A Gorgias AI Agent demo usually goes the same way. Someone types "how long does shipping take?" and the agent answers in a second, correctly and politely. Everyone in the room nods.
Two weeks after launch, a customer writes in about a damaged gift that someone else bought with a discount code. The same agent now has nothing to go on, and what it does next decides whether the brand keeps it switched on.
Both moments are real. The first is what AI Agent does well out of the box. The second is where the work sits. This guide is for Shopify brands on Gorgias deciding whether to turn AI Agent on, and what to prepare before they do. Our team's track record includes 60+ DTC brands on retainer, and everything below comes from what our team has seen inside those inboxes.
Every helpdesk now sells the same promise
Gorgias has AI Agent. Intercom has Fin. Zendesk and Front sell their own AI layers. The decks look alike, with a large automation percentage on the first slide and a smaller payroll on the second.
The figure we're comfortable quoting is narrower. On properly configured accounts, the median is 50-60% of repeat tickets handled by the AI Agent. Three words in that sentence do a lot of work:
- Repeat covers the questions that come back every day. One-off problems sit outside it.
- Median means half the accounts land below that range.
- Properly configured means someone did the preparation described further down.
Quote the number without those three words and you're quoting a sales deck.
Five checks before you switch it on
We run these on an inbox before recommending AI Agent. A "no" on any of them is a thing to fix first, and fixing it is cheaper than a launch that gets rolled back.
- Is the store big enough? Below €2M GMV, a solid set of macros and a good Help Center usually pay back sooner. AI Agent needs volume to earn its setup.
- Do the same questions keep arriving? Pull last quarter's tags. If a handful of categories (order status, address changes, sizing, returns) make up most of the inbox, there's a pattern to automate. If volume is scattered across dozens of small topics, there isn't.
- Are your policies on paper? Return windows, refund rules, what counts as a defect, what happens when a parcel shows as delivered and never arrived. If today's answer is "ask whoever's on shift", the agent has nobody to ask.
- Can someone describe your voice? A tone guide helps. So do a few months of tickets resolved by your best agents, because we write the agent's tone instructions from them.
- Who owns the knowledge after launch? A carrier moves its cut-off. The sale season brings a temporary return rule. Someone has to update the source the agent reads, and if nobody's name is on that job, answers go stale without anyone noticing.
One more, harder to score: if your team is split on whether to use AI at all, have that conversation before launch. An agent that half the team wants to fail usually does.
Three places it goes wrong
Rules that live in someone's head
AI Agent follows instructions. It can't follow a policy that shifts depending on who's answering. Brands that skip writing the rules end up with an agent that refunds one customer and refuses the next for the same problem, and the team stops trusting it fast.
A usable rule reads like this: replace a damaged item if the customer sends a photo and the order arrived within the last month; otherwise, hand it to a person. A page of rules written that plainly does more for automation than any setting inside Gorgias.
A voice built on quirks
Some brands write to customers in deliberate lowercase, with running jokes and their own slang. AI Agent can get close. Getting all the way takes far more configuration and review than a neutral voice, and the cost can outrun the savings for a long time.
Sometimes the honest call is to keep people on the conversations where the voice is the product, and automate only the plain logistics around them.
Products where the conversation is part of the sale
Think fine jewellery, or made-to-measure furniture: anything where the buyer expects someone to know their order. In those categories, a bot closing a complaint reads as indifference. AI can still do the prep work, tagging the ticket and pulling the order details so the person who answers starts with context.
What "configured" actually involves
Turning AI Agent on is quick. Add a few FAQ articles, ask it about shipping, get a correct answer. That's the demo from the opening, and it's where a lot of launches stop.
The setups that last have five pieces in place:
| Piece | What it is | What happens without it |
|---|---|---|
| Knowledge source | One maintained document covering returns, refunds, shipping cut-offs, sizing by category, defects and exceptions. Dated, with an owner. | The agent fills gaps with generic answers. |
| Tone instructions | Written from your best resolved tickets: greeting, sign-off, how formal to be, whether exclamation marks are allowed. | Replies that are correct and sound like a different brand. |
| Action limits | What it may do alone, such as refunds below a ceiling you choose or address edits before fulfilment. Anything outside goes to a person. | One wrong decision repeated at scale. |
| Ticket classification | Checking that incoming tickets land in the right category before the agent replies. | A good answer to the wrong question. |
| Weekly review | A person reads the AI-handled tickets flagged for tone or accuracy. | Drift nobody spots until a customer posts about it. |
If you're still choosing what to automate with rules versus AI, our guide to Shopify and Gorgias automations covers the rule-based side.
Can it make up a policy?
It's the first question finance asks, and a fair one. Gorgias AI Agent answers from the sources you connect, such as your Help Center articles and the guidance you write, so an answer can be traced back to a document. That lowers the risk, and a vague or outdated source can still produce a confident wrong answer.
How we keep a mistake small:
- When the agent isn't sure, it hands the ticket to a person instead of guessing.
- Action limits cap what a single error can cost.
- The weekly review catches a bad pattern before it spreads.
- Someone on your side knows how to pause AI Agent, and has done it once in a test.
What it changes for your team
Vendors often pitch AI Agent as a way to halve the support team. What tends to happen is quieter. AI Agent takes the predictable tickets: where's my order, change my address, which size fits, a simple return. People keep the harder ones, like escalations, defects, messy returns and wholesale questions, and each of those takes more effort than the tickets that left.
So headcount usually holds while the job changes. Triage shrinks. Investigation and policy work grow. The saving shows up later, when order volume rises and you don't have to hire for it.
Cutting the team on launch day removes the people AI Agent hands its hard cases to, which is the quickest way to make a working setup look broken. Let AI handle the routine. People handle the rest.
A rollout order that keeps mistakes small
- Write the source. The policy document from the checks above, signed off by ops and CX together.
- Set the voice and the limits. Tone instructions and action ceilings, plus the rules for passing a ticket to a person.
- Test where customers can't see it. Run the agent against real past tickets and read what it would have sent. Where an answer is wrong, fix the source rather than the answer.
- Go live on the safest categories. Order status and address changes first. Widen the scope only while the weekly review stays clean.
- Keep reviewing. Policies change every season, and the document and limits change with them.
Our AI Agent service follows this order.
The 30-minute version
On a call, we look at your top ticket categories, how much of your policy is written down and who would own the knowledge after launch. You leave with one of three answers: start now, fix a short list first, or don't automate this yet. If it's the second, you get the list.
CTRL runs on Gorgias. Our team was named Best European Gorgias Partner 2024.
Book a demo. 30 minutes, no deck.
