put claude on your helpdesk.

Goal: Know your real cost per AI-handled conversation before changing anything. (10 min)

1. Find every AI line on the invoice: resolutions, outcomes, sessions, and copilot seats. Each vendor bills a different unit.
2. From reporting, note how many conversations the AI touched and how many it says it resolved.
3. Divide the AI fees by the conversations it actually helped with. That's your real cost per AI-handled ticket.
4. Remove names, emails and order numbers before pasting anything into Claude.

Here are last month's AI-related charges from our helpdesk invoice, plus the AI stats from our reporting:

[paste, with customer details removed]

Work out: 1) total AI cost, 2) cost per AI-touched conversation, 3) cost per conversation the AI claims to have resolved, 4) what the bill becomes if ticket volume doubles. Show the maths.

Done when: You know your cost per AI-handled ticket, and what happens to it if volume doubles.

Goal: Give Claude your policies, past answers and tone, so its drafts sound like your best agent. (20 min)

1. Create a Project called "Support brain".
2. Add your help-centre articles, saved replies (macros), refund and returns policy, and a short tone guide as project knowledge.
3. Write project instructions: tone, what never to promise (refunds outside policy, legal statements), and when to hand over to a human.
4. Test it with 10 real past tickets, with personal details removed, and compare against what your team sent.

You draft replies for [Company] customer support.
Tone: warm, direct, short paragraphs, no jargon.
Always: use the attached policies; quote order/return windows exactly; end with one clear next step.
Never: promise refunds or compensation outside policy, give legal or medical advice, or guess. If unsure, say "needs human" and why.
Escalate to a human: legal threats, safety issues, chargebacks, VIP accounts, anything emotional.

Done when: 8 of 10 test drafts are usable with light edits.

Goal: Sort a whole inbox by topic and urgency, with risky tickets sent to a person. (15 min)

1. Pick 4–6 categories your team already uses, plus "human review".
2. Write urgency rules: outages and double charges are urgent; legal or emotional messages go to human review.
3. Paste 20 recent tickets and ask for a triage table.
4. Compare with how your team actually routed them, then tighten the rules.

Triage these tickets. Categories: Urgent, Billing, How-to, Bug, Human review.
Rules: outages or double charges = Urgent. Legal threats, safety issues, angry VIPs or anything you're unsure about = Human review.
Return a table: ticket #, category, urgency 1–3, one-line reason.

[paste 20 tickets, personal details removed]

Done when: At least 18 of 20 match how your team would have routed them.

Goal: Cut reply time on routine tickets without letting anything risky go out unchecked. (20 min)

1. For each routine ticket, paste it into the project and ask for a draft.
2. Approve, edit or escalate. Every edit tells you which instruction to improve.
3. Time 10 tickets with and without drafts.
4. Once it works, the next step is connecting Claude to the helpdesk itself, via its API or MCP server where one exists. Intercom's is mostly read-only today. That's where we come in.

Draft a reply to this ticket using our policies. Under 120 words. If it hits any escalation rule, don't draft: reply "needs human" and name the rule.

[paste ticket]

Done when: Drafts save at least 2 minutes per ticket across 10 tickets.

Goal: Get your own resolution rate, not the vendor dashboard's. (15 min)

1. Define "resolved": the customer confirmed, or didn't reply in 7 days and didn't open a new ticket on the same issue.
2. Export 50 AI-touched conversations with personal details removed.
3. Ask Claude to classify each one against your definition.
4. Compare with the vendor's number. The gap is what you're paying for but not getting.

Classify each conversation: Confirmed resolved / Left, came back within 7 days / Left, unknown / Handed to human. Use the ticket history and timestamps. Then give totals and our honest resolution rate.

[paste export]

Done when: You have your own resolution rate next to the vendor's, and you know the gap.