Skip to content
AI Agent

Deskhand: AI customer support agent

An agent that answers from approved documentation, performs account actions within strict limits and hands over cleanly to people.

Sector
SaaS
Typical client
Subscription software company with a global customer base
Typical timeline
12 weeks
Typical team
3 engineers, 1 AI lead
Support agent resolving a billing query with visible tool calls and approval
Agent analytics covering resolution rate, handovers and customer satisfaction
Quality review of sampled conversations with accuracy scores and outcomes

About this project

Support teams spend much of the day on questions that follow a known process. Basic chatbots frustrate customers because they can answer questions but cannot actually do anything.

Deskhand can read account data and take limited actions through approved tools. Riskier steps need a person's approval, every action is logged, and evaluation suites run before each release.

Objectives

  • Resolve routine tickets without compromising accuracy
  • Allow safe account actions under explicit permissions
  • Make every agent decision reviewable by the support team

What’s delivered

  • Permission-aware retrieval across help content
  • Scoped tools for billing, plans and account lookups
  • Evaluation suite built from historic tickets
  • Approval gates and human escalation paths
  • Quality, cost and escalation dashboards

Target outcomes

50%

automated resolution goal

95%

evaluation accuracy bar

100%

actions logged for audit

<30s

first response target

Technology

  • Claude
  • Python
  • pgvector
  • LangGraph
  • Langfuse

Services involved

Have a similar challenge?

Book a free consultation. We’ll scope your project, pressure-test the plan and give you an honest estimate, with no commitment required.