Skip to content

Customer support

Reference build

First response in under a minute, without a confidently wrong answer

A support agent that resolves routine tickets end to end through narrowly scoped tools, states its confidence before acting, and escalates rather than guessing.

<60sfirst-response time

At a glance

Duration
17 weeks
Team size
1 person
Engagement
New build
Project type
AI & automation
Industry
B2B SaaS

The situation

The challenge

Routine tickets — order status, refunds, password resets — consume the same triage attention as complex ones. But an agent that is occasionally confidently wrong damages customer trust more than a slow human queue does, so autonomy had to be earned rather than assumed.

What we did

Narrow the tools, enforce the limits in code, and make escalation the default whenever confidence is low.

The calls that mattered

  • Tool-scoped access, not system access

    Order lookup, account lookup, refund issuance with a hard cap, ticket updates, escalation — each validated in code independent of the model's reasoning. The limits are not instructions the model can talk itself out of.

  • Confidence stated before consequences

    The agent declares confidence before any consequential action, and anything below a calibrated threshold escalates automatically instead of proceeding.

What changed

of tickets auto-resolved
68%of tickets auto-resolved
first-response time
<60sfirst-response time
escalation accuracy
96%escalation accuracy
customer satisfaction
+22%customer satisfaction
  • 68% — no human involvement
  • <60s — down from over four hours
  • +22% — post-rollout

68% of tickets resolved without a human, first response under a minute instead of four hours, 96% escalation accuracy and a 22-point satisfaction improvement.

Services used

  • Tool-scoped support agent with code-enforced limits
  • Confidence-based escalation and human handoff
  • Full reasoning and tool-call audit trail

What we would do differently

Every project has one of these. Publishing it is the point — a case study with no regrets in it is marketing, not evidence.

The valuable engineering was not the reasoning quality, it was the tool design and the hard limits enforced outside the model. Starting with conservative thresholds and relaxing them against real evaluation data is what made it deployable.

Services used

More work

Ready to start?

Start a project like thisWork