Shopify Runs Your Store. An AI Agent Handles the Conversations.
Shopify is a genuinely good platform for running an online store. It handles the catalogue, the checkout, the payments, the fulfilment — the mechanics of commerce.
What Shopify doesn't handle is the conversation layer. The customer who messages at 10 PM asking where their order is. The shopper trying to figure out if the medium will fit. The buyer who needs to start a return. The lead who abandoned their cart after adding three items.
These conversations happen in your inbox, your chat widget, your WhatsApp, and your email — and in the Shopify stores we've worked with, they stack up faster than most small teams can keep up with. A store doing 200 orders a week can easily generate 60–80 support messages per week, the majority of which are variations of the same five questions. That is a significant chunk of a small team's week spent on work that follows predictable, answerable patterns.
An AI agent connected to your Shopify store handles them automatically, drawing from your live product and order data to give accurate, specific answers at any hour. The agent doesn't know less than your support team about order status — it knows more, because it reads the API directly rather than logging into the admin panel to copy-paste tracking numbers.
What a Shopify AI Agent Can Do
Order Status and Tracking
Shopify stores the order. The agent reads it.
When a customer asks about their order, the agent looks up their purchase by email or order number, retrieves the current status and tracking, and gives them a specific update: "Your order shipped on Tuesday with Royal Mail. The last tracking update was this morning at the East Midlands sorting facility. Estimated delivery is tomorrow."
No copy-paste. No manual lookup. No "we open Monday at 9."
For most Shopify stores we've audited, order status sits at 30–40% of all support volume. Automating it is the single highest-ROI starting point. A homewares brand we worked with was spending roughly 12 hours per week on order-status queries alone across a team of three. Within the first week of the agent going live, that dropped to under two hours — the remaining two hours being edge cases the agent correctly escalated rather than guessed at.
The agent also handles the related questions that spiral off a tracking query: "It says delivered but I haven't received it." "The tracking hasn't updated in three days." "The delivery address is wrong — can you redirect it?" These require decision logic, not just data retrieval, and a well-built agent handles each one differently depending on what the courier and Shopify data actually shows.
Returns and Exchanges
A customer wants to return an item. The agent checks their order against your return policy settings, determines eligibility, generates a return label through your returns app (Loop, AfterShip, WeSupply, or similar), and sends the instructions — all inside the conversation.
Returns that fall outside policy — past the window, damaged by the customer, missing tags — get handled honestly and routed to your team with full context if an exception review is needed. The agent doesn't try to bluff its way through awkward cases; it hands them off.
What this looks like in practice: a customer contacts at 11:30 PM wanting to return a jacket bought 28 days ago. Your policy allows 30 days. The agent confirms eligibility, asks whether they'd prefer a refund or exchange, generates the return label, and sends the prepaid postage instructions — the whole thing resolved before the customer goes to sleep, with zero involvement from your team. The same customer contacting at day 32 gets a clear explanation of the policy, an escalation to the team flagged as "outside window, needs exception review," and a realistic expectation of when they'll hear back. No false promises, no unhelpful silence.
Product Questions and Recommendations
"Does this come in a size 12?" "What fabric is this made from?" "Is this suitable for outdoor use?" "What's the difference between these two models?"
The agent answers from your Shopify product metadata — variants, descriptions, materials, dimensions, care instructions. The accuracy is bounded by the quality of your product data, which turns out to be a useful forcing function to actually keep it complete.
For recommendation queries — "I'm looking for a gift for someone who runs, budget about £60" — the agent searches your catalogue and presents the most relevant options with a short reason each fits. This is where the agent earns its keep beyond support: it behaves like a knowledgeable sales assistant who has read every product page and can match customer intent to inventory without needing to be prompted by a human colleague.
A clothing retailer we worked with found that 18% of product questions were being answered incorrectly by their human team — not from negligence, but because product descriptions on the site contradicted variant details in the backend. Building the agent exposed the inconsistencies. Fixing them improved both the agent's accuracy and the human team's answers.
Abandoned Cart Recovery
A customer added items and left. Shopify captures this. The agent can follow up.
Inside a defined window (typically 1–4 hours), the agent sends a message via email or WhatsApp referencing the specific items they left behind. Not "you left something behind" — a message that acknowledges what they were looking at and asks if they had a question.
Some people left because they had a question; answering it converts them. Some left on price — the agent can offer a targeted discount at this point if your policy allows. Some left to think — a nudge at the right moment brings them back. Well-timed, personalised recovery typically runs 10–20%. Generic email blasts sit at 3–5%.
The distinction that matters here is conversational versus broadcast. A broadcast says "you left something in your cart." A conversational recovery opens a two-way exchange: "I noticed you were looking at the trail running shoes — do you have any questions about sizing or delivery before you decide?" If the customer replies, the agent is ready to continue. If they don't, the recovery attempt closes cleanly without repeated messages.
Post-Purchase Follow-Up
After delivery, the agent checks in. A short message: "Your order arrived earlier this week — how are you finding it?" This opens a feedback loop, catches issues before they end up on a public review, and creates a natural moment to suggest complementary products.
Customers who respond positively get a gentle nudge toward a review. Customers who flag a problem get immediate attention from your team.
The timing matters. Contact too soon after delivery and the customer hasn't had time to form an opinion. Contact after two weeks and the moment has passed. Shopify delivery data gives the agent the actual delivery timestamp — it follows up based on when the item arrived, not a fixed number of days after dispatch.
Customer Account and Profile Queries
"Can I change my delivery address?" "I need to update my payment method." "I've moved house — can you update my details?"
Standard account queries your team currently handles manually can be managed automatically through Shopify's Customer API — reading and updating customer records, confirming changes, and triggering any downstream actions. For stores with loyalty programmes or subscription products, the agent can also field queries about point balances, renewal dates, and subscription management without those going to a human inbox.
The Technical Integration
The Shopify integration uses the Shopify Admin API and Storefront API:
Shopify Admin API — reads order data (status, items, shipping, tracking), customer data (email, address, order history), and product data (variants, inventory, descriptions). Writes customer updates (address changes, tags). Triggers refunds and cancellations within your defined policy rules.
Returns app API — your returns management app (Loop Returns, AfterShip Returns, WeSupply) has its own API. The agent connects to this to check eligibility and generate return labels.
Courier tracking — tracking is retrieved from the courier API (Royal Mail, DPD, Hermes, FedEx, etc.) using the tracking number from the Shopify order.
Communication channels — the agent lives in your website chat widget, WhatsApp Business, email, or all three at once.
Authentication for the Admin API requires a private app credential scoped to the data the agent actually needs — not full admin access. We scope down to read-only on orders and products by default, with write permissions added only for the specific operations the agent is built to perform. This reduces the blast radius if credentials are ever compromised.
Shopify Apps vs Custom Build
Several Shopify apps offer AI-powered customer service — Tidio AI, Gorgias AI, Reamaze. They're worth evaluating for straightforward use cases.
| Factor | App (Tidio, Gorgias AI) | Custom Build |
|---|---|---|
| Setup time | 1–3 days | 4–5 weeks |
| Return on investment | Faster initially | Higher ceiling over time |
| Return policy logic | Generic rules | Matches your exact policy |
| WhatsApp integration | Limited or paid tier | Native, full two-way |
| Actions (process returns, update accounts) | Read-only or partial | Full write access |
| Product data accuracy | Trained on scraped content | Direct API access to live data |
| Escalation routing | Into app inbox | Into your existing tool (Gorgias, Freshdesk, inbox) |
| Monthly cost | £50–£300 depending on volume | Build cost amortised; no per-conversation fee |
A custom build makes sense when:
- You want tighter integration with your specific product data and return logic
- Your workflows are specific enough that off-the-shelf tools can't accommodate them
- You want the agent on WhatsApp or a channel the apps don't support well
- You need the agent to actually take actions (process returns, update accounts), not just answer
- Off-the-shelf tools have produced underwhelming results because of generic training
The honest answer for most small Shopify stores: start with an app. The honest answer for stores doing £500K+ a year with specific workflows: a custom agent will outperform an app by a meaningful margin. We've replaced both extremes — sometimes telling a smaller store they don't need us yet.
What to Expect in Practice
The first two weeks after go-live are not passive. We monitor every conversation the agent handles, review escalations, and tune the decision logic based on what real customers actually ask — not what the client predicted they'd ask.
Stores consistently underestimate two things: the volume of edge-case queries that arrive in the first week, and how much customers diverge from the assumed conversation pattern. A customer asking "where is my stuff" is a reasonable test case. A customer writing "hi my mum ordered something and I need to change the address she can't figure it out" is not — but it arrives on day two, and the agent needs to handle it cleanly.
The calibration period is where the agent gets genuinely useful rather than just functional. After 30 days of live operation, a well-built agent handles 70–80% of contact volume without human involvement. The remaining 20–30% are the exceptions — complaints needing judgment, queries outside scope, edge cases where the right answer is "a human should handle this."
Common Mistakes and What Goes Wrong
Building before cleaning product data. The single most common problem. An agent that confidently tells a customer the wrong fabric content or cites a variant that doesn't exist is worse than no agent — it erodes trust. Product data review is not optional pre-work; it is part of the build.
No defined escalation path. Some teams build the agent and assume it will handle everything. It won't and shouldn't. Without a clear escalation route, queries that need a human go nowhere — the agent ends the conversation or loops, and the customer goes to a review site. Define what escalates, where it goes, and who picks it up.
Deploying on every channel at once. Starting with your website chat widget and one other channel is the right approach. Deploying across website, WhatsApp, email, and Instagram DMs simultaneously on day one creates too many variables to debug if something goes wrong.
Treating the agent as a cost centre, not a revenue tool. Stores that build purely for support deflection leave money on the table. The same agent handling order queries can run abandoned cart recovery, post-purchase follow-up, and recommendation conversations. The infrastructure is the same; the scope just needs to include it.
Expecting it to work without feedback loops. An agent deployed and forgotten degrades over time as your product catalogue changes, your return policy updates, and customer query patterns shift. Monthly review of escalations and agent accuracy keeps it calibrated.
What You Need Before You Start
Complete product data. The agent's product answers are only as good as your Shopify product metadata. Before building, review your descriptions, variant details, materials, dimensions, and care instructions. Incomplete data produces vague, unhelpful answers — and customers can tell.
A documented return policy. The agent applies your return policy programmatically. It needs to be unambiguous: what's the window, what condition must items be in, what's the process for exceptions. If your policy has grey areas, resolve them before building, not after.
A clear escalation path. Define what goes to a human: queries outside the agent's scope, complaints needing compensation, returns outside policy that need exception review. The escalation should plug into your existing support workflow (Gorgias, Freshdesk, or just your inbox).
Your communication channels. Decide where the agent lives — website chat, WhatsApp, email, or some combination. The integration approach differs per channel.
Build Timeline
A Shopify AI agent covering order status, returns, and product queries typically goes live in 4–5 weeks:
- Week 1: Shopify API integration, product data review, return policy documentation
- Week 2–3: Agent build and channel integration
- Week 4: Testing with real Shopify data — realistic order scenarios, edge cases
- Week 5: Go live and monitoring
Most stores see a measurable reduction in support volume within the first week of live operation.
Related guides
- WooCommerce AI agent: automate support, orders, and returns
- AI agents for e-commerce
- AI agents for retail
- How we built a support agent for an e-commerce brand
- AI agent development services
Where This Doesn't Fit
If your differentiation is white-glove human service — small luxury brands, bespoke products, anything where the conversation itself is part of what people are paying for — an agent in front of customer messages can undercut the brand more than it saves on support cost. Use it on transactional queries only, and keep humans on the rest.
The other place we've watched this go sideways: stores where product data is genuinely unreliable. If variants are wrong, stock counts lag reality, and descriptions contradict each other, the agent will confidently surface that mess at scale. Fix the source of truth first.
If you want a realistic picture of what automation would look like for your specific store — including the parts we'd push back on — we'll talk it through.
Talk to us about your store — tell us your monthly order volume and top support query types, and we'll be honest about what's worth automating.
Frequently Asked Questions
How much does a Shopify AI agent cost to build?
A custom Shopify AI agent covering order status, returns, and product questions typically runs £8,000–£20,000 depending on complexity, the number of integrations, and which communication channels are included. Off-the-shelf app alternatives like Tidio AI or Gorgias AI cost £50–£300 per month. For stores doing under £300K annually, the app route usually makes more financial sense. For stores above that threshold with specific workflows, the build cost pays back within 6–12 months through reduced support headcount and improved cart recovery.
Will the agent work with my existing Shopify apps?
In most cases, yes. The agent connects to Shopify via the Admin API, which sits alongside your existing apps rather than replacing them. Returns integrations (Loop, AfterShip, WeSupply) connect via their own APIs. Your existing support inbox tool (Gorgias, Freshdesk) becomes the escalation destination for queries the agent hands off. The main constraint is that some Shopify Plus features and certain third-party app data require additional API scopes or separate integration work.
Can the agent handle queries in languages other than English?
A custom-built agent can be configured to handle multiple languages. The most practical approach for UK and US stores is English with Spanish as a secondary language, since those cover the majority of customer bases. The agent detects the language of the incoming message and responds accordingly. Product data and policy documentation need to exist in the target languages for the agent to answer accurately — it can't translate inaccurate source data into a correct answer.
What happens when the agent gets something wrong?
The agent is built with guardrails: it escalates to a human rather than guessing when it encounters a query it cannot answer confidently. During the first 30 days, we review every escalation and every conversation where the customer indicated dissatisfaction. Errors in the early period are almost always traceable to gaps in product data or policy documentation rather than model failure — which is why the pre-build data review matters. The agent does not fabricate information; it either answers from the data it has access to or it escalates.
How do I handle returns if my policy has exceptions?
Returns that fall outside standard policy — outside the time window, items with specific conditions, unusual circumstances — are routed to your team with full context: what the customer said, what the order data shows, and a flag indicating why the agent didn't auto-resolve it. Your team gets everything they need to make a judgment call without asking the customer to repeat themselves. Some stores build a simple exception approval flow: the team member approves the exception in a Slack message and the agent generates the return label automatically once confirmed.
Does the agent work on WhatsApp as well as the website?
Yes, and for many Shopify stores, WhatsApp is actually the higher-value channel. WhatsApp Business API integration means the agent handles the same queries — order status, returns, product questions, cart recovery — through WhatsApp conversations. The abandoned cart recovery flow is particularly effective on WhatsApp because open rates are significantly higher than email. WhatsApp integration requires a verified WhatsApp Business API account, which adds 1–2 weeks to setup but is worth doing from the start if your customers already message you there.
How long before the agent reduces my support volume noticeably?
Most stores see a measurable change in the first week. The queries that drop immediately are the high-volume, pattern-driven ones: order status, tracking updates, basic returns eligibility. The more nuanced queries — product comparisons, complaints, edge-case returns — take longer to handle well because the agent needs calibration against real conversation data. After 30 days, a well-built agent typically handles 70–80% of contact volume without human involvement. The remaining 20–30% are the cases that genuinely need a person, which is the correct outcome.