What is an AI shopping agent, and how is it different from a chatbot?

Most of us have used a retail chatbot. You open a small window in the corner of a website, ask where your parcel is, and receive a link to the returns policy. It is useful up to a point, but it rarely finishes anything for you.

An AI shopping agent is a different kind of tool. It is built to take on a task, such as "find me a pair of waterproof walking boots under £120 that will arrive before Saturday", and work through it until there is something concrete for you to approve. This article explains what that means in practice, where the line sits between a chatbot and an agent, and what safeguards you should expect before letting any software spend money on your behalf.

The short answer

A chatbot responds. An agent acts.

A chatbot is designed around a conversation: you ask, it answers, and the work of deciding and doing stays with you. A shopping agent is designed around an outcome. It can search, compare, ask you a clarifying question, prepare a purchase and, once you approve, complete the checkout and follow the order through to delivery.

The difference is not that one uses artificial intelligence and the other does not. Many modern chatbots use the same large language models as agents. The difference is in what the system is allowed to do, and how it is held accountable for doing it.

What a chatbot usually does

Most retail chatbots are tied to a single shop and a narrow set of jobs:

• answering questions from a help centre
• looking up the status of an order
• suggesting products from that retailer's own range
• handing you over to a human when things get complicated

They are helpful for quick answers, but they are not working for you. They are working for the shop, inside the shop's website, and the moment you need to compare options elsewhere you are on your own again.

What a shopping agent does

A shopping agent sits on your side of the counter. With Ottr, for example, you message it like a contact, describe what you need in your own words, and it takes the task from there.

It plans and uses tools

An agent breaks a request into steps and uses tools to carry them out. For a request like the walking boots above, that might mean searching for products, filtering by price and delivery date, checking stock, and comparing the shortlist on the things you said mattered.

It remembers context

A good agent keeps track of the conversation, so you can say "show me the second one again" or "what about in brown?" without starting over. In Ottr's case, the conversation belongs to you rather than to an app, so you can begin on the web and carry on in Telegram without repeating yourself.

It can complete the task

This is the real dividing line. When you have chosen something, an agent can prepare the purchase, check the final price, and, after you approve it, pay and place the order. It then keeps you posted on confirmation, dispatch and delivery, and can help with a return.

Agent, assistant, chatbot: a quick comparison

	Retail chatbot	Shopping agent
Works for	The shop	You
Scope	One retailer	Across retailers it can reach
Main job	Answer questions	Complete a task
Can it buy?	Rarely	Yes, with your approval
Memory	Usually per session	Across the conversation and channels

Why safeguards matter more for agents

Once software can spend money, the question changes from "is the answer correct?" to "should this action happen at all?". Language models are capable, but they can misunderstand a request, and a casual reply in a chat is easy to misread. An agent that can buy things needs firm boundaries that do not depend on the model getting everything right.

At OTTR we have built those boundaries into the platform itself rather than into the model's instructions:

1. The agent cannot approve or pay. Its tools can search, compare and prepare a purchase. Approval and payment sit in separate parts of the system the agent has no access to.
2. You approve every purchase on a secure OTTR page. Saying "yes" in a chat only ever produces an approval link. The approval itself happens on a page that requires your signed-in session, and is tied to the exact item, price, retailer and delivery address. We explain why in Why a 'yes' in a chat should never be enough to buy something.
3. Each purchase gets its own card. Ottr pays with a single-use virtual card limited to the amount and retailer you approved, so your own card number is never shared. See Single-use virtual cards: how Ottr pays without exposing your card.

These are not features layered on top. They are the reason an agent can be trusted with a task at all.

What a shopping agent is not

It is worth being clear about the limits, too.

• It is not autonomous by default. A well-designed agent does the legwork and leaves the decision with you. Standing permissions, such as "reorder coffee when I run low", are a future possibility at OTTR, but they are modelled and switched off for now.
• It is not all-knowing. An agent can only compare what it can reach. If a retailer is not available to it, it should tell you rather than guess.
• It is not a substitute for judgement. You still decide what you want. The agent's job is to make that decision easier and the follow-through effortless.

Where Ottr is today

OTTR is in development. You can talk to Ottr on the web, and on Telegram where it is enabled. WhatsApp is waiting on Meta's approval, and Apple Messages is planned. Purchasing, payments and card issuing currently run in a sandbox against a test merchant, so no real money moves while we complete the regulatory and partner work needed to go live. Stripe is our acquirer of choice for card payments.

We think that is the right order of things: prove the safeguards first, then open the doors.

Other agents can use it too

One more difference worth knowing about. Because a shopping agent is built around tools rather than a chat window, other AI systems can use it as well. Your own assistant could ask Ottr to find and prepare a purchase, while the final approval still comes to you. We cover this in Agent-to-agent commerce, and you can see it working on our page for AI agents.

The takeaway

A chatbot is a helpful way to get answers. An AI shopping agent is a way to get things done: it searches, compares, prepares and, with your explicit approval, buys. The value is in the follow-through, and the trust comes from the guard rails around it.

If you would like to see how that feels, you can try Ottr in sandbox today, and for practical tips on getting good results, read How to brief an AI shopping assistant.