What is an AI agent? A plain-English guide for business owners

What is an AI agent, and does your business need one? You have probably heard that agents will handle your inbox, your leads and your admin. Some of that is true and some is sales talk. This guide explains how agents work, where they help, where they fail, and how to check whether you are ready.

What is an AI agent, in plain English?

An AI agent is a program that uses a language model to decide what to do next, then does it through the tools it is connected to. IBM describes an agent as a system that autonomously performs tasks by designing workflows with available tools. The key word is performs. The agent does not just write text for you to copy. It reads, decides and acts.

Think of it as a junior assistant with a clear job description: a goal, tools it may use, rules about what it may not do, and a record of what it has done. When it hits something outside its rules, it stops and asks.

The four parts of every AI agent

• A model that reads the situation and chooses the next step. This is the language model, the same kind of technology behind AI chat tools.
• Tools it can call: your email, calendar, CRM, helpdesk, spreadsheet or order system, each connected through an API with set permissions.
• Instructions and rules: the goal, the playbook, approval limits and the cases it must hand to a person.
• Memory and logs: what it has done so far in this task, and a record you can audit later.

How does an AI agent work, step by step?

An AI agent works in a loop: it receives a trigger, plans, acts through a tool, checks the result and repeats until the goal is met or a rule tells it to stop. That loop is what separates an agent from a one-shot AI reply.

Here is a simple example. A new enquiry arrives through your website form at 9:40 at night.

1. Trigger: the form submission reaches the agent.
2. Plan: the agent reads the enquiry and decides it is a sales lead for your consulting package, not a support request.
3. Act: it checks your CRM, finds no existing contact, and creates one with the details from the form.
4. Act again: it drafts a reply with two open meeting slots taken from your calendar.
5. Check: your rules say first replies to new leads need approval, so it places the draft in a review queue instead of sending it.
6. Log: every step is recorded, so you can see what it did and why.

IBM notes that agents can create subtasks on their own to reach a complex goal, and that an agent’s ability to store past interactions in memory and plan future actions shapes what it does next.

AI agent vs chatbot vs automation: what is the difference?

A chatbot answers, a traditional automation follows fixed rules, and an AI agent makes judgment calls inside rules you set. Gartner gives the same split as advice: use agents where decisions are needed, automation for routine workflows and assistants for simple retrieval. It also warns about agent washing, the rebranding of existing chatbots, assistants and automation tools as agents, and estimates that only about 130 of the thousands of agentic AI vendors are real.

	AI chatbot	Rule-based automation	AI agent
What it does	Answers questions in a conversation	Moves data when a trigger fires	Works toward a goal across several steps
Handles new situations	Only within its content	No, it follows the rule or fails	Yes, within its rules and tools
Takes actions in your systems	Rarely	Yes, the same action every time	Yes, chosen case by case
Main risk	Wrong answer	Breaks when inputs change	Wrong action
Typical example	Website FAQ bot	New order copied to a spreadsheet	Lead triaged, CRM updated, reply drafted

A quick way to tell which one you need is the judgment test. Write the process out step by step. If every step can be written as “if this, then that”, you need automation, which is cheaper and more predictable. If at least one step needs reading and judgment, such as “is this a sales lead or a complaint?”, an agent may earn its cost. If nothing needs to change in any system, a chatbot is enough. Our comparison of AI chatbots and AI agents for customer service applies this to customer questions with a 100-ticket test.

What can AI agents do for a small business today?

AI agents work well today on repeatable, multi-step office work where the rules are clear and a mistake is small or easy to undo. They are weaker on one-off judgment calls, relationships and anything where the right answer is not written down anywhere.

• Inbox triage: reading a shared inbox and sorting messages by type and urgency before anyone opens them.
• Project follow-up: watching tasks and deadlines, nudging owners and posting status before you ask.
• Lead handling: qualifying enquiries, updating the CRM and drafting first replies for approval.
• Order and support questions: answering from live order data and making small, approved changes.
• Back-office steps: moving records between systems, checking them against what is already there, and flagging mismatches.

A worked example: is a lead-reply agent worth it?

For illustration: say your business receives 120 website enquiries a month. Logging each one in the CRM, checking for duplicates and writing a first reply takes a person about 6 minutes. That is 720 minutes, or 12 hours a month, and replies sent after hours wait until the next morning.

An agent that logs the lead, checks the CRM and drafts the reply cuts the human part to a 1-minute review per enquiry: 120 minutes, or 2 hours a month. If that person’s time costs the 10 hours saved are worth about $400 a month. Set that against the running cost, which for our single-process agent builds is typically $50 to $250 a month in model usage and hosting. Speed is the other gain: the draft is ready minutes after the enquiry lands. Run the sum with your own numbers before you buy.

If the sum comes out thin, the process may not be the right first candidate. Our scoring method for choosing which processes to automate first helps you compare options, and the guide to AI running costs shows what goes into the monthly bill.

Smaller firms are moving more slowly than large ones. McKinsey’s State of AI in 2026 survey, published on 25 August 2026, found that 40% of organizations with more than $1 billion in revenue were scaling AI agents, up from 27% a year earlier. Among smaller organizations, the share stayed flat at 22%. Starting with one well-chosen process is how a small business closes that gap without a large team.

What are the risks, and how do you keep an agent in check?

The main risk is an agent taking a wrong action with real consequences, and the main defense is to limit what it can touch and require a person to approve anything high-impact. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.

The OWASP security project lists excessive agency as a top risk for AI applications. It names three root causes, and each one has a plain fix:

Root cause	What it means	The fix
Excessive functionality	The agent can use tools it does not need	Give it only the tools for its one job
Excessive permissions	Its access is wider than the task	A dedicated account with the minimum permissions needed
Excessive autonomy	High-impact actions run without a check	A human approval step before they take effect

OWASP’s own advice is to use human-in-the-loop control so a person approves high-impact actions before they are taken. NIST makes a related point: human-AI configurations can span from fully autonomous to fully manual, and human roles in decisions and oversight need to be clearly defined. Our guide to human-in-the-loop AI and where people should stay in control shows how to choose the right level for each action.

Is your business ready for an AI agent? A 7-point check

You are ready for an AI agent when you can describe one process, its rules and its exceptions on a single page. Work through these seven points and mark each one pass or fail.

1. Pick one process, not a department. “Reply to new website leads”, not “sales”. Pass: it runs often enough to matter, as a rule of thumb at least 30 times a month.
2. Write down the steps a person follows today, with each system they open and each decision they make. Pass: the steps fit on one page and two people who do the job agree on them.
3. List the rules and the exceptions: approval limits, refund caps, who owns what, and the cases that always go to a person. Pass: every exception has a named owner.
4. Collect 10 real examples: real emails, tickets or requests. Pass: at least 3 of them are awkward cases, because those become the hardest tests.
5. Check access. Pass: every system involved has an API or admin access you control, and you can create a limited account for the agent.
6. Mark where a person signs off. Pass: every step that sends money, contacts a customer in a new way, or cannot be undone within a day has an approver.
7. Record a baseline. Pass: you have two or three measures, such as time to first reply or hours spent, written down before anything is built.

Seven passes: you are ready to scope a build. Five or six: fix the gaps first, which are usually points 2 and 3. Four or fewer: start with a simpler automation or a tidy-up of the process itself. If you are stuck at point 1, our buyer’s checklist for AI services without scope creep helps you narrow a big idea into a buildable first project.

How Upmark Labs can help

We build AI agents as fixed-scope packages, with scope and price shown on a Scope Card before you pay, built in your own accounts so you own the workflows, data and AI provider relationship.

A first agent usually means one process handled end to end. That is what our AI Multi-Step Workflow Automation build does, from $6,900. It runs one business process across up to two systems with read and write access, makes AI-assisted decisions inside your approval rules, and sends anything outside those rules to a named approver.

After you order, you provide access and information within 72 hours: admin or API access to the systems involved (test records are fine), your rules and approval limits, 10 real examples, and the name of the person who approves actions. The build runs to a fixed plan of 20–30 working days, including 50 documented test scenarios, your review with two rounds of changes, a recorded 60-minute handoff and 30 days of support. Express and Rush delivery times are shown in the price panel before you order. Optional monthly maintenance follows.

It is not the right choice for everything. If your process is really one spreadsheet updating from one source, or you only need alerts routed, a smaller automation costs less, and the AI agents store has single-job agents for things like project follow-up. If your 7-point check came out at four or fewer, start with a conversation, not a build. You can see the full process on how it works and compare packages on our pricing page.

Frequently asked questions

Start with the 7-point check above. Then book a free expert consultation and bring your answers: we will tell you which agent package fits, or whether a simpler automation would do the job. If you need several agents working together across your business, use custom AI for projects above $15,000.