AI customer service for ecommerce: what to automate and what to keep human

A bot trained on your FAQ page can tell a customer what the return policy says. It can't start the return. The tickets worth automating end with something changing in another system: a label created, an address edited, a subscription skipped. That means the agent needs a connection to your store, your subscription platform, your returns tool, and your carrier tracking.

Split that access into two levels and grant each one deliberately.

• Read access covers order status, tracking, order history, and whether an item is still inside the return window. Low risk, and enough to close most order status tickets.
• Write access covers editing an address, cancelling an unfulfilled order, issuing a return label, skipping or swapping a subscription, and refunding. Give it one action at a time, with a dollar limit on anything that moves money, and log every action the agent takes.

Subscription changes deserve extra care. When a customer asks to cancel, the agent can offer a skip or a pause first, and some of those customers stay. The churn rate guide covers how to measure whether it's working.

Deflection means the customer didn't reach a person. Resolution means their problem got fixed. A customer who gives up on the bot and files a chargeback counts as deflected.

Vendors bill on resolution, and each one defines it its own way. Intercom counts an outcome for its Fin agent when the customer confirms the issue is resolved, when they don't ask for more help after Fin responds, or when Fin completes a workflow, handoffs included. Gorgias defines an automated interaction as a request fully resolved without a human agent, and if a human replies within 72 hours of an AI Agent resolution, it bills that conversation as a regular helpdesk ticket instead. Zendesk counts an automated resolution when the AI agent resolves a request without escalation to a human.

So a customer who goes quiet can count as resolved. Before you trust the vendor's dashboard, measure it yourself: of the conversations marked resolved, how many came back about the same order within a week?

Write these down before launch, and have your CX manager own them. They are policy, and the agent follows them the same way every time.

• Any refund, credit, or reshipment above a dollar limit you pick.
• The customer asks for a person. Hand off on the first request.
• The same customer contacting you a second time about the same order.
• Words like chargeback, lawyer, allergic reaction, injury, or a mention of the press.
• Anything the policy doesn't cover. The agent should say it's checking with the team, and stop there.
• Your highest-value customers, if you can identify them in the helpdesk.

The handoff should carry the order, what the customer asked, and what the agent already tried, so the person picking it up doesn't ask the customer to repeat everything.

Get the policy right before you automate it. In 2024 a Canadian tribunal ordered Air Canada to pay a customer $812.02 after its website chatbot described a bereavement fare refund the airline didn't actually offer. The tribunal's view was that it makes no difference whether wrong information comes from a static page or a chatbot.

Track the agent the way you'd track a new hire, on a weekly sheet, split by ticket type.

• CSAT on agent-resolved conversations compared with human-resolved ones in the same ticket type. A gap on returns tells you something different from a gap on complaints.
• Reopen rate: resolved conversations the customer came back to. This is the number that catches false resolutions.
• Escalation rate by reason. A sudden jump usually means a policy changed and nobody told the agent.
• Wrong actions: refunds, labels, or edits a person had to undo.

Read the conversations too. A human QA team can only sample a slice of tickets, but software can review all of them and flag the ones that look wrong for a person to read. That's why the QA seat goes away on our Zero Hire Org Chart and an agent operations seat appears: someone has to watch what the agents do and fix them when they break.

The options fall into two groups. We don't recommend a specific vendor here.

• Helpdesk-native AI, sold by the helpdesk you already use. It sees your ticket history and macros out of the box, and it's usually the shortest path to launch.
• Standalone AI agents that connect to your helpdesk and store. They can go deeper on custom workflows, at the cost of another integration to maintain.

Most price by the resolution. Current list prices from the vendors' own pricing pages:

• Intercom: $0.99 per Fin outcome, charged once per conversation.
• Gorgias: $0.90 per AI Agent interaction on annual plans or $1.00 on monthly plans, with overage above the plan's allowance at $1.50. Each interaction also counts as a helpdesk ticket.
• Zendesk: a set number of automated resolutions included per agent per month (5 or 10 depending on plan), then $1.50 each committed or $2.00 pay as you go.

Compare those with what a human-handled ticket costs you today: support payroll divided by tickets closed. Then remember that the vendor's resolution count sets your bill, which is one more reason to check it against your own reopen rate.

On the Zero Hire Org Chart, a made-up $50M subscription brand goes from 12 support agents to 3. The three keep refunds above the limit, angry customers, and anything strange. The CX manager stays and owns the policies the agent follows. Most brands get there by not backfilling as people leave.

To see what your own support headcount costs against your revenue, run it through the org chart roast, or plan the next year in the headcount planning template.

Roast your org chart