Gorgias learns from your support history. We start from how the job should be done

Precedent or procedure: where the agent gets its judgement

It would be easy to write that Gorgias handles tickets while we handle the sale, and it would be out of date: their AI agent recommends products against live inventory and offers upsells as well as processing returns. The real difference is older than either product's feature list, and it comes from where each company started. Gorgias was a helpdesk first. Years of your tickets, macros, and help centre articles sit inside it, and that archive is a genuinely excellent thing to teach an agent from, because it is a record of what your team actually said to real customers. If you have it, use it. It is an advantage we cannot manufacture.

We start from the other end, because most of the businesses we work with do not have that archive. They are newer, or they have grown faster than their documentation, or they have just added a product line nobody has answered a question about yet. So instead of mining what was said, we write down how the job should be done: the questions to ask before recommending, the checks that have to pass before something is promised, when to hand to a person, what never to say. Then we build the agent to follow it. It is closer to training a new hire from a procedure than to learning from a transcript pile, and it is available on day one rather than after a year of accumulating tickets.

The distinction matters most exactly where support archives are weakest. "Does this fit a 2018 Golf estate with rails?" is not in the product description, not in the help centre, and not returned by any Shopify field, because it is a join between the customer's car and compatibility data that usually is not structured yet. History cannot teach an agent an answer nobody has ever worked out correctly, and a ticket archive will happily teach it the confident wrong ones your team gave under time pressure.

Two meters, and what they encourage

Gorgias bills on two axes: a ticket fee whenever something leaves the helpdesk, and an automation fee when the AI resolves a conversation on its own, with their billing documentation noting that since May 2025 a single AI-resolved ticket can draw both. That is a coherent model for a support operation, where every conversation you avoid is money saved, and it makes the automation case easy to put to a finance team. It is a strange fit for the conversations we are built for, because those are not costs to be deflected. A ten-turn exchange that ends in the right roof bar going into the basket is the most valuable conversation of the day, and any model that prices it as an expensive incident is pointed the wrong way. We bill per conversation regardless of how long it runs, which is the same idea from the other end: the conversation you want most should not be the one you pay most for.

You probably keep both

This is not a rip-and-replace pitch and the page would be less useful if it pretended to be. Most stores that run ConnectAI keep the helpdesk they already have: the agent handles the storefront conversation, resolves what it can against live catalog and order data, and escalates with the full transcript into the existing inbox when a person is needed. The helpdesk stays the system of record for human work, which is what it is good at. What changes is the volume and the quality of what reaches it: fewer wrong-product returns arriving as tickets, and the ones that do arrive carrying the whole conversation instead of a customer repeating themselves for the third time.

What this looks like in production

Two live deployments carry this argument, and they are the only two we cite. PTC Auto sells over 250,000 automotive parts and supports customers in seven languages across web, social, and their business phone; fitment questions decide most of those conversations, and the agent cut support operations cost by 40%. Dakdragerwinkel.nl matches roof racks, boxes, and bike carriers to a specific car. The customer gives a Dutch licence plate, the agent verifies it against the RDW national registry, checks the shop's own compatibility data, and recommends only carriers that genuinely fit. It resolves 73% of engaged conversations without a human, handles order tracking in 28% of them, and puts products straight into the basket. Both numbers come from production data, not a pilot.