AI for E-commerce: Where It Pays and Where It Doesn't

The verdict. If you run a small store doing under roughly 300 support tickets a month, the AI worth paying for in 2026 is the AI you’re already paying for: the copy and store assistants bundled into your platform. A dedicated per-resolution AI support agent is a volume play, and below that line the invoice tends to land within a rounding error of the labour it replaces. Automate order-status questions first, or don’t automate at all yet.

This is a Workflow Decision Lab piece, not a “15 best AI tools for Shopify” list. Named operator, real published prices with links, the arithmetic that actually decides this, and the two places it reliably goes wrong.

Proof status. Prices below are quoted from vendor pricing pages as of 21 August 2026 and linked at each figure. Where a number comes from a secondary source rather than the vendor’s own page, it is labelled (secondary). The worked example uses clearly-labelled assumptions for ticket mix and labour cost. Those are inputs you should replace with your own, not measurements we took. Vendor pricing changes often; re-check the links before you commit.

1. The operator, precisely

Maya runs a home-goods store on Shopify. Roughly 400 SKUs, about 1,800 orders a month, two people handling support between other jobs. She is not drowning. She’s being pitched.

Her actual leak isn’t “not enough AI.” It’s that the same four questions eat her mornings:

• “Where is my order?”. By a wide margin the most common, and entirely answerable from data she already has.
• “Can I change/cancel this?”. Time-sensitive, and wrong answers cost real money.
• “Does this fit / what’s it made of?”. Pre-purchase, and the answer is sitting in a product page nobody reads.
• Returns and exchanges. Policy-driven, repetitive, occasionally emotional.

That mix matters more than any tool comparison. Three of those four are low-judgement and high-repetition, which is exactly what AI is good at. One of them is time-sensitive and expensive to get wrong, which is exactly what it’s bad at.

2. What the AI actually does

Strip the marketing and there are three distinct jobs, and they have completely different economics:

1. Catalogue and copy. Generate product descriptions, alt text, meta descriptions, email subject lines. Cheap, low-risk, already bundled.
2. Support deflection. An agent reads the customer’s message, looks up the order, and answers without a human. Priced per resolution, and this is where the money goes.
3. Order-status automation (WISMO). A narrow slice of #2. Deterministic: look up the order, read the tracking, reply. It is the highest-volume, lowest-risk automation in the whole store.

Most stores buy #2 when what they needed was #3.

3. The four ways to do it

Approach	Good for	Cost shape	The catch
Built-in platform AI (Shopify Magic / Sidekick)	Product copy, store questions, admin tasks	$0 extra — included in your existing plan	Shallow on support; won’t resolve tickets for you
Helpdesk + AI agent (Gorgias, Intercom Fin)	Real ticket deflection at volume	Base plan plus a per-resolution fee	You pay twice per AI-handled ticket — see §4
Chat-first AI (Tidio/Lyro and similar)	Pre-purchase questions, lower volumes	Lower monthly floor, capped conversations	Weaker when it must act on live order data
Fix the process first	Almost everyone under ~300 tickets/mo	$0	Unglamorous. Also usually correct.

That last row is not a joke, and §6 is about it.

4. Cost: the real 2026 figures

Platform AI — already paid for. Shopify’s plans run $39 / $105 / $399 per month for Basic / Grow / Advanced (US pricing, monthly billing; annual is roughly 25% less). Shopify Magic and Sidekick are included at no additional cost on every tier. Their pricing page lists “Sidekick, AI assistant for commerce” as a core feature across plans. (Shopify pricing, note the page geolocates, so it may show your local currency; the USD plan figures are secondary.)

Per-resolution support AI, where the invoice comes from.

• Intercom Fin: $0.99 per outcome, on a base plan of $49/month that includes 50 resolutions. A billable outcome is a resolution or a procedure handoff; lead qualification is billed higher. You’re charged once per conversation no matter how many actions it takes. (Fin pricing docs, primary source.)
• Gorgias: helpdesk tiers around $10 / $60 / $360 / $900 per month for 50 / 300 / 2,000 / 5,000 tickets, priced by ticket volume rather than per seat. AI Agent adds roughly $0.90–$1.00 per conversation on top of the ticket you’re already being billed for. Overage runs about $0.32–$0.40 per extra ticket. (secondary, Gorgias’ own pricing page renders its table client-side; verify at gorgias.com/pricing before committing.)
• Tidio: Free tier at $0 (50 conversations), Starter $24.17/mo, Growth from $49.17/mo (up to 2,000 conversations), Plus from $300/mo. The Lyro AI agent as a standalone add-on starts at $32.50/mo. (Tidio pricing, primary source, verbatim.)

Now the arithmetic that actually decides it. Maya’s 1,800 orders produce roughly 270 tickets a month (assumption: a 15% contact rate — measure your own). On a Gorgias-style plan that’s the $60 tier. Suppose the AI resolves 45% of them, Intercom’s published case studies put real-world Fin resolution rates in the 42–50% band (secondary), so that’s a fair planning number rather than the 80% in the sales deck.

• 270 tickets × 45% ≈ 122 AI resolutions
• 122 × ~$0.95 ≈ $116/month in AI fees, plus the $60 base ≈ $176/month
• Those same 122 tickets at (assumption) 4 minutes each ≈ 8.1 hours, at (assumption) $20/hour ≈ $161

So at Maya’s volume the tool costs slightly more than the labour it removes. Not catastrophically, but nowhere near the “cut support costs 40%” claim, and that’s before the time she spends configuring it.

The break-even moves in your favour on exactly two things: higher volume (the base fee amortises) and a more repetitive ticket mix (resolution rate climbs). It moves against you when tickets need judgement, because a failed AI attempt still costs you the human touch afterwards.

5. Where it fails

• Double-billing surprise. With helpdesk-plus-agent pricing you pay the ticket fee and the resolution fee for the same conversation. Model your bill as base + (tickets × resolution rate × per-resolution fee), not as a flat subscription.
• Confident wrong answers on order changes. “Can you cancel it?” has a real deadline attached to it. An AI that says yes after the warehouse picked the order creates a refund, a return shipment, and an angry customer. Put a hard human gate on anything that mutates an order.
• Resolution counted, problem not solved. Most vendors count a resolution when the customer stops replying. Customers stop replying when they give up, too. Watch your re-open rate and your follow-on ticket rate, not the vendor’s deflection dashboard.
• Stale catalogue data. AI answering “is this in stock” from an index that syncs hourly will lie to customers during exactly the sales spike where it matters.
• Generated copy at scale, indexed at scale. Bulk-generated product descriptions across hundreds of near-identical SKUs is how stores end up with thin, duplicative pages. Generate a draft, then edit, don’t publish raw.

6. When NOT to use it (do this first)

If you’re under roughly 300 tickets a month, these cost $0 and beat the tool:

1. Put tracking where they’ll actually look. A large majority of “where is my order” traffic is a shipping-confirmation email that got buried. Fix the email, add an order-lookup link in the header, and put tracking in the order-status page. WISMO volume drops before you’ve bought anything.
2. Answer the fit/material question on the product page. If the same sizing question arrives twenty times a month, that’s a product-page defect, not a support workload. Update the page once.
3. Write the returns policy in plain English and link it three places. Policy tickets are a documentation problem wearing a support-ticket costume.
4. Use macros before agents. Canned responses in your existing helpdesk resolve repetitive tickets at zero marginal cost, and they never invent a refund policy.

Do these first. If ticket volume is still painful afterwards, then the per-resolution math is worth running, and you’ll have a cleaner ticket mix, which makes the AI perform better anyway.

7. The 30-day test

Baseline to record before you switch anything on (one week is enough):

• Tickets per week, split by the four categories in §1
• Median first-response time and median time-to-resolution
• Your actual contact rate (tickets ÷ orders)

Then run the tool for 30 days and watch four numbers:

Metric	Keep it if
AI resolution rate	Holds at or above ~40% on your real ticket mix
Re-open / follow-on rate	Does not rise versus baseline
Total monthly cost	Comes in under your labour-equivalent (§4 arithmetic)
CSAT on AI-handled tickets	Within a point of human-handled tickets

The keep/drop rule: keep it only if resolution rate holds and re-opens didn’t rise and the invoice beats the labour it replaced. Two out of three means you bought a dashboard.

Bottom line

For a small store, the honest ranking is: use the platform AI you already pay for, fix the four process leaks that generate most of your tickets, and only then price a per-resolution agent against your own measured ticket mix. The tools are real and they work. The question is whether your volume is high enough for the arithmetic to land on your side of the line. Below ~300 tickets a month, it usually isn’t.

If you want this run against your actual numbers, send us the decision, order volume, contact rate, and ticket mix is enough to do the math.

---

More on this decision, three ways to look at it: