Support chatbot with KB
HIGH-VOLUME E-COMMERCE · Consumer e-commerce

A support agent trained on two years of tickets and the product KB. Resolves four in ten tier-one queries on its own.

Tier-1 tickets fully resolved

~40%

3 weeks · 1 PM + 1 engineer

First-token latency

< 2s

CSAT on resolved chats

~92%

Kick-off to production

3 weeks

“We were ready for it to be okay. We were not ready for it to be better than us on the boring stuff.”

Head of Customer Service, consumer e-commerce

Before

Where the team was when we picked this up.

  • The same handful of questions made up most of the support inbox.
  • Off-the-shelf bots gave answers that contradicted product policy.
  • When the bot escalated, agents lost context and had to start over.

What we built

Knowledge grounding

Indexed the product KB and the last two years of resolved tickets. Answers cite a passage. Anything not in the corpus is escalated, not guessed.

Clean handover

When the bot escalates, the human gets the transcript, the proposed answer and the reason it stopped. No restarts.

Tone match

Style-tuned on the brand’s actual support replies. Reads like the team, not like a generic helpdesk.

What changed

  • Tier-one queue cleared faster, agents spent more time on edge cases.
  • CSAT on bot-resolved tickets sits above the human baseline.
  • Refund and policy answers are now consistent across the team.

After

Same team. Same week. Different shape of work.

Stack

Anthropic ClaudeOpenAI embeddingspgvectorZendesk APITypeScript

Timeline & team

3 weeks · 1 PM + 1 engineer

Got a workflow like this one?

Book a working session. We will tell you whether this is a four-week build or something bigger, and what it would take to ship it.

Book a working session

Cookies. Sadly not chocolate chip.

We use cookies to keep the site working, understand what is useful, and avoid shouting ads into the void. You can accept all, reject non-essential, or choose your own settings.

More detail lives in our Privacy Policy and Terms.