Skip to content

E-commerce

Support agent

A support agent that resolves the routine tickets end-to-end and hands the rest to a human with full context — scoped tightly enough to be trusted with real customer conversations.

  • LangGraph
  • Anthropic
  • River
47%

tickets auto-resolved

first-response time

-38%

In short

  • Routine tickets resolve end-to-end; only what needs judgement reaches a human.
  • Every escalation arrives with the full thread and order context — no repeating.
  • Agent runs are durable, so a campaign spike queues instead of dropping customers.

The engagement

Sector
Consumer e-commerce
Company shape
Scale-up, high ticket volume
Engagement
Agent build + handoff design

Anonymised by policy

Every engagement runs under NDA — we protect our clients' confidentiality as a matter of policy, and we'd extend the same discretion to your work. The constraints, the shape, and the outcome here are real; the identity is redacted.

The challenge

The support queue grew with every campaign, and hiring couldn't keep pace. Most tickets were routine and repetitive, but the long tail needed a human — so a blunt 'automate everything' bot would have made the experience worse.

The approach

  1. 1
    Scope to what's safe to automate

    We mapped the ticket taxonomy and automated only the categories the agent could resolve reliably, with tools scoped to safe actions.

  2. 2
    Context-rich handoff

    When the agent hands off, the human inherits the full conversation, the attempted resolution, and the relevant order context — no repeating themselves.

  3. 3
    Durable, observable runs

    Agent runs are durable jobs (River) so a slow tool call or retry never drops a customer, and every run is traceable.

The solution

A LangGraph agent handles the routine categories with tools scoped to safe actions, resolves what it can end-to-end, and escalates the rest with full context attached. Runs are durable background jobs so nothing is lost to a timeout, and the whole thing is observable per conversation.

The system we built

  1. Classifyticket → category
  2. Resolvescoped tools
  3. Confidence checkcan we close it?
  4. Escalatecontext-rich handoff
Abstract by design — the architecture we built, not a client screen. No confidential data is shown.
Illustrative
A context-rich handoff — the human inherits the full thread
Illustrative
Walkthrough: classify → resolve → escalate with full context

The results

tickets auto-resolved

47%

of incoming tickets, end-to-end

first-response time

-38%

vs. the pre-agent baseline

Data & safety

Guardrails
Tools are scoped to safe actions; anything outside the automatable envelope escalates to a human.
Evaluation
Per-category resolution and escalation quality are tracked so the safe-to-automate set stays honest.
Data handling
Customer data stays within the support stack; every run is durable and auditable.

The stack

  • LangGraph
  • Anthropic
  • River
Capability: AI Agents & Automation

Questions we get

We automate only the ticket categories it can resolve reliably, scope its tools to safe actions, and escalate anything outside that envelope to a human with full context — the agent's job is bounded on purpose.

The full conversation, the agent's attempted resolution, and the relevant order context — so the customer never has to repeat themselves and the human starts with everything already in hand.

Agent runs are durable background jobs, so a campaign spike or a slow tool call queues and retries instead of dropping a customer or taking the feature down.

More selected work

// Let's build

Want an outcome like this on your problem?

Tell us where you are. We reply within a day with a concrete next step.