AI agents for e-commerce: a real case, costs, pitfalls
An AI agent helps an online shop mostly by answering repetitive questions (where is my order, how do I return an item, which size) and by preparing the work of the humans. In the example calculated below, the language model costs only a few dollars a month; the real budget is the tool, the setup and the review time.
The case: one shop, 600 conversations a month
Take a modest shop. This is not a testimonial: it is a calculation scenario, with assumptions stated plainly so you can replace the figures with your own.
- 600 support conversations a month, by chat and email;
- about 60% concern four topics: parcel tracking, delivery times, returns and exchanges, availability and sizes;
- the other 40% are disputes, lost parcels, requests for a goodwill gesture or angry customers;
- one person handles support, between other tasks.
The 60% of repetitive questions are the target. The remaining 40% should not be automated at first. This split is a working assumption: export three months of messages and sort them by hand before deciding anything.
What the agent does, step by step
You do not give it everything at once. Three levels, from least to most risky:
- Draft for the support person. The agent reads the message, checks the return policy and product pages, and proposes a reply. Nothing goes out without a human click. This is the test phase.
- Automatic replies on marked topics. Order tracking (read-only in the shop tool), delivery times, return rules. Off-topic or in doubt: hand over to the person with the context already summarised.
- Simple actions. Generate a return label, change an address before shipping. Refunds stay human.
This progression follows the one described in our article on AI agents for customer service, which shows why a correct answer does not prove you can let the agent act. For the difference between answering and acting, see also AI agent or chatbot.
What it costs: ready-made tools
Two families of offers exist. The prices below come from the official pages consulted on 5 November 2026; they change, so check before deciding.
| Option | Listed price | What it covers | Watch out for |
|---|---|---|---|
| Tidio, Starter plan | $24.17 a month | 100 billable conversations, 50 one-off Lyro conversations | No price per extra conversation is displayed |
| Tidio, Lyro AI Agent alone | from $32.50 a month | from 50 Lyro conversations a month | Quotas go up to 1,000; beyond that you must contact sales |
| Intercom Fin | $0.99 per billable outcome | An outcome does not always require customer confirmation | Monthly minimum commitment if you keep your current helpdesk |
| Zendesk AI, Team suite | €55 per agent per month (annual billing) | Includes AI agents | Usage and some features may cost extra |
At Tidio, the Lyro conversation quota is set in tiers (from 50 to 1,000 a month): you have to estimate your volume rather than pay as you go. With Fin, if the agent handled 360 conversations (60%) and each counted as one outcome, the bill would be 360 × 0.99 = $356.40 a month. That is an order of magnitude from an assumption, not a promise: the number of billed outcomes depends on the vendor’s definition. Our article on the displayed resolution explains why that definition matters as much as the price.
What it costs: building it yourself
With an automation tool such as n8n (€20 a month on the Starter Cloud plan billed annually, or free self-hosted), you connect your shop, your mailbox and a language model. The model is paid by usage, per million tokens.
Calculation assumption per conversation: 8,000 input tokens (the return policy and history are sent again at each exchange) and 1,500 output tokens. Prices per million tokens taken from our Models page:
| Model | Input / output (per million tokens) | Cost per conversation | For 600 conversations |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 / $0.50 | about $0.0016 | about $0.93 |
| GPT-5.6 Terra | $2 / $12 | about $0.034 | about $20.40 |
Even with the more expensive model, the bill stays under 25 dollars. The same logic appears in the article on the real cost of an AI agent: the model weighs little; integration, testing and human time are what cost. To cut the bill further, the guide on reducing the cost of an AI agent covers caching and model choice.
What the table leaves out: the time to write the knowledge base, connect the shop tool, test on real cases and review the first weeks. Count it in working hours, not token dollars.
A real example: what Klarna announced
Klarna is not a shop, but its case is the best documented. In a press release of 27 February 2024, the company said its assistant had handled 2.3 million conversations in one month, two thirds of its chats. It is a company statement, not an independent measurement, and it dates from 2024. Keep the method rather than the figure: volumes handled say little about the accuracy of answers. The full analysis is in our customer service review.
Pitfalls specific to e-commerce
- A vague return policy. If your returns page contradicts itself, the agent will too. Fix the text before connecting anything.
- Outdated stock and prices. An agent that says “available” based on yesterday’s cache does more damage than “I’ll check”. Connect it to a live read or have it say it does not know.
- Commercial promises. An agent that grants a discount code or a refund “to be nice” commits you. Limit its tools to the strict minimum, as the guide on agent permissions explains.
- Invented answers. Without a source document, a model fills the gaps. Measure that rate on your past conversations, as in our article on agent hallucinations.
- Customer data. Names, addresses and order numbers go through a third-party service. See AI agents and GDPR.
- The wrong metric. The number of “resolved” conversations is reassuring. Track reopenings, complaints and product returns instead.
Deciding in one week
- Export three months of messages and sort them into four or five topics.
- Write the rules (returns, delivery times, fees) in a single reviewed document.
- Test in draft mode on 50 past conversations, sending nothing.
- Count the errors, then open automatic replies on the most reliable topic.
- Recalculate the real cost after a month: subscription, tokens, review hours.
For other business uses, see the use cases for small businesses and the guide to evaluating an AI agent.
What to remember
- Aim first at repetitive questions whose answer is written down.
- The language model is the cheapest item; the tool and human time weigh more.
- Start with drafts, open topic by topic, keep refunds human.
- Redo the calculation with your own volumes and today’s prices.
Frequently asked questions
Can an AI agent run customer service for a small shop on its own?
It can answer alone on questions whose answer is written down: delivery times, tracking, return policy, sizes. Disputes, refunds and angry customers must be able to reach a person. Start in draft mode, then open automatic replies one category at a time.
How much does an AI agent cost for an online shop?
Three items: the tool (from 32.50 dollars a month for Tidio's Lyro agent, or 0.99 dollars per outcome with Intercom Fin), the language model if you build it yourself, and setup and review time. The model is often the smallest item.
Which language model should I pick to answer customers?
A small, fast model is enough for simple questions with documents supplied. Move to a stronger model only for cases that need reasoning, and test on your real conversations before deciding.
Can the agent refund or change an order?
Technically yes, if it is connected to your shop tool. In practice, start with read-only access (order status) and keep human approval for any money movement.
Does my customers' data go to the vendor?
Yes, as soon as the agent reads a message or an order, that data goes through the service. Check the hosting location, retention period and processing agreement before connecting anything.
Add Info Agent IA as a preferred source on Google Our articles show up more often in your Google “Top stories”.