Mike HeapMike is an experienced Product Manager who focuses on all the “non-development” areas of My AskAI, from finance and customer success to product design, copywriting, testing and more.
Table of Contents
- What does the brand do?
- Which helpdesk does the brand use?
- How did the brand train their AI customer service agent?
- When did the brand decide to turn on 'direct replies' to customers?
- What was the biggest thing the brand did to improve their AI agent's resolution?
- How does the brand customize their AI agent setup to work for their business?
- Communication and brand voice
- Context and clarification
- Handover and escalation
- What impact is the brand's AI customer service agent having now?
- Where does the brand go from here?
Created time
May 27, 2026 02:57 PM
Title length (<60)
Author
Last optimised
Ecomm?
Tags
Image
Publish date
Jul 15, 2026
Video
Slug
my-askai-premium-accessories-case-study-2026
Featured
Type
Case study
Ready to Publish
Ready to Publish
A premium accessories brand resolves ~1,856 of ~2,900 monthly Gorgias tickets with AI, holding 85% AI CSAT and saving ~153 hours every month across email Copilot and chat direct replies.
This premium accessories brand sells direct on Shopify, and the support load looks like every modern accessories brand at scale: order status, returns, exchanges, charger compatibility, "does this fit my device". The CX team carrying that load is small relative to the volume.
Across the chat widget on the storefront and the email tickets routed into Gorgias, ~2,900 conversations a month now land with our AI. The split is deliberate. Chat answers go direct to customers; email tickets pass through Copilot so a human still signs off on the reply.
64% of the routed tickets resolve without a human, customer satisfaction sits at 85%, and the team saves roughly 153 hours a month.
Here's how it came together.
What does the brand do?
The brand makes premium tech accessories for one of the major device ecosystems, aimed at the kind of buyer who cares how their charging cable looks on a kitchen counter.
The catalog spans charging (braided cables, magnetic charging stands, multi-device docks, travel kits), phone and earbud cases, and a smaller everyday-carry line. The brand is design-led: muted palettes, natural materials, considered packaging, a price tier above the commodity-accessories shelf.
Distribution runs through their direct storefront (Shopify-powered) plus major electronics retailers. Brand collaborations are a recurring marketing motif rather than the core revenue line.
A broad catalog lives in Shopify. The size of it matters less as a marketing fact than as a support fact: every product in it is a potential "is this compatible with my new phone", "what cable length do I have", "which dock fits my laptop" question. The catalog is the first thing the AI has to know.
Which helpdesk does the brand use?
The brand runs support inside Gorgias, with chat embedded on the storefront and email tickets flowing into the same inbox. The Shopify backbone behind the storefront keeps order and customer data accessible from the same place the support team works.
The deployment mode is split by channel rather than collapsed into a single setting. Email tickets route through our AI in Copilot mode: we draft the reply as an internal note, surface the supporting context, and a human reads it before the customer sees it. Chat traffic is the opposite: our AI replies directly to the customer in the widget.
Chat and email aren't the same channel, and treating them like one is where most rollouts go wrong. Chat customers are mid-task, often standing in front of a charger that won't sync or a dock that won't power, and they expect a reply now. The cost of waiting is higher than the cost of an imperfect AI answer the human can correct on the follow-up.
Email is the opposite. The customer has already accepted that a reply takes longer, the ticket usually carries more context, and the reply itself is going out under the brand's own voice. Copilot keeps a human in the loop on every email reply without losing the speed-up of our AI drafting it first.
The Gorgias integration is the canvas the rest of the configuration sits on. See our Gorgias integration page for more.
How did the brand train their AI customer service agent?
Luckily, most of the brand's knowledge already existed in writing somewhere; it just lived in several different places. The training plan connected each of them in turn:
- Gorgias help center went in first via our Knowledge Base connector. Years of public-facing answers (compatibility tables, return windows, shipping policy, care instructions) were already written down for customers. Routing them into the AI's knowledge base was a one-step task rather than a content project.
- Shopify wired in next. With the full catalog and live order data behind every WISMO and warranty question, Shopify is the most important knowledge source on this build. The AI doesn't just know about the catalog from a help-center summary; it queries the actual product data when a customer asks whether a specific cable fits a specific phone, and the actual order data when a customer asks where their shipment is.
- User Data API added the customer-specific layer on top of that. Order history, status, what's been bought, what's in flight: every "where is my order" question turns into a real lookup against the customer's own record (rather than a templated apology).
- File uploads filled in the gaps the public help center doesn't carry. Internal notes on edge-case returns, supplier lead times for less-common SKUs, policy detail that lives on internal docs rather than the customer-facing site, all routed in as supplementary context the AI can reach when a question touches them.
- Self-Learning was switched on across the lot. Every time a human agent closes a ticket the AI couldn't, our AI drafts a new knowledge article from that exchange. Over the period reported, around 400 of those auto-drafted articles have shown up in AI replies (a working measure of how much the corpus has grown on its own, without a scheduled content-writing sprint to keep up with new product drops).
The set of sources reads as a layered stack rather than a single ingest job: structured product and order data underneath (Shopify + User Data API), curated public content in the middle (help center + uploads), and a self-extending layer on top (Self-Learning) that fills the gaps the first two miss.
When did the brand decide to turn on 'direct replies' to customers?
The answer is two answers, because the brand runs two channels and made a separate call on each.
Chat went direct early. Chat customers expect immediacy by definition: they opened a chat widget rather than a contact form. The questions that come in via chat skew toward lower-stakes ground: order status, return policy, "does this work with my phone".
The cost of a slightly imperfect AI reply that a human follows up on is lower than the cost of leaving the customer staring at a typing indicator for twenty minutes.
Email stayed in Copilot mode and has stayed there. Email tickets carry more context and more brand exposure: longer customer messages, more detail on the order, replies that read in the customer's inbox alongside marketing emails and order confirmations.
A misjudged reply on email costs more than the equivalent slip on chat, and the saved minutes per ticket are still substantial when the AI drafts and a human approves (rather than the human composing from scratch).
The split is the deployment decision the post turns on. It's not bravery vs. caution; it's matching the AI's reply mode to the channel's tolerance for a wrong reply.
What was the biggest thing the brand did to improve their AI agent's resolution?
The strongest lever in the brand's setup is the live-data layer (Shopify and the User Data API working together) and what it does to the largest slice of incoming tickets.
A premium accessories brand running ~2,900 tickets a month against a catalog this wide ends up with a long tail of compatibility questions, an even longer tail of order-status questions, and a thick middle of returns, exchanges and "I bought the wrong one". The class of question that used to default to a templated "please contact support" (or worse, a templated "check your shipping confirmation email") turned into a class of question the AI can close.
The AI sees the order date, the SKU, the cable length, the dock variant, the customer's purchase history. The reply moves from generic to specific because the data underneath it is specific.
Layered on top of that is the Self-Learning loop. Every ticket where the AI didn't know the answer becomes a piece of training data the moment a human closes it. Our AI compares its own attempt against the human reply, identifies the gap, and drafts a knowledge article that feeds back into the next answer.
~400 of those auto-drafted articles are now in active use across AI replies (a working count of how much the corpus has self-extended since launch). For a brand running through frequent product drops, seasonal SKUs and ongoing accessory-line refreshes, that's the difference between a knowledge base that ages and one that keeps up.
The two layers are complementary rather than alternative. The live-data layer handles the bulk of the well-shaped tickets where the customer is asking about a specific record the AI can look up. Self-Learning handles the long tail of newer questions where the answer isn't yet in the help center but a human agent has already composed it.
Resolution rate is the headline; the architecture underneath it is two specific things working in concert.
How does the brand customize their AI agent setup to work for their business?
The customizations sit inside our Guidance layer. All three Guidance types (Communication, Context & Clarification, and Handover & Escalation) are configured.
Communication and brand voice
The brand's voice on email and chat is the same one that runs across the product photography and the packaging: considered and design-led, never loud.
Communication Guidance is where that gets enforced inside AI replies: sign-off conventions, tone calibration, the structure of a typical answer, when to use a bulleted list and when to keep things conversational. The AI doesn't guess at house style; it has a configured definition of it.
Context and clarification
The Context & Clarification layer pulls its weight on a catalog this wide. A customer asking "does this charger work with my phone" is asking a question the AI can't answer until it knows which charger and which phone.
Rather than commit to a confident-but-wrong reply, the guidance routes the AI to ask the clarifying question first: which model, which year, which cable length. The cost of one extra back-and-forth is much lower than the cost of a misfit recommendation that becomes a return.
The same logic applies to ambiguous "where is my order" questions when a customer has multiple recent orders, or "I want to return this" without a SKU specified. Clarification before commitment, by design.
Handover and escalation
The Handover & Escalation layer is where the high-stakes lanes get fenced off. Wholesale and B2B inquiries, complaint escalations, suspected fraud, refund cases that fall outside the standard window: our guidance routes each of these to a human before the AI gets to reply.
The escalation reads intent rather than keywords. That matters on a long-form email where the customer hasn't typed the word "refund" but is clearly working their way toward one.
The three Guidance layers cover voice, judgment, and routing in sequence: what to sound like, when to clarify, and when to step back. None of them are off-the-shelf; each is configured to the brand's own policy specifically.
What impact is the brand's AI customer service agent having now?
- 64% AI resolution rate across the routed Gorgias tickets
- ~2,900 tickets handled per month by the AI (~1,856 resolved without a human, ~1,044 escalated)
- ~153 hours saved per month (at ~5 minutes per AI-resolved ticket)
- 85% AI CSAT across resolved tickets
- ~400 auto-drafted Self-Learning articles in use across AI replies, a compounding knowledge asset
The headline isn't that our AI handles every ticket; it doesn't, by design. Email runs through Copilot mode so a human is in the loop on every reply that goes out under the brand's name.
The headline is what 64% looks like on a catalog this wide, running through a split-channel deployment that protects email tone while taking the chat load directly: a small CX team getting back the equivalent of a full work-week every month, with our AI holding customer satisfaction above 85%.
Where does the brand go from here?
A few threads are next.
The first is more autonomous Tasks layered on top of the same Shopify and User Data API plumbing already in place. Refunds, exchanges, address changes and replacement orders are natural extensions of the live-data layer; the AI is already reading the customer's order. The next step is letting it act on the order where the policy allows.
A warranty-claim Task that checks the Shopify order date against the warranty window (and either processes the claim or escalates cleanly) is the kind of workflow the data layer is already half-built for.
The second is the Self-Learning corpus, which keeps growing in the background. As new accessory lines launch and seasonal SKUs cycle through, the AI's grasp of the current catalog only gets better; the 400 articles already in use are a midpoint rather than a ceiling.
I've been on enough of these rollouts to know the frame that lands hardest: 100% automation isn't the goal. The goal is the right tickets handled well, on the right channel, at the right level of human oversight.
64% resolution at ~2,900 tickets a month, with email running on Copilot and chat running direct, is a working answer to that, and the trajectory is up. (For an eCommerce step-change story on a different helpdesk, see how Edel Optics moved from 25% to 79% AI resolution on Zendesk once the User Data API was wired in.)
If you'd like to see more customer stories like this one, browse all our case studies here. Or, if you want the same setup on your own Gorgias inbox, check our Gorgias integration page and pricing.
