Chatbots answer, agents act: a plain-English guide to agentic AI

For most businesses, the first encounter with AI was a chatbot: a box in the corner of a website that answers questions. Useful, sometimes. But when the customer asks it to actually do something — change a delivery address, issue a refund, book a call — the chatbot usually hands over to a human or, worse, pretends.

An AI agent is the next step. It doesn't just respond; it works towards a goal. That one shift changes what you can hand over to software.

The four things that make something an agent

Strip away the marketing and an agent is a language model with four extra abilities:

• A goal, not a prompt. You tell it the outcome you want ("resolve this ticket", "qualify this lead"), not the exact words to say.
• Tools. It can call real systems: look up an order in Shopify, update a deal in HubSpot, send an email, query a database.
• A loop. It plans a step, takes it, looks at the result and decides what to do next — repeating until the goal is met or it gets stuck.
• Memory. It remembers what happened earlier in the task, and often what happened with this customer last month.

A chatbot has, at most, the first half of one of those. That's why chatbots feel clever but rarely save much time.

A concrete example

A customer writes: "My order hasn't arrived and I need it by Friday."

A chatbot replies with a link to the shipping policy.

An agent looks up the order, checks the carrier's tracking, sees the parcel is stuck at a depot, checks your policy on reshipping, confirms stock, creates a replacement order with express shipping, emails the customer the new tracking number and leaves a note on the original order for your team. If the order value is above a limit you set, it pauses and asks a person to approve the reshipment first.

Same model underneath. Completely different outcome.

Why now?

Agents aren't a new idea, but three things changed recently. Models became reliable enough at following multi-step instructions. They became good at using tools through structured function calls rather than guessing. And the cost per task dropped far enough that it makes sense to let an agent take ten small steps instead of one.

What agents are good at — and what they aren't

Agents shine on work that is frequent, rule-guided and spread across several tools: triaging tickets, researching leads, processing invoices, reconciling spreadsheets, preparing reports. This is the work your team describes as "important but mind-numbing".

They are a poor fit, at least on their own, for decisions that are rare, high-stakes and mostly judgement: pricing a major contract, handling a crisis, deciding who to hire. In those cases the right pattern is an agent that prepares everything — research, drafts, options — and a person who decides.

A good rule of thumb: if you could write a clear training document for a new hire to do the task, an agent can probably learn it too.

The part people forget: governance

An agent that can take actions can also take wrong ones. Before any agent goes live, you need to decide three things: which actions it may take on its own, which need a human's approval, and what it should never do. Then you need a log of everything it did, so you can see why.

That isn't a reason to avoid agents. It's simply the difference between a demo and a system you can trust with customers.

Where to go from here

If you're curious whether an agent could take something off your team's plate, start by listing the tasks that are repeated dozens of times a week and touch more than one tool. That list is usually the best map of where agents will pay off first. In our next article we share a simple four-question test for picking the right one.