How it started
The customer service team at our client — a mid-sized e-commerce company — was exhausted.
They weren't doing the work they'd signed up for. Instead of helping customers with genuinely tricky problems, they were spending 70% of their time answering the same five questions: where's my order, how do I return it, did you get my email, what's your address, can I change my delivery.
The inbox was a permanent 400-ticket backlog. Response times had crept out to 24 hours. Two senior agents had quit in six months. The director, Emily, was very politely asking us if AI could help — and also, quietly, whether we'd tell her if it couldn't.
What we promised not to do
Before anything else, we agreed on what we weren't building:
- No customer-facing bot pretending to be human. If you're talking to a machine, you should know.
- No "deflection at all costs." Success wasn't fewer human conversations. Success was fewer frustrating conversations.
- No dark patterns hiding the hand-off. If the AI wasn't sure, it said so, and a human took over within two minutes.
Emily loved this list. It mattered that we cared about the same things she did.
What we built
A calm little system:
- A domain-grounded assistant that answers using the company's own help centre, return policy, and order system — not the open internet.
- A clear tone of voice, co-written with the support team over two workshops. Warm, plain, slightly dry. Not perky.
- Confidence-based hand-offs — below a threshold, the assistant opens a ticket and a human replies. Above it, the assistant answers.
- Evals on every question type, reviewed weekly by the team, not just the data scientist.
We went live with a polite message — "Hi, I'm a support assistant. I can help with orders, returns, and shipping. For anything else, I'll bring in a colleague." — and the hand-off button was one tap.
What changed
Three months in:
- 60% lift in customer satisfaction scores.
- 80% faster first response on routine queries (from a day to under a minute).
- 85% first-contact resolution — the assistant fully resolving the query by itself.
- ~45% lower cost per ticket — not from firing anyone, but from handling more volume with the same team.
The team's workload shifted. They stopped answering "where's my order" and started handling genuinely novel cases — the angry email, the edge-case refund, the order that went to the wrong continent. That's work that needs a person, and the people were better at it because they weren't worn out.
"The numbers are great. But the real change is that my team is excited about their work again."
— Emily, Customer Service Director
What we got wrong the first time
Our initial hand-off threshold was too high. We wanted the assistant to "lean in" and try. It turned out customers hated that — they could smell uncertainty in the answers.
We rebuilt the threshold in week three based on real conversations. Lower confidence, faster hand-off, better outcomes. Humility is a useful feature.
Where it's going
The team is now using the same infrastructure to draft internal replies — the assistant writes a first pass of the trickier tickets, and a human edits. It's cut per-ticket time further, without taking a single warm conversation away from a customer.
If you're looking at support, the question to ask isn't "can AI replace my team." It's "what would my team love to stop doing?" Start there. Come talk to us when you're ready.