Customer Support

Support is where AI assistance meets real customers at volume, so this track is built around one principle: Claude drafts, agents decide, and uncertainty is never hidden. A workflow that makes agents faster on routine tickets and reliably steps aside on hard ones will earn trust. One that quietly guesses will lose it on the first bad day.

Workflow patterns

Ticket triage

Classification is only as good as the label definitions. Define each label with examples, and give Claude an explicit way out when it is unsure:

Classify the ticket into exactly one category: Billing, Technical,
Shipping, or Account. Also rate urgency (P1–P3) and sentiment.

Definitions with examples are attached. If the ticket fits two
categories, or your confidence is low, or the customer sounds distressed,
set category to NEEDS_HUMAN and say why in one sentence.

That NEEDS_HUMAN path is the difference between a triage assistant and a liability. Low confidence routing to a person is the system working, not failing.

Response drafting

Ground every draft in the approved knowledge base and the customer's actual history — never in Claude's general knowledge of how products usually work. Jonas, Northstar's support lead, adds one rule that pays for the whole workflow: the draft must name any fact it is missing before it proposes a resolution. A draft that says "I could not confirm whether the unit is under warranty — check before sending" is honest scaffolding an agent can finish in a minute. A draft that assumes the warranty is a complaint waiting to happen. Control tone with two or three example replies from your best agent, not with adjectives.

Escalation logic

Encode observable triggers, not vibes: safety concern, legal threat, security issue, vulnerable customer, repeated contact about the same failure, refund above threshold, policy exception. Two requirements make the logic auditable — Claude states which trigger it detected and quotes the phrase that fired it, and the list itself is owned by a person and reviewed as policy, because escalation rules are policy.

Connected support workflow

An approved MCP connection can pull the relevant order, ticket history, or account record from Zendesk, Intercom-style tools, or an internal platform, so agents stop tab-hopping. Begin read-only. Add write actions — tagging, status changes — only after accuracy, permissions, and a recovery procedure are proven. "Undo" must exist before "do" is allowed.

Track lab

Build a triage-and-draft workflow for one ticket category from your own queue. Test three tickets:

1. A routine request — is the draft sendable with light edits?
2. An incomplete request — does the draft surface the missing fact?
3. A case that must escalate — does it stop and name the trigger?

Score classification accuracy, evidence use, tone, completeness, and escalation behavior against the answer key from your workflow packet.

Guardrails

• Minimize customer and personal data in context — the draft rarely needs the full history.
• Never promise an outcome that policy does not support; the KB is the boundary.
• Refunds, account changes, legal statements, and case closure require human approval unless policy explicitly automates them.
• Preserve the original ticket and the human decision in the audit trail.

Definition of done

Agents accept the drafts for routine cases, the workflow reliably stops on the cases that need a person, and everyone can see which of those two things happened and why.