approach: ChatGPT Work needs your approval for important actions by default, and on web and mobile you approve or decline each one on screen. Its predecessor, the retired ChatGPT agent, handed the browser back to you whenever a login was needed so you typed passwords yourself, and its help page listed habits that still apply to any agent: enable only the apps a task needs, avoid vague prompts like "check my email and handle everything", and stop a task if something looks wrong.

A five-rung ladder for how much an agent may do alone

The most useful decision you'll make about any agent is how far up this ladder to let it climb. Start every new agent on a low rung and move it up only after it has proved itself on your own work.

1. Suggest. The agent researches or drafts; a person does everything else. A dental practice asking ChatGPT Work to gather three suppliers' prices into a spreadsheet sits here.
2. Draft and queue. The agent prepares actions, such as reply drafts or journal matches, and a person approves each one before it happens. Review replies and supplier emails belong here permanently for most businesses.
3. Act with step-by-step approval. The agent works through a task but asks before each action, as Claude in Chrome does in "Manually approve" mode. A nursery manager watching it fill a reorder basket on a wholesaler's website is on this rung.
4. Act within limits and report. The agent acts on its own inside set boundaries and shows you what it did. Automatic bank reconciliation, where Xero's JAX matches only what it's highly confident about and leaves the rest for you, is an example.
5. Act and flag exceptions. The agent handles the routine and only raises what it can't. A customer-reply agent answering opening hours and prices out of hours, and handing everything else to staff in the morning, runs here.

Most small-business agents should live on rungs 2 to 4. Rung 5 suits narrow, reversible, low-stakes work; anything involving money, contracts, health or a customer's complaint stays lower.

What an agent costs beyond the price tag

Agent pricing looks small per action, so do the sum for your volume before switching one on. Two quick illustrations using list prices:

• HubSpot Customer Agent. At 50 credits per resolved conversation and $10 per 1,000 credits, 200 resolutions a month use 10,000 credits, about $100. Service Hub Professional includes roughly 3,000 credits a month, and unused credits don't roll over, so the extra 7,000 cost about $70, on top of the Professional subscription itself, which HubSpot lists at $90 a seat a month on annual billing plus a $1,500 onboarding fee.
• Meta Business Agent. About 400 AI replies a month at 4-5 cents each is $16-$20. Meta's own example puts a simple question answered in four messages at 16-20 cents, so long conversations cost more.

Then add the costs that don't appear on any invoice: an afternoon writing the agent's instructions and limits, the time someone spends reviewing its work (budget 10 to 20 minutes a week for the first month), and any plan upgrade the agent requires. An agent that saves three hours a week but needs a $50-a-month upgrade and an hour of checking is still usually worth it; one that saves twenty minutes rarely is.

Four beliefs about agents that don't hold up

"An agent is just a better chatbot." The model inside may be the same; the difference is permission to act. That changes the questions you ask before using one, from "is the answer good?" to "what can it touch, and what happens if it's wrong?"

"Agents replace staff." In small businesses they replace tasks, usually the tedious middle of a job: gathering, matching, drafting. A tutoring agency that switches on automatic reconciliation still needs someone to review what the agent wasn't confident about and to deal with the parent who paid the wrong amount.

"If it's built into my software, it's safe to leave alone." Built-in agents are usually well-limited, but they still make mistakes. Xero says JAX only reconciles automatically when it's highly confident and that you can challenge or reject its matches, which is another way of saying you should look.

"More autonomy is always better." The most useful agents in small businesses run on a short leash: draft, don't send; fill the basket, don't pay; suggest the match, let a person confirm. Autonomy is something you extend once the agent has earned it on your own data.

Where agents still stumble

• Messy inputs. An agent reconciling payments against invoices is only as good as your invoice references. Garbage in, confident garbage out.
• Vague goals. "Sort out our suppliers" gives an agent nothing to finish. It will either stop early or wander.
• Websites that fight back. Logins, pop-ups, prices hidden behind demo forms and pages that change layout all trip agents up, as vendor B did in the example above.
• Instructions hidden in content. This is called prompt injection: a web page or email containing text designed to trick an agent into doing something else. OpenAI's own help page gives the example of an agent researching a restaurant booking that meets a malicious comment telling it to fetch a password reset code. The defence is limiting what the agent can reach, and approving sensitive steps yourself.
• Costs that scale with volume. Per-message and per-resolution pricing is cheap at 50 conversations a month and noticeable at 5,000. Check the definition of what's billed before you switch anything on.

For the fuller list of what can go wrong once agents start taking actions, see what can go wrong when AI agents act for you.

Agent words you'll meet, in plain English

• Agent mode. A setting in a chat product that lets it take actions rather than only reply. ChatGPT Work is OpenAI's current version.
• Tools, apps or connectors. The accounts and services an agent is allowed to use, such as your Google Drive, email or CRM. Each one is a permission you grant.
• MCP (Model Context Protocol). An open standard that lets AI assistants plug into other software in a consistent way. Airtable, for example, runs an MCP server that Claude and ChatGPT can connect to, limited to what your own account can see.
• Human in the loop. A step where a person must approve before the agent continues. The rungs 2 and 3 above are human-in-the-loop designs.
• Resolution. The unit many customer-service agents charge for. Always read the vendor's definition, because some count a customer going quiet as resolved.
• Prompt injection. Text hidden in a web page, email or document that tries to give an agent new instructions. Limiting what the agent can reach is the main defence.

Giving an agent its first job

The safest first job is one where the agent gathers or drafts and a person decides. A good brief has four parts: the goal, the finish line, the limits and the check. A veterinary practice's first brief, for ChatGPT Work, looked like this:

Goal:        Build a spreadsheet of our 25 most-ordered consumables
             with current prices from our two approved suppliers.
Finish line: One row per item: name, pack size, price per unit at
             each supplier, link to the product page.
Limits:      Only the two supplier websites. Don't log in, don't
             add anything to a basket, don't contact anyone.
Check:       Flag any item where pack sizes differ between
             suppliers; I'll compare those by hand.

Run it, check every flagged line and three unflagged ones at random, and note how long it took against doing it yourself. If it was right and faster, give it a slightly longer leash next time. If it wasn't, the brief usually needs a clearer finish line before the agent needs replacing.

Two other starting points suit most small businesses. If you already use software with a built-in agent, such as Xero on the Growing plan or HubSpot on a Professional tier, switching that agent on for a month with a weekly review is the lowest-effort trial there is. And if the job is really "do the same thing every time X happens", you may not need an agent at all; agent vs chatbot vs automation helps you tell which one a job needs. If you'd rather map the candidates with someone, that's what my AI implementation consultation covers.

AI agents: questions owners ask

Is ChatGPT an AI agent?

Ordinary ChatGPT conversations aren't; they answer and wait. ChatGPT Work is: OpenAI describes it as an agent for longer, multi-step work and finished deliverables such as spreadsheets, documents and presentations, working across connected apps and files. It replaced the older ChatGPT agent, which was retired on 9 July 2026, and it's included on Plus, Pro, Business and Enterprise plans.

Do AI agents need a developer to set up?

Not the ones most small businesses start with. Agents built into tools you use, such as Xero's JAX reconciliation, HubSpot's Customer Agent or Meta's Business Agent, are switched on in settings. General agents such as ChatGPT Work and Claude in Chrome take a plain-English instruction. Custom agents that act inside your own systems are where a developer or consultant usually comes in.

Can an AI agent spend money or sign contracts for me?

It shouldn't, and the mainstream tools try to stop it. Anthropic says Claude in Chrome won't make purchases or create accounts, and advises against using it to manage financial accounts. OpenAI's agent tools ask for confirmation before high-impact actions. Keep payments, contracts and anything irreversible as steps a person takes.

Further reads

• How to Connect ChatGPT or Claude to Your Business Apps — Connecting assistants to your apps, the step before agents act.
• Can ChatGPT Agent Handle Business Admin Tasks Unsupervised? — How far ChatGPT's agent tools can run admin unsupervised.
• HubSpot AI Agents and Breeze: What Small Teams Should Switch On — Which HubSpot agents are worth switching on.
• What Is RAG? A Plain-English Explainer for Business Owners — The retrieval idea behind agents that read your documents.
• How to Add Human Approval Steps to AI Automations — Designing the approval step every agent needs.
• AI Email Triage for Shared Inboxes: Sales, Support, and Invoices — A lower-risk place to start: triaging a shared inbox.
• How to Prepare Your Small Business for AI Agents — Six things to have ready before an AI agent acts for you, with an approval table, a photography studio's four-week prep and a checklist.
• Can AI Manage Your Calendar and Book Meetings for You? — The five kinds of AI scheduling tool, which one fits who books whom, and the diary rules to write before letting any of them near your calendar.
• Can AI Fill In Forms and Supplier Portals for You? — Browser agents, recorded workflows, integrations or AI-prepared data: how to choose for each form and portal, with a supervised-run prompt and a ten-form test.
• What Are AI Marketing Agents and Should a Small Business Use One? — What separates a marketing agent from a chatbot or automation, the agents small businesses can switch on now, what they cost and the guardrails to set.
• How to Pilot Your First AI Agent Without Risking Customers — Run your first AI agent in shadow mode, then behind approvals, then live in a narrow window. A café's catering agent shows each stage, cost and stop rule.
• What Is Prompt Injection and Should a Small Business Worry? — How hidden instructions in emails, web pages and CVs hijack AI assistants, a five-minute exposure check, and the controls that work without an IT team.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help articles on ChatGPT Work and the retired ChatGPT agent; Claude help 'Claude in Chrome permissions guide'; Xero's bank reconciliation page on JAX; HubSpot, Meta, Intercom and Microsoft facts checked on their own pages (September 2026).