A production customer-support assistant, three Claude agents deep, and a bill nobody has examined.
Your Scenario
You have inherited a customer support assistant that is already in
production. Every ticket passes through three Claude agents in sequence: 5,000 tickets
a month, $5,000 on the bill, exactly $1.00 a ticket. Nobody has looked at
that number since the system was built, and your CFO has just worked out it is 2% of
revenue.
Your job is to bring it down without breaking it. One constraint holds
all the way through: resolution must never fall below 85%. The sections below
walk you from understanding the bill to designing a replacement.
What you will do
Your Scenario — you are here. How the application is wired,
and which agent spends what.
Understand the Baseline — why one of the three agents carries 42% of the
bill, and what swapping its model would cost you in resolution.
Optimization Leversscored — cut at least
$1,500 a month without dropping below 85% resolution.
Scaling Challengescored — decide what has to
change when volume grows ten times.
Design Your Systemscored — build one from
scratch, to a budget that tightens per ticket as volume grows.
The first two sections are reading. The last
three are scored, and the sidebar keeps the count as you go.
💡 Where the money is. Agent 2 takes 42% of the bill,
half as much again as Agent 1 — and not because it runs more often, since all three see
every ticket. The next section explains what actually makes it expensive.
Understand the Baseline
Goal: analyze where money goes. No changes yet.
Question 1 · Which agent costs most?
Answer: Agent 2 (Analysis) at $2,100/month = 42% of budget. Why? It has the longest context (ticket + category + history + similar cases) — more context means more input tokens, and Opus input is $15/1M tokens.
Question 2 · What's your cost per ticket?
Answer: $1.00/ticket (5,000 tickets / $5,000). Tokens per ticket: 2,100 total (1,700 in + 400 out).
Question 3 · What if you switch Agent 2 to Sonnet?
Current (Opus)
Agent 2 Cost$2,100
Quality95%
Resolution87%
If Sonnet ($3/1M)
Agent 2 Cost$420
Quality90%
Resolution85%
Savings
-$1,680/mo
💡 Key insight: you could save $1,680/month by switching to Sonnet, but you'd lose 2% resolution rate. That's the tradeoff.
Exercise 1 · Optimization Levers
Challenge: reduce costs 30% without dropping below 85% resolution.
Real-time results
New Cost
$5,000
Savings
-$0
Resolution
87%
Status
✅ Target Met!
Before → After
Cost$5,000 → $5,000
Per ticket$1.00 → $1.00
Resolution87% → 87%
💡 Hint: try Haiku for Triage (simple categorization) + Sonnet for Analysis & Resolution to hit 30% savings.
Exercise 2 · Scaling Challenge
Challenge: your optimized system scales to 50K tickets/month (10x growth). What changes?
At 5K/mo
$3,500
At 50K/mo
$35,000
New Problems
Quotas, SLA, Infra
Choose your scale strategy
👆 Click a strategy above to see the impact…
💡 Reality check: scaling 10x isn't just multiplying by 10 — you hit API quotas, need infrastructure changes, and operational complexity increases. Your architecture needs to adapt.
Exercise 3 · Design Your Optimal System
Challenge: design from scratch. Pick your volume, then fit the budget that comes with it — $6K at 5K tickets, $11K at 10K, $45K at 50K. Min resolution: 85%. Note the allowance per ticket falls as you scale, from $1.20 to $0.90: the business expects volume to buy efficiency.