Search for an ecommerce AI chatbot and you get ten lists of tools. They rank roughly the same eight vendors on features and starting price, and not one of them explains the thing that decides whether the purchase pays for itself: how the vendor decides a conversation was a success.
That definition lives on a billing page most owners never open. It is worth ten minutes.
An ecommerce chatbot does one job well. It answers questions whose answers already exist inside your systems. Where is my order. Can I still return this. Do you have it in a 42. Is it in stock in the Berlin warehouse.
None of those needs cleverness. All of them need a live, correct lookup. Salesforce spells out the mechanism in its own guide to order-status tickets: the bot asks for an order number, queries your order management system, and comes back with the real shipment status.
So the quality of the bot is mostly not a property of the bot. It is a property of the plumbing behind it. A store syncing stock and order state overnight will get an assistant that answers confidently and wrongly at three in the morning, to a customer who then emails you anyway.
What these bots are bad at is any judgement call your policy does not already contain. Goodwill on a damaged item. A shipping exception with a wedding date attached. A B2B buyer asking about terms. Those need to reach a person fast, and how fast is a configuration decision rather than a model capability.
Returns sit in between. Reading a return status is easy; checking eligibility, creating the return, and issuing the label is where the ticket actually closes, so where AI actually helps with returns depends on how much of that flow your systems expose.
Intercom's Fin is the clearest case to study, because it publishes the mechanics instead of saving them for a sales call. Fin is priced at $0.99 per outcome, on top of a base plan of $49 a month that includes 50 resolutions. Unused resolutions do not roll over.
Fair enough. The question is what counts as an outcome.
Fin records a resolution when the customer confirms the answer was satisfactory, or when the customer "exits the conversation without requesting further assistance." Go quiet for 24 hours after Fin's last answer and that is logged as an assumed resolution. Intercom's own pricing documentation is unusually direct about the uncomfortable case: a customer who leaves after Fin's answer without asking for more help "is counted as an Assumed Resolution and billed at $0.99."
Read that as an owner rather than as a support manager. A shopper who asks about a late order, gets an answer that does not help, gives up, closes the tab, and emails you instead has just been recorded as a success and charged for.
To be fair to Intercom, the same page lists what is not billed, and the list is reasonable. A customer who asks outright for a human. A customer whose frustration Fin detects, triggering its own escalation. A procedure that fails technically. A conversation abandoned after Fin asked a clarifying question rather than gave an answer. Intercom even names the countermeasure: configure escalation rules so the assistant hands off when it senses frustration, instead of letting the conversation die quietly into a paid resolution.
That setting, not the model, is what your evaluation call should be about.
The arithmetic matters more than the unit price, too. A thousand billed outcomes in a month is $990 plus the base plan, and the share of your chat volume that turns into billed outcomes is partly a configuration decision and partly not. Run a trial on real traffic and bill yourself before you model anything.
Two figures from Zendesk's 2026 roundup of AI customer service data sit awkwardly next to each other. 51% of consumers say they prefer interacting with bots over humans when they want immediate service. 68% believe chatbots should have the same level of expertise and quality as highly skilled human agents.
The first number is your business case. The second is the bar. People will take the bot over the queue, then hold it to the standard of your best agent. Nobody grants a discount for it being software.
Here is the part most owners miss: the assistant is rarely the expensive line on the invoice.
In the AI integration work we do at Encomage, a project like this splits the way data projects always split. Choosing and configuring the assistant takes days. Making order status, stock levels, shipment tracking, and returns state reachable in real time, and trustworthy enough to show a customer unprompted, takes weeks. Stores that already hold a clean live link between storefront and ERP get a good chatbot quickly. Stores moving that data in a nightly batch do not, and no amount of model quality closes the gap.
It is the same failure surface as any other back-office integration, and it fails in the same four places most Magento ERP integrations break. If your stock figures are already an hour stale for your own staff, putting them in front of a customer in a confident sentence makes the problem louder rather than smaller.
The last one gets skipped most. If your team answers four hundred order-status emails a month at two minutes each, you know exactly what the assistant has to beat. If you have never measured it, you will not be able to tell a good deployment from an expensive one when the invoice arrives next quarter.
Most of the difficulty here sits in the systems behind the chat window rather than in the chat window itself. That is the kind of work we take on at Encomage. It usually starts with an honest look at what your order, stock, and returns data can answer in real time today, before anyone picks a vendor.
Let’s build something powerful together - with AI and strategy.
