We cut our agent costs 73% by removing intelligence, not adding efficiency

We launched our customer service agent with a simple pricing model: $47 per agent per month. Clean. Predictable. Wrong.

Three months later, we're looking at a $3,200 monthly bill for what should have been $188. The agent works beautifully — customers love it, response times dropped 80%, satisfaction scores went up. But the unit economics are completely broken.

Here's what nobody tells you about agent pricing: the work expands to fill the intelligence available.

Our agent started simple — answer FAQ questions, escalate complex issues. But because it could do more, it started doing more. Research customer history. Cross-reference order details. Generate detailed summaries. Write follow-up emails. Each conversation that should have cost $0.15 in API calls was burning $2.40.

The trap: per-agent pricing makes you think linearly, but agent costs compound exponentially with capability.

We fixed it with a cost ceiling architecture:

# Agent cost controls
max_tokens_per_conversation: 8000
max_api_calls_per_session: 12
escalation_threshold: $0.25

# Route by cost, not capability
simple_queries: "gpt-4o-mini"  # $0.02/conversation
complex_queries: "claude-sonnet"  # $0.18/conversation
research_tasks: "escalate"  # $0/conversation

The breakthrough was realizing that most customer service work is actually simple work. Password resets, order status checks, basic troubleshooting — these don't need a $50/hour AI brain. They need a $2/hour AI brain with good routing.

Now we measure cost per resolution, not cost per token:

• Tier 1 (80% of queries): GPT-4o-mini, $0.03 average cost
• Tier 2 (15% of queries): Claude Sonnet, $0.22 average cost
• Tier 3 (5% of queries): Human escalation, $0 AI cost

The agent now asks itself: "Is this worth $0.25 to solve perfectly, or should I solve it adequately for $0.03?" Most of the time, adequate wins.

We also added conversation value gating. If a customer's account value is under $100, the agent gets a $0.15 budget. Enterprise customers get $2.00. The agent quality scales with customer value, not customer need.

Result: Same customer satisfaction scores, 73% lower AI costs, and we actually understand our unit economics now.

The hard lesson: Flat agent pricing is a lie. Your real costs are tokens × complexity × conversation length × model choice × retry loops. If you're not measuring cost-per-outcome, you're flying blind into a budget explosion.

Most agent builders discover this the expensive way. We did too. Now our agents have spending discipline built in from day one.