AI chatbot vs. AI agent in support: the difference that costs money

A customer writes on Sunday evening: “Hi, my order 4711 was supposed to arrive Friday, but it's not here. What's going on?” Two replies are possible. The first: “Thank you for your message. You can check the status of your shipment at any time via the tracking link in your shipping confirmation. Is there anything else I can help with?” The second: “Hi Anna, your parcel has been at the Cologne delivery depot since Friday noon and hasn't gone out for delivery yet. According to DHL it arrives tomorrow. If it's not there by Tuesday evening, just reply to this email and we'll open an investigation.”

Both replies were written by an AI. The first comes from a chatbot, the second from an agent. The difference isn't the writing style. The difference is that the second reply looked something up.

This article explains what separates an AI chatbot from an AI agent, how to spot the difference in a product demo, and why it shows up directly in your follow-up questions, your working hours and your reviews.

What a chatbot does

A chatbot is a conversation window. It takes an input and produces an output. With older chatbots the output comes from a decision tree: “Press 1 for delivery, 2 for returns.” With newer ones it comes from a language model that has read your FAQ and forms fluent sentences from it.

What both have in common: they don't know your orders. The chatbot knows you offer parcel tracking because it says so in the FAQ. It doesn't know where shipment 4711 is right now, because it doesn't look in Shopify or at the carrier. So it replies with what it has: a general pointer.

That's also why chatbots so often end up in a loop. The customer asks a specific question, gets a general answer, asks again, gets the same answer. According to a Gartner survey from 2026 of 3,566 customers, only 27 percent would try a chatbot again after a negative experience. A chatbot usually gets one chance.

What an agent does

An AI agent isn't a conversation window; it's a case handler. It receives the request, and before it replies, it does what a good support person would do:

1. It classifies the request. Delivery question, return, address change, complaint. This categorisation decides which data it needs and which rules apply.
2. It fetches the data. The order via email address or order number from Shopify, the customer history, the live shipment status from the carrier.
3. It checks your rules. Return period, goodwill limits, what to do with “delivered but not here”, which tone you write in.
4. It writes a finished reply. With name, order, status, date and the next step.
5. It decides whether the reply may go out. For clear cases in approved categories, automatically. For everything else, as a draft for you.

The last point is the most important and the most often overlooked. An agent isn't better because it sends more automatically. It's better because it knows when it shouldn't.

Here's how it works at Humane: the agent reads every incoming email, categorises it, pulls order and customer data from Shopify and the shipment status from the carrier. Then it writes a reply in your brand's tone. Whether it goes out automatically or waits as a draft is set per category, and the first week is always a training week. If the carrier has no scan yet, the agent says so: registered, no status yet. It doesn't claim the parcel is on its way. What that looks like for delivery questions specifically is in Answering tracking questions automatically.

The difference on four typical requests

It's clearest when you let both loose on the same cases.

“Where is my order?” Chatbot: link to parcel tracking. Agent: status, location, expected date, and what the customer should do if the date passes. The chatbot creates a follow-up, the agent closes the ticket.

“Can I still change the address?” Chatbot: “Address changes are possible as long as the order hasn't shipped. Please contact our support.” That's a circle: the customer has just contacted support. Agent: it sees the order hasn't shipped, changes the address or presents the change to you as a draft, and confirms the new address to the customer.

“I'd like to return this.” Chatbot: link to the returns policy. Agent: it checks whether the return window is still open, sends the customer to the returns portal or explains the process, and if the window has closed, applies your goodwill rule or presents the case to you.

“This is the third time I'm writing.” Chatbot: the same reply as the first and second time. Agent: it recognises the repetition and the frustration, doesn't reply with a template, and escalates to a human. This is exactly where the reviews that hurt you come from, and exactly where no automation should kick in.

Why the difference costs money

The maths is simple once you count the hidden costs.

Suppose you get 300 requests a month, and a third of them are delivery questions. A chatbot answers these 100 requests with a tracking link. Suppose half the customers are satisfied because they were only looking for the link. The other half write again, this time to you, now with a day's delay and a little less patience. So you've saved 50 tickets and made 50 tickets harder.

An agent answers the same 100 requests with the actual status. The cases where the status shows a problem, it presents to you as drafts. The rest is done. What remains are the cases that actually need your time.

On top of that come three costs that appear in no price comparison:

• Follow-up questions. Every general answer to a specific question creates a follow-up. It costs you more than the first one, because the customer has already been let down once.
• Reviews. A chatbot that doesn't help is a frequent trigger for bad reviews that have nothing to do with the product.
• The path to a human. According to Gartner, 87 percent of customers say companies using generative AI in customer service must provide a way to reach a human. A chatbot that blocks that path costs you customers, not just tickets.

The reverse also holds: for tasks where the answer comes from a system, customers don't want a human at all. According to Zendesk CX Trends 2025, 67 percent of consumers are ready to delegate tasks like order tracking to an AI assistant. The condition is that it actually gets them done.

How to spot the difference in a demo

Vendors now use the words “chatbot”, “assistant” and “agent” interchangeably. Don't rely on the label. Ask three questions.

• “Show me a reply to a real delivery question.” If the reply contains no shipment status and no date, but a link, it's a chatbot.
• “What happens when the order number doesn't exist?” A chatbot replies politely anyway. An agent says it can't find the order and asks, or presents the case to you.
• “Where do I define what may go out automatically?” If the answer is “all or nothing”, the most important control is missing. You want to decide per category.

A fourth question is worth asking about website widgets: how does the request reach you at all? A chatbot lives in the window on the shop page. But your customers write emails, reply to shipping confirmations, use the contact form. An agent has to work where the requests actually arrive: in the inbox.

Where even an agent has to stop

An agent isn't the solution for everything, and a vendor who claims it is, is selling you a chatbot with better marketing. Three limits remain:

• Missing data. If the carrier delivers no scan, the agent can only say there's no scan. That's the right answer, but it doesn't solve the problem.
• Discretion. Whether a customer may still return something after the window has closed is a decision. The agent can apply your rule. The exception stays with you.
• Anger. A customer who's furious doesn't want the best answer. They want someone to be responsible. The agent can recognise that and pass it on, not replace it.

Why we at Humane therefore didn't build a chatbot for the widget but an agent for the inbox is in Why Humane doesn't build ticket bots. And where the limit of automation lies overall, in The 90 percent limit.

In short

A chatbot answers questions with what it knows, and it only knows what's in your FAQ. An agent answers questions with what it looks up: order, customer, shipment status, your rules. The chatbot is cheaper to buy and more expensive to run, because it creates follow-ups and alienates customers. The agent handles the standard cases and presents you with the rest. When you evaluate a tool, let it answer a real delivery question and search for an order number that doesn't exist. Those two tests say more than any feature list. The overview of the whole topic is in AI in customer service: what actually works in 2026.

If you want to see the difference on your own emails: try it free for 14 days, no credit card, every reply as a draft first.