Customer service is where most small online businesses first meet AI in earnest, often because their helpdesk software switched it on. Automation can answer “where is my order?” at 2am and clear a post-sale backlog, but it also speaks for the business in a way a product-copy tool does not. This guide covers what to automate, how to design the hand-over to people, how to measure quality, and the legal points that apply as of November 2026.
Four levels of automation
It helps to separate customer-service automation into levels, because risk rises sharply as human review falls away.
| Level |
What the AI does |
Who speaks to the customer |
Typical risk |
| 1. Triage |
Tags, prioritises and routes incoming messages |
A person |
Low: misrouting delays a reply |
| 2. Drafting |
Suggests a reply; an agent edits and sends |
A person |
Low to moderate: depends on review discipline |
| 3. Bounded bot |
Answers a fixed set of topics (order status, delivery times, returns process) from your own data |
The bot, with a route to a person |
Moderate |
| 4. Open bot |
Answers anything, generating responses from a general model plus your content |
The bot |
High: invented policies, commitments, tone failures |
For most small shops, levels 1 and 2 deliver the bulk of the benefit. Triage gets urgent messages (failed payments, damaged deliveries, complaints) seen first. Drafting removes the blank-page problem for routine replies. Our practical guide to AI for small ecommerce suggests starting there and moving up only once you have quality data.
Making drafting work
Drafting succeeds or fails on whether agents actually read the drafts. Two habits help. First, ground the drafting tool in your own policies (returns, delivery, warranty) rather than letting it rely on general knowledge, so drafts quote your terms rather than an average retailer’s. Second, make editing easy and expected: if agents feel they are approving rather than writing, they skim. Periodically compare sent replies with the original drafts; if almost nothing is ever changed, either the drafts are excellent or nobody is checking.
Bounding a bot
If you deploy a customer-facing bot, restrict what it may discuss. A bot that can look up an order, explain your published returns process and hand over to a person is far easier to keep accurate than one that will answer any question. Explicitly exclude topics where a wrong answer creates a commitment: refunds outside policy, compensation, price matching, legal rights, product safety. For those, the bot’s only job is to collect details and pass the conversation on.
Designing escalation
Escalation is the part of the system customers judge you on, so design it before you design the bot.
- Define triggers. Hand over to a person when the customer asks for one, when the bot fails to resolve the issue after two attempts, when a message mentions a complaint, a legal right, injury, a safety issue or a payment dispute, and when sentiment is clearly negative.
- Make the human route visible. A “talk to a person” option should be offered, not hidden behind repeated deflection. Customers who feel trapped by a bot escalate elsewhere, publicly.
- Carry the context. The agent should receive the full transcript and any order details, so the customer does not repeat themselves.
- Set honest expectations. If no one is available until Monday, the bot should say so and confirm how the customer will be contacted.
- Close the loop. Feed escalated cases back into your knowledge base and the bot’s allowed topics.
Returns and refunds deserve special care because consumer rights apply regardless of what a bot says; our guide to returns, refunds and cross-border selling sets out the baseline the bot must not contradict.
Measuring quality
Deflection rate, the share of conversations that never reach a person, is the metric vendors lead with, and it is easily misread. A conversation the customer abandons in frustration counts as deflected. Better measures for a small team:
- Resolution confirmed by the customer, or no repeat contact on the same issue within, say, seven days.
- Accuracy on sampled conversations. Each week, read a fixed number of bot conversations and AI-drafted replies against a short checklist: was the information correct, did it match current policy, was escalation offered when it should have been, was the tone right?
- Escalation quality. How long escalated customers waited, and whether the agent had the context needed.
- Error log. Record every wrong answer found, its cause (outdated knowledge base, invented content, misunderstood question) and the fix.
- Complaints mentioning the bot, tracked separately.
In our view, sampling is non-negotiable. Language models fail unpredictably, and aggregate metrics can look healthy while a specific wrong answer is being repeated to every customer who asks a particular question.
Telling customers they are talking to AI
Article 50 of the AI Act has applied since 2 August 2026; that date was not changed by the 2026 AI Omnibus. Article 50(1) requires providers of AI systems intended to interact directly with people to design them so that the people concerned are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed, observant and circumspect person given the circumstances. Under Article 50(5) the information must be clear and distinguishable and given at the latest at the first interaction.
The formal duty sits with the provider, which is usually the chatbot vendor. But a shop that builds its own bot on a model and runs it under its own name may be the provider, and in any case a deployer controls how the bot is presented. The Commission’s guidelines on the Article 50 transparency obligations explain the scope of “directly interactive” systems with examples. Our practical view: do not rely on the “obvious” exception. Name the assistant as an automated assistant in the chat window header and in its first message. It costs nothing and removes the argument.
Breaches of the Article 50 obligations can attract fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher; for SMEs, the lower of the two applies.
GDPR Article 22 and automated decisions
Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or similarly significantly affects them. Such decisions are allowed only where necessary for a contract, authorised by law, or based on explicit consent, and even then the business must provide safeguards including the right to obtain human intervention, to express a view and to contest the decision.
Answering an order-status question is not that kind of decision. But some customer-service automations come closer than shops realise:
- automatically rejecting a refund or return request;
- closing or suspending an account for suspected abuse;
- blocking an order flagged as fraudulent with no human review;
- deciding eligibility for credit-style payment options.
The guidelines on automated decision-making and profiling adopted under the GDPR stress that human involvement must be meaningful: a person rubber-stamping the system’s output does not take a decision out of Article 22. The Court of Justice has also read “decision” broadly; in Case C-634/21 (SCHUFA) it held that an automatically generated credit score can itself be a decision where a third party relies on it heavily. The safest design is for the bot to gather information and a person to make any adverse decision, with the customer told how to challenge it. Proposals to amend parts of the GDPR were still under negotiation at the time of writing; until anything is adopted, the current text applies. For broader context, see our guide to GDPR obligations for online businesses.
Bots also handle personal data in every conversation, so the usual rules on processors, retention and security apply. Account-related flows deserve particular care: a bot that will change a delivery address or reveal order details must authenticate the customer first. Our guide to protecting ecommerce customer accounts covers the risks.
When the chatbot gets it wrong: Moffatt v. Air Canada
The best-documented example of liability for chatbot statements is Moffatt v. Air Canada, 2024 BCCRT 149, decided by the British Columbia Civil Resolution Tribunal on 14 February 2024. A customer booking flights after a family bereavement asked the airline’s website chatbot about bereavement fares. The chatbot told him he could apply for the reduced fare retrospectively, after travel. The airline’s own policy page said otherwise, and his later claim was refused.
The airline argued that it could not be held liable for information provided by its chatbot, at one point suggesting the chatbot was a separate legal entity responsible for its own actions. The tribunal rejected that, holding that the chatbot was part of the airline’s website and the airline was responsible for all the information on it. It found negligent misrepresentation: the airline had not taken reasonable care to ensure the chatbot was accurate, and it was reasonable for the customer to rely on it without cross-checking another page. He was awarded C$650.88 in damages, plus interest and tribunal fees.
The amounts were small and this is a Canadian small-claims tribunal decision, not binding precedent in Europe. Its value is as a clear illustration of a principle that consumer and contract law in most jurisdictions would likely reach by their own routes: what your bot tells customers is treated as what you told them. That is the strongest argument for bounding the bot’s topics and grounding its answers in your published policies.
What to do next
- Map your current tools against the four levels above, including AI features your helpdesk enabled by default.
- Write escalation triggers and a list of topics the bot must never decide.
- Label any customer-facing assistant clearly as automated, from the first message.
- Move any adverse decision about an individual (refund refusals, account closures, fraud blocks) to a person, with a route to contest it.
- Start a weekly sample review and an error log before expanding automation.
Our guide to building customer trust as a small online retailer explains why the hand-over to a person is often the moment that decides whether a customer comes back.
This article is general information, not legal advice.