How Can AI Improve E-Commerce Customer Service ?

Most e-commerce contacts are the same handful of questions: where is my order, can I return this, why was I charged twice. An AI agent connected to your order system can answer those from real data and actually complete the action. Anything unusual goes to a person with the conversation and the order history already attached, so the customer never repeats themselves.

Scripted Bot vs. AI Support Agent

Deflection Is the Wrong Target

Plenty of support tools report a deflection rate, meaning the share of contacts that did not reach a human. It is a number that improves when customers give up, which is not the same as being helped.

Measure resolution instead — the share of contacts fully handled without a person, where the customer did not come back about the same thing within a week. That second clause matters. A bot that closes a conversation and generates a second contact tomorrow has moved work, not removed it.

Then watch your CSAT split by whether the contact was handled by the agent or escalated. If the agent's satisfaction is materially worse, you are trading customer experience for cost and should know it.

Where to Start

Start with order status. It is the highest-volume contact type in almost every store, it needs no judgement, and the answer is sitting in a system the agent can read. Get that resolving properly before adding anything else.

Returns come next, and they are a step up because eligibility depends on your policy and the specific order. Encode the policy as rules the agent checks, not as text it interprets — you want a customer's return approved because it met the rule, not because the model found the wording persuasive.

The Rules This Has to Satisfy

If you sell into the EU, Article 50(1) of the EU Artificial Intelligence Act, Regulation (EU) 2024/1689 puts the duty on the provider: “Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system” — unless that is obvious to a reasonably well-informed and observant person in context. You become that provider under Article 3(3) the moment the assistant ships under your own name or trademark, which is the normal case for a storefront widget, so this is your design duty rather than your vendor's. "Obvious" is not a defence you want to litigate over a support widget branded with a person's name. Build the disclosure as a first-turn message the customer cannot miss, make it survive a hand-off between sessions, and keep it as a system property rather than a prompt instruction — a prompt-level instruction is one jailbreak away from being absent.

Returns logic has a statutory floor underneath it. The Consumer Rights Directive 2011/83/EU gives consumers 14 days to withdraw from a distance contract, with no reason required and all payments including standard delivery costs reimbursed. So eligibility has to be computed as a rule against the contract date and the goods category — including the Article 16 exceptions — rather than inferred by the model from your returns policy page. A policy document stricter than the Directive will teach the model the wrong answer. Compute the statutory entitlement deterministically, and leave the model to explain the outcome and complete the action.

Selling into India adds a clock. Under the Consumer Protection (E-Commerce) Rules, 2020, every e-commerce entity must appoint a Grievance Officer who acknowledges a consumer complaint within forty-eight hours and redresses it within one month of receipt. That makes escalation a timed workflow with a named owner, not a queue. Design the hand-off so it stamps that clock the moment the agent classifies a contact as a grievance — a bot that keeps trying to resolve for three days has already spent the acknowledgement window.

Which Model We'd Shortlist for This

Rates below are per million tokens, input then output, taken from each provider's own pricing documentation and checked on 7 August 2026. Each name links to that model's page, where the source and the capture time are shown in full.

Claude Haiku 4.5 — $1/$5 with a 200,000-token window and a 50% batch rate. That window holds an order history alongside the returns policy without chunking either of them.

Gemini 2.5 Flash-Lite — $0.10/$0.40 for intent classification and deflection in front of everything else, so only the tickets that need a costlier model reach one.

Mistral Small 4 — $0.15/$0.60 with up to 90% off cached input. The policy and tone preamble repeats on every single ticket, which is exactly where that discount lands. Apache 2.0 weights if EU customer data has to stay in-house.

GPT-5.4-mini — $0.75/$4.50 with a 400,000-token window and no long-context surcharge published, so a long conversation thread does not change the rate mid-chat.

Where This Fits

This is one part of our work in AI for E Commerce. See the full set of AI use cases for the equivalent in other industries and functions.

Frequently Asked Questions