AI for ecommerce brands: what to automate first in 2026
Short answer
An ecommerce brand should automate order status and returns replies from its Shopify data first in 2026, then product descriptions, review replies, supplier emails, Klaviyo flows and ad creative tests. A person approves every refund. Resolution rates vary. Intercom claims a 76% average for its Fin agent across 12,000+ customers, while one independent 500-ticket small-business test found 38%. Online sales were 28.3% of retail sales in Great Britain in September 2026, according to the ONS.
Deniz SancarFounder of Yarigai, London. Updated .
Yarigai is a done-for-you AI agency that builds these systems, so read this with that in mind. Every price and figure links to its source below. Read this page as Markdown.
At a glance
| Workflow | What it does | Tools it connects | What stays human |
|---|---|---|---|
| Order status replies | Answers "where is my order" from order and tracking data | Shopify, carrier tracking, helpdesk such as Gorgias, Zendesk or Intercom | Parcels delayed past an agreed limit |
| Returns and refunds | Checks the order against the returns window, sends the label, prepares the refund | Shopify, returns app, payment provider | Approving every refund |
| Product descriptions | Drafts copy from supplier specs in the brand's voice | Shopify admin, supplier spec sheets | Checking every claim on size, material and safety |
| Review replies | Drafts replies to new reviews | Review platform, helpdesk | Replies to negative reviews |
| Stock and supplier emails | Reads supplier emails, updates expected dates, drafts purchase orders at reorder points | Email, Shopify inventory, stock sheet or ERP | Approving every purchase order |
| Email flows | Writes and segments welcome, post-purchase, back-in-stock and win-back emails | Klaviyo, Shopify | Approving copy and checking consent |
| Ad creative tests | Produces ad variants, runs split tests, reports the winners | Meta Ads, creative library | Approving claims and budget changes |
Resolution rates above come from a vendor (Intercom) and an independent test reported by Superframeworks, sources 4 and 6.
Customer service on order status and returns
Order status and returns questions are the first build, because the answer sits in Shopify and the carrier's tracking data. The AI reads the order number or email address, looks up the order and the tracking events, and replies with the real status. It hands the conversation to a person when a parcel is late past the brand's limit, when the customer is angry, or when the question falls outside order data.
What resolution rates to expect
Vendor figures run well above independent ones, and vendors define "resolved" in different ways.
| Figure | Who says it | Type |
|---|---|---|
| 76% average resolution across 12,000+ customers | Intercom, for its Fin agent | Vendor claim |
| "above 50% automation" | Orthofeet, quoted on Gorgias's pricing page | Vendor-published customer claim |
| 38% resolution in an independent 500-ticket small-business test | Superframeworks, June 2026 | Independent, small sample |
| 38% to 53% across independent tests and Intercom's own case studies | Superframeworks, June 2026 | Independent summary |
Read the definition before comparing. Intercom counts an outcome when the customer confirms a fix, stops asking for help after Fin replies, or Fin completes a workflow, including a handoff. Plan on the independent range until your own data says otherwise.
Returns and refunds stay with a person
For online orders, GOV.UK says the customer has 14 days from receiving the item to say they want to cancel, then another 14 days to send it back. The brand "must refund the customer within 14 days of receiving the item back", including standard delivery if the customer paid for it. An AI can check the dates, send the returns label and prepare the refund in the payment provider. A person approves each refund before money moves.
Worked example for support
These figures are assumptions apart from the list prices, so replace them with your own helpdesk data.
- Assume 2,000 tickets a month and the independent 38% resolution rate, so the AI resolves 760 tickets.
- On Intercom Fin at $0.99 an outcome, 760 outcomes cost $752.40 a month, or $9,028.80 a year. On Gorgias's AI Agent at $0.90 an automated interaction on an annual contract, they cost $684 a month, or $8,208 a year, if all 760 fall inside the plan's included bundle. Gorgias charges $1.50 for each interaction above the bundle, so a plan that is too small costs more.
- On a custom build, model usage at an assumed 3,000 input and 500 output tokens per conversation on Claude Sonnet 5.5 ($2 and $10 per million tokens) is about $0.011 a conversation. That is $22 a month for all 2,000 conversations, before hosting and maintenance.
- If each resolved ticket saves an assumed 4 minutes of agent time at £14 an hour, 760 tickets save about 50.7 hours, or £709 a month.
The vendor fees and the staff saving are in different currencies, and this guide does not convert them. At low volume the per-resolution tools usually win, because a custom build has a fixed cost to recover.
Product descriptions and review replies
AI drafts product copy and review replies quickly, but UK consumer law makes the brand responsible for both, so a person approves them.
Product descriptions
The build drafts each description from the supplier's spec sheet, the brand's style guide and the product's Shopify fields, and marks every factual claim. A person checks size, materials, care, safety and any environmental claim before it goes live. Since 6 April 2025 the Digital Markets, Competition and Consumers Act 2024 governs unfair commercial practices, and the CMA says its provisions "replace and update the existing Consumer Protection from Unfair Trading Regulations 2008". A wrong material or a made-up feature in an AI draft can be a misleading action by the brand once it is published.
Review replies, never reviews
The CMA's fake reviews guidance defines a fake review as one that "purports to be, but is not, based on a person's genuine experience". Its examples of banned conduct include buying reviews "generated by software applications (such as bots)" and offering a refund or gift card to a customer who changes a negative review so it no longer reflects their experience. It says suppressing genuine negative reviews is "problematic", and that traders must take "such reasonable and proportionate steps as are necessary to prevent and remove from publication banned reviews". Encouraging only satisfied customers to leave reviews can count as cherry-picking.
That sets clear limits for an AI build.
- The AI drafts replies. It never writes, edits or hides a review.
- Replies to negative reviews go to a person, and no reply offers anything in return for changing the review.
- Review requests go to every customer, not only the happy ones.
Stock and supplier emails
Supplier email suits brands that buy stock in batches, because the same few facts arrive in every email. The build reads each supplier email, pulls out confirmed ship dates, quantities and price changes, and updates the stock sheet or inventory system. When stock falls below a reorder point, it drafts a purchase order with the last agreed price. A person approves every purchase order, because it commits money.
The same build can flag a supplier price rise against the product's margin before the next order goes out.
Email flows and ad creative tests
Email flows and ad tests are marketing builds where AI writes many variants and the data picks the winner.
Klaviyo flows
The build writes welcome, post-purchase, back-in-stock and win-back flows in Klaviyo from the brand's own past emails and order data, and segments by product and purchase date. A person approves the copy. Consent rules still apply. The ICO's PECR guidance lets a brand email an existing customer who bought, or negotiated to buy, a similar product, provided it gave "a simple way to opt out both when you first collected their details and in every message". The ICO says this guidance is under review after the Data (Use and Access) Act.
Ad creative tests
An ad testing build turns the brand's product photos and winning past ads into new variants, runs them as split tests, and reports which ones lower the cost per purchase. A person approves every product claim in the creative and any budget change.
What AI costs an ecommerce brand
AI support comes as a per-resolution fee from a helpdesk vendor, or as a one-off build plus model usage. One London agency lists a custom chatbot with CRM integration from £6,500 plus £399 a month (Softomate, updated 1 October 2026), and a simple automation from £1,500 plus £99 a month. These are "from" prices before VAT, and no independent survey of UK AI build prices exists.
The AI runs on top of the store's existing plan. Shopify's UK prices, checked on 7 November 2026, are below.
| Shopify plan | Annual billing | Monthly billing | UK online card rate | Fee on third-party payments |
|---|---|---|---|---|
| Basic | £19 a month | £25 a month | 2% + 25p | 2% |
| Grow | £49 a month | £65 a month | 1.7% + 25p | 1% |
| Advanced | £259 a month | £344 a month | 1.5% + 25p | 0.6% |
| Plus | From £1,800 a month | From £1,800 a month | 1.3% + 25p | 0.2% |
The guides to AI chatbot costs in the UK and AI agent costs compare more vendors, and the guide to Zapier, Make, n8n or custom code covers the automation platforms many stores already use.
How Yarigai builds this for ecommerce brands
Yarigai is a done-for-you AI agency in London. It builds support agents that answer from a client's own docs and order data, inbox triage, internal tools that replace spreadsheets and AI agents, inside the brand's own Shopify, helpdesk and email accounts. The brand owns the code, the data and the accounts from day one.
Yarigai also built and runs two marketing products that match the ad testing workflow. Mockingbird remakes winning Meta ads in a brand's style. Outtest is a set of AI agents that run split tests, listed on the Stripe App Marketplace.
A project starts with a free 30-minute scope call on Google Meet, and the brand keeps the notes whether it hires Yarigai or not. A blueprint of about a week, credited to the build, sets out a written plan, a clickable prototype, the monthly running cost, a fixed price and the payback maths. The build takes 2 to 6 weeks with a demo of working software every week, and a single workflow, such as order status replies, can go live in about two weeks.
A person approves anything that sends money, emails a customer or deletes data, so every refund and purchase order waits for a person. AI and hosting bills are paid by the brand at cost. Yarigai's guarantee is one sentence: "If the build isn't profitable for you, we won't build it." Its customer support work is described on the services pages, and Yarigai has no published ecommerce case study.
Common questions
What should an ecommerce brand automate first with AI?
Order status and returns replies, because the answers sit in Shopify and tracking data, so the AI replies with facts instead of guesses. After that come product descriptions, review replies, supplier emails and email flows. Refunds and purchase orders stay with a person.
What resolution rate does an AI support agent get?
Intercom claims a 76% average across 12,000+ Fin customers, which is a vendor figure. One independent 500-ticket small-business test found 38%, and Superframeworks puts independent tests and Intercom's own case studies at 38% to 53%. Definitions of "resolved" differ, so test on your own tickets.
How much does an AI customer service agent cost for a Shopify store?
Off-the-shelf agents charge per resolution, $0.99 on Intercom Fin and $0.90 to $1.00 on Gorgias's AI Agent (USD, November 2026). A custom build costs from £6,500 plus £399 a month on one London agency's list, with model usage around $0.011 per conversation on Claude Sonnet 5.5.
Can AI write product descriptions for my store?
Yes, as drafts from the supplier's spec sheet and your style guide. A person should check every factual claim before publishing, because the DMCC Act's unfair commercial practices rules apply to what the brand publishes from 6 April 2025, whoever drafted it.
Can AI reply to customer reviews?
It can draft replies, and a person should approve replies to negative reviews. The CMA's fake reviews guidance bans fake reviews, including bot-generated ones, and treats offering a refund to change a negative review as banned conduct. An AI should never write, edit or hide a review.
Should AI approve refunds automatically?
No. A person should approve each refund before money moves. The AI can check the order against the 14-day cancellation window, send the returns label and prepare the refund, so the approval takes seconds.
Sources
Checked on or before 7 November 2026. Prices change, so check the live page before you decide.
- Internet sales as a percentage of total retail sales (ratio) (%), Office for National Statistics, released 23 October 2026.
- Shopify pricing, Shopify, checked 7 November 2026.
- Intercom pricing, Intercom, checked 7 November 2026.
- Fin, the AI agent, Intercom, checked 7 November 2026.
- Gorgias pricing, Gorgias, checked 7 November 2026.
- 10 Best AI Customer Support Tools for 2026 (Real Resolution Rates, Real Per-Ticket Costs), Superframeworks, published 12 June 2026.
- Unfair commercial practices (CMA207), Competition and Markets Authority, published 4 April 2025, updated 18 November 2025.
- Fake reviews (CMA208), Competition and Markets Authority, published 4 April 2025.
- Accepting returns and giving refunds, GOV.UK, checked 7 November 2026.
- Electronic mail marketing, Information Commissioner's Office, checked 7 November 2026.
- AI Automation Pricing for UK Businesses: What Every Service Costs and What ROI to Expect in 2026, Softomate, updated 1 October 2026.
- Claude pricing, Anthropic, checked 7 November 2026.
Keep reading
How much does an AI customer service chatbot cost in the UK?
A custom UK support chatbot costs £5,000 to £20,000 to build in 2026, while off-the-shelf bots charge $0.99 to $2.00 per resolution plus seats.
Read moreHow much does it cost to build an AI agent for a small business?
A first AI agent costs £6,500 to £8,000 to build on published UK agency price lists in 2026, plus about $6 to $240 a month in model usage at 2,000 runs.
Read moreZapier, Make, n8n or custom code: which should a small business use?
A two-action workflow run 50,000 times a month costs $489 on Zapier, $107 on Make and $5 as custom code on Cloudflare (November 2026). Here is when each wins.
Read moreIs it cheaper to build your own tools than pay for SaaS?
Building beats SaaS only when the subscriptions you cancel cost more than upkeep at 15% to 25% of the build a year. Here is the payback formula and 10 real receipts.
Read moreWant the payback maths for your store?
Book a free 30-minute scope call. If the build isn't profitable for you, we won't build it.