Case pattern: a shop that used AI for support without losing customers
A composite pattern showing how an online store put AI on first-line support, what it answered, where it handed over, and what changed as a result.
The short answer
This is a composite pattern rather than a named client. The shape: an online store whose support inbox and chat were dominated by five question types, where is my order, what is your returns policy, can I change my address or order, is this item available in another size or variant, and how do I return something. Together those made up the clear majority of contacts, and every one of them could be answered from the order record and the published policies. The intervention put an assistant on first-line contact that answered exactly those five, grounded strictly in the store’s own data and policies, announced itself as automated in its first message, and handed over to a person immediately on anything else, on any expression of frustration, and on any question its grounding did not clearly answer. What protected satisfaction was speed: routine answers arrived instantly at any hour, and the customers who needed a person reached one faster because the queue was shorter. The decision that made it work was the hand-over design, not the assistant.
What was automated and what was not
| Contact type | Handled by | Why |
|---|---|---|
| Where is my order | Assistant, from order and carrier data | Factual, verifiable, high volume |
| Returns policy questions | Assistant, from the published policy | Factual, published, high volume |
| Address or order changes | Assistant proposes; system applies within rules; person for anything outside | Rules decide what is changeable |
| Availability and variants | Assistant, from product data | Factual and current |
| How to return | Assistant, with the returns link and steps | Procedural |
| Complaints and dissatisfaction | Person, immediately | Judgement and goodwill |
| Refund and goodwill decisions | Person | Money and discretion |
| Damaged or wrong items | Person, with the assistant collecting photos first | Judgement, with preparation automated |
| Anything unclear in the data | Person | No guessing |
| Repeat contacts about the same issue | Person, escalated | A pattern the assistant should not continue |
How it was built
- Classified a month of contacts to find the five types and their share of volume.
- Wrote the answer material: policies in clear language, the rules on what may be changed and when.
- Connected order lookup with least privilege and identity checks appropriate to the data shown.
- Wrote the hand-over rules before the answer rules: what always goes to a person.
- Set the tone and the opening line, stating plainly that this is an automated assistant and how to reach a person.
- Tested against real past conversations, including the difficult ones, before launch.
- Launched on chat first, with a person monitoring every conversation for the first week.
- Reviewed weekly: resolution rate, hand-over rate, satisfaction, and a sample of transcripts.
What changed
Routine questions were answered in seconds, including at night and at weekends, which had not been possible before. Human handling time went to the contacts that needed judgement. Waiting times for a person fell because the queue was shorter. The complaints that mentioned the assistant were about the rare cases where it answered something it should have escalated, which the weekly transcript review caught and fixed.
What this means for you
An assistant on first-line support works when a few question types dominate the volume and each can be answered from your own order data and policies. Write the hand-over rules before the answer rules, say plainly that it is automated, refuse rather than guess, escalate on any sign of frustration, and review transcripts weekly. Done that way, satisfaction rises because everyone gets served faster, including the customers who needed a person all along.
Frequently asked questions
Did customers complain about talking to a bot?
In the pattern, the complaints that follow are about a rare wrong answer rather than about the assistant existing, for two reasons. The assistant said what it was at the start, so nobody felt deceived. And because it removed the routine volume, the people waiting for a human waited less. Customers object to being trapped and to being misled, not to a fast correct answer from a machine.
What did it not answer?
Anything not in its five categories, anything involving a complaint or a refund decision, anything where the customer expressed frustration, and anything where its grounding data did not contain a clear answer. In all of those it handed over immediately with the conversation so far, rather than attempting a guess. The list of what it would not attempt was written before the list of what it would.
How long did it take to build?
Weeks rather than months, because the hard parts were not technical: writing the five answer sets accurately, connecting order lookup securely, agreeing the hand-over rules and the tone, and testing against real past conversations. The assistant itself was a small piece grounded in that material.