Cost Per Outcome Benchmarks: Customer Service

Customer support is the proving ground for AI agents. Hundreds of companies have deployed AI customer support in the last 24 months, creating a real data set of production costs. This article compiles the actual cost-per-outcome benchmarks for AI customer service agents—from Klarna's famous $0.19 per resolved ticket to production costs at Intercom Fin, Sierra, and others—and contrasts them with manual baseline costs. Use these numbers to calibrate your own forecast and understand where realistic AI economics sit.

Baseline Cost Per Ticket (Manual/Human)

Before you can measure AI ROI, you need to know what you're starting from.

For a typical customer support team in the US, baseline cost per resolved ticket is between $18 and $35, depending on complexity and industry.

Cost breakdown for a typical CSR:

• Annual salary: $32,000 (entry-level support rep)
• Benefits (health, 401k, etc.): $8,000
• Overhead (facility, tools, training): $10,000
• Total fully loaded: $50,000 per CSR per year

Productivity calculation:

• Support hours per year: 1,800 (assuming 40-hour weeks, minus PTO and training)
• Average handle time (AHT) per ticket: 6 minutes (industry standard)
• Productive tickets per person per year: 1,800 hours × 60 minutes / 6 minutes = 18,000 tickets per year
• Cost per ticket: $50,000 / 18,000 = $2.78 per ticket (hourly labor only)

But this is the optimistic case. In reality:

• First-call resolution rates are often <70%, so 30% of customers call back. That's 1.3-1.5x the cost per issue resolved.
• Overhead (facility, tools, training, management) adds another 20-30%.
• Turnover is high in support (25-40% annually), and replacement cost is substantial.

More realistic baseline: $4.50-$7.50 per first-time call, $9.00-$14.00 per issue fully resolved.

For mid-market healthcare or financial services support (more complex work, higher salaries):

• Annual fully loaded cost per CSR: $65,000-$85,000
• AHT: 8-10 minutes (more complex)
• Productive tickets per year: 9,000-12,000
• Realistic cost per resolved issue: $6.00-$10.00 per ticket (conservative model)

Let's use $8.00 per resolved ticket as a representative baseline for a mid-market support organization.

Klarna: $0.19 Per Resolved Ticket

Klarna's AI customer support agent (built in-house) has become the most-cited benchmark in the industry. Klarna reported handling 2.3 million conversations in Q4 2024, with an average cost of $0.19 per resolved conversation.

This is real production data from a high-volume, public company. But it's important to understand what's driving the unit economics.

Why Klarna's number is so low:

1. Massive volume. 2.3 million conversations per month means infrastructure cost (gateway, monitoring, caching) is amortized over an enormous base. A 1,000-conversation-per-month customer service team will have different per-unit economics.

2. Financial services domain specificity. Klarna handles payment disputes, refunds, and balance inquiries—work that has clear, deterministic resolution paths. There's less ambiguity than in general customer support.

3. High escalation acceptance. Klarna accepts that 15-20% of conversations escalate to humans (Klarna hasn't disclosed exact escalation rate). The $0.19 number includes human review cost amortized, not pure API cost.

4. Prompt caching leverage. By using the same customer context across multiple conversations, Klarna achieves very high cache hit rates (likely 50%+), which reduces API cost by 80-90% on cached calls.

5. Multiple attempts and iteration. Klarna's agent likely ran for 18+ months before reaching $0.19. Early iterations probably cost $2-5 per conversation. The $0.19 represents optimization at scale.

Klarna's economics at your scale:

If Klarna's $0.19 includes all cost (API, human review, infrastructure), then for a 10,000-conversation-per-month operation:

• Total monthly cost: $1,900
• Implied cost before infrastructure amortization: $0.10 (API + retries + review)
• Implied infrastructure overhead: $900/month = $0.09 per conversation

Your infrastructure cost might be $2,000-3,000/month, not $900. So your realistic cost per conversation = $0.10 (API + review) + $2,500 / 10,000 = $0.35 per conversation.

Takeaway: Klarna's $0.19 is real but not portable to a smaller operation without similar volume and domain specificity. Expect 2-3x that cost at mid-market scale.

Intercom Fin: ~$0.99 Per Resolution

Intercom (a customer communication platform) launched Intercom Fin, an AI agent for financial services support. Published cost is approximately $0.99 per customer issue resolved, including API, validation, and human review.

This is a better comparable for mid-market companies because Intercom designs for SaaS companies (smaller scale than Klarna).

Intercom Fin's cost drivers:

1. Higher escalation rate. Intercom Fin targets more complex financial issues (account disputes, subscription issues, billing questions), with an estimated 25-35% escalation to human.

2. Third-party API integration. Intercom Fin calls external APIs (Stripe for payment data, custom payment processor integrations) to verify customer information and take action. Each API call adds latency and cost.

3. Multi-turn conversations. Average conversation length is 3-4 turns (customer and agent exchange messages multiple times), not single-turn resolution.

4. Built-in human fallback. Human support is embedded in the conversation flow—the system is designed to escalate smoothly, not to minimize escalation.

Intercom Fin's cost at your scale:

If your support operation is 5,000 tickets per month (similar to a mid-market B2B SaaS company), Intercom Fin's model suggests:

• AI-resolved cost: $0.40-0.60 per ticket (API + inference)
• Human review/escalation cost: $0.35-0.50 per ticket (25-30% escalation at ~$4 per escalation)
• Total cost per ticket: $0.75-$1.10

This is 10-15% of your baseline cost-per-ticket ($8), so ROI is strong.

Sierra: ~$1.50 Per Resolution

Sierra (an agentic AI platform for customer service) has disclosed production cost data for its customers. Average cost per resolved customer interaction is approximately $1.50, including API, infrastructure, and human review.

Sierra's model is instructive because it's designed for enterprise customer support (higher complexity, more integrations, higher touch):

Sierra's cost drivers:

1. Rich tool integration. Sierra agents integrate with CRM, ticketing systems, knowledge bases, and payment systems. Each integration adds infrastructure cost.

2. Higher accuracy requirement. Enterprise customers require 85%+ accuracy before deploying to production, which means more training, more evaluation, and more human-in-the-loop feedback.

3. Variable escalation. Escalation rates vary by customer vertical (SaaS support ~15-20% escalation; healthcare/insurance ~35-45% escalation). Sierra's $1.50 average reflects a mixed portfolio.

4. Optimization for retention. Sierra is designed to reduce churn by improving customer experience, not just cost. That means higher accuracy and more natural conversation flow, which costs more.

Sierra's cost at your scale:

If you're deploying a Sierra-like agent in mid-market healthcare or insurance (higher complexity, 40%+ escalation):

• AI-resolved cost: $0.50-0.80 per ticket
• Human review/escalation cost: $1.00-1.50 per ticket (40-45% escalation)
• Total cost per ticket: $1.50-$2.30

This is still 20-30% of your baseline cost ($8-10 in healthcare), so ROI remains positive.

Decagon and Devin: Task-Based Pricing

Some AI platforms (Decagon, a focus on back-office; Devin, an AI for coding tasks) use per-task pricing instead of per-token. This is useful for cost benchmarking because it's outcome-based.

Decagon charges approximately $1.50-2.50 per completed back-office task (invoice processing, data entry, reconciliation). Devin charges approximately $2.25 per task for coding work (bug fixes, refactoring, documentation).

These are higher than customer support because back-office and coding tasks are longer, more complex, and require more reasoning. But the benchmark is useful: if a task-specific AI service charges $2-3 per outcome, that's your reference point for true all-in cost, including the vendor's margin.

Extracting the true cost: If Devin charges $2.25 per task and takes 40% margin, the true cost is roughly $1.35 per task (API + human review + Devin's infrastructure cost, amortized).

Cost Per Outcome By Service Type

Customer support breaks down into a few service types, each with different cost economics:

| Service Type | Complexity | Typical AHT | AI Cost | Human Review | Total Cost | Baseline | ROI | |---|---|---|---|---|---|---|---| | Password reset / simple account issues | Low | 2 min | $0.05 | $0.10 (5% escalation) | $0.15 | $2.00 | 92% savings | | General knowledge questions | Medium | 5 min | $0.12 | $0.40 (15% escalation) | $0.52 | $4.50 | 88% savings | | Billing and subscription issues | High | 8 min | $0.20 | $1.00 (35% escalation) | $1.20 | $7.50 | 84% savings | | Complaints and escalations | Very high | 12 min | $0.30 | $3.00 (70% escalation) | $3.30 | $10.00 | 67% savings |

Note: All numbers above are illustrative based on industry patterns. Your actual costs depend on your baseline productivity, fully-loaded labor cost, and agent accuracy in your specific domain.

Building Your Own Benchmark

To calibrate your forecast, answer these five questions:

1. What's your current cost per resolved ticket? Work backwards from payroll / tickets handled. Don't forget overhead and turnover cost.

2. What's your target use case? Simple (password resets, FAQs) or complex (billing disputes, complaints)? Simple cases run $0.10-0.40 total cost. Complex cases run $1.50-3.00.

3. What's your realistic accuracy rate? Most first-generation agents achieve 70-85% first-touch resolution. Plan for 15-30% escalation to humans.

4. What's your volume? Klarna's per-unit cost is low because of massive volume. Your cost per unit will be higher at lower volume—infrastructure is fixed.

5. What integrations do you need? Pure text-based support (simpler, cheaper) vs. integrated with CRM/billing (more complex, more expensive). Factor in 5-15% additional cost for integrations.

Your formula:

• API cost: Research the model you'll use (Claude, GPT-4, Llama). Estimate tokens per ticket. Multiply by per-token cost. Typical: $0.05-0.20 per ticket.
• Human review cost: Escalation rate × time per escalation × loaded labor rate. Typical: $0.20-0.80 per ticket.
• Infrastructure cost: Variable by volume, but typically $1,000-3,000/month. Divide by monthly ticket volume. Typical: $0.20-0.50 per ticket.
• Total: $0.45-$1.50 per ticket for a realistic mid-market deployment.

If your baseline is $8-10 per ticket and your AI cost is $0.50-1.50, you're looking at 80-95% cost reduction—board-grade ROI.

The benchmarks in this article are the external sanity check. Build your own cost model, then compare against these public data points. If your forecast is >2x higher than comparable companies, dig into why.

For the full ROI calculation framework, see How to Actually Measure AI ROI (With Numbers). If you want to run the numbers interactively, use the AI ROI Calculator.