Cost Per Ticket: How AI Customer Service Is Actually Measured

The metric that matters for AI customer service is cost per ticket: the fully loaded cost of an AI agent or AI-assisted team to resolve one customer support request. This includes not just the API bill but embeddings, human review, integrations, and observability—the entire cost iceberg beneath the waterline.

Klarna's public benchmarks set the gold standard: their AI customer service agent resolves tickets at $0.19 per interaction. Sierra, a newer entrant in the space, reports approximately $1.50 per resolution. The range tells you something important: cost per ticket depends heavily on whether you're doing full auto-resolution or human-in-the-loop, which models you're using, how much domain knowledge you're encoding, and how much human review you've built in. There's no universal number, but there are benchmarks.

The Full Cost Stack for a Customer Support Ticket

Here's what cost per ticket actually includes:

Model cost. A typical customer support query might involve 1-3 API calls to an LLM. If you're using Claude 3.5 Sonnet (as of early 2025), each call might cost $0.01-$0.03 depending on the length and the model. If you're using GPT-4 Turbo, it might be $0.04-$0.08. If you're using a smaller, cheaper model like Claude 3 Haiku, it might be $0.001-$0.005. This is the visible part.

Embeddings and vector database. Most customer support AI systems use retrieval-augmented generation (RAG) to ground the agent in your company's knowledge base. Every ticket might trigger 2-5 embedding lookups (searching your internal docs, previous ticket history, product specifications). Embedding APIs cost $0.00001-$0.0001 per thousand tokens, and vector database storage is typically billed per million vectors. For a large support team handling 10,000 tickets per month, embedding and vector costs might add $0.05-$0.12 per ticket.

Failed retries and rate limiting. Not every API call succeeds on the first try. Transient failures, rate limits, and timeouts cause retries. Depending on your infrastructure, this might add 5-15% overhead on API costs.

Human review time. If your process requires a human to review the AI's response before it goes out (common in financial services, healthcare, or when handling high-value customers), you need to allocate labor cost. A compliance officer in healthcare spending 5 minutes reviewing a ticket at a $60K annual salary (fully burdened) costs about $0.50 per ticket reviewed. This is often the largest single cost component. Klarna's $0.19 per ticket benchmark suggests they're doing mostly auto-resolution; Sierra's $1.50 suggests more human review.

Observability and logging. Every API call, retry, and decision path needs to be logged so that you can audit the agent's behavior and calculate costs. Logging platforms (Langfuse, custom CloudWatch implementations) might add $0.02-$0.05 per ticket.

Third-party integrations. If your support agent integrates with Stripe to check customer payment status, Slack to escalate complex tickets, or Twilio to send SMS notifications, each integration has a per-call cost. For a typical support ticket, this might add $0.02-$0.08.

Infrastructure and amortization. If you're running custom embedding infrastructure, prompt caching layers, or rate-limiting gateways, amortize that infrastructure cost across your ticket volume. Depending on your architecture, this might add $0.01-$0.10 per ticket.

Summing across a typical ticket: $0.03 (model) + $0.08 (embeddings) + $0.04 (retries) + $0.25 (human review) + $0.03 (observability) + $0.04 (integrations) + $0.05 (infrastructure) = $0.52 per ticket. That's in the middle of the Klarna-to-Sierra range, suggesting a reasonable baseline.

Why Your Current Measurement Is Incomplete

Most companies trying to measure customer support AI cost only track the model cost ($0.03 in the example above) and call it done. They say "our support AI costs $0.03 per ticket, wow, that's cheap!" But that ignores 85% of the real cost stack. The human review step alone—$0.25—is 8x the model cost.

This is why the AI Cost Iceberg matters. The visible API cost is only the tip. If you're benchmarking your customer support AI against Klarna at $0.19 per ticket and only measuring model cost at $0.03, you're missing the point. The question isn't "how cheap is the model API?" It's "what's the all-in cost of a resolved ticket to my business?"

This matters for vendor evaluation. If you're comparing CloudZero's support AI (hypothetically $0.15 per ticket) against Intercom Fin ($0.99 per resolution), and you're only looking at model costs, you'll make the wrong decision. You need the full stack.

Measuring the Full Stack

To get accurate cost per ticket, instrument your support agent to emit a structured cost trace for every ticket it handles. Include:

• Ticket ID
• Model selection and API cost
• Number of embedding lookups and cost
• Human review time (if any) and cost
• Third-party integration calls and costs
• Success/failure on first attempt, retries, latency
• Final outcome (auto-resolved, escalated to human, bounced to another system)

Aggregate these traces weekly or monthly to calculate average cost per ticket, broken down by cost category, by ticket type, by customer segment, and by outcome. Once you have that, you can answer the real questions:

"Is customer support AI actually cheaper than our current process?" Compare $0.52 per AI ticket against the fully loaded cost of a human support agent ($150-$300 per ticket, including overhead). The answer is almost always yes.

"Which customer segments are we losing money on?" If support tickets for your highest-tier customers cost $1.20 per ticket (due to human review overhead) and you're pricing support at a flat $0.50 per ticket in your SLA, you're losing money on them.

"Where should we optimize?" If human review is 50% of cost, that's your target for automation or process improvement.

Cost Per Ticket as a Vendor Evaluation Tool

When you're evaluating a new customer support AI vendor, ask them for their cost per ticket benchmark and insist on seeing the full cost stack. If they quote only model cost, ask them to include human review time, embeddings, integrations, and observability. If they won't, assume they're either cherry-picking numbers or they don't actually measure it.

The strongest vendors (Klarna, Intercom, Sierra) publish full cost numbers because they're competitive on the full stack, not just on model tokens.

The Model Selection Lever

One of the highest-impact levers in cost per ticket is model selection. Using GPT-4 costs significantly more per token than Claude 3 Haiku, which costs more than smaller, open-source models. But cheaper models sometimes produce lower-quality responses that require more human review, making the total cost per ticket higher.

The question you should be asking: "For each ticket type or customer segment, which model minimizes cost per ticket while maintaining quality standards?"

Some support organizations implement a tiered model strategy: GPT-4 or Claude 3.5 Sonnet for complex tickets that require nuanced understanding, Claude 3 Haiku for routine tickets and first-contact resolution, and open-source models for classification and routing tasks that don't require deep reasoning.

This requires measuring cost per ticket by model, by ticket type, and by outcome quality. Once you have that data, the ROI of using different models becomes obvious.

Real-World Math: Comparing Two Support AIs

Support AI A:

• Uses Claude 3.5 Sonnet exclusively
• Cost per ticket: $0.34
• Auto-resolution rate: 62%
• Customer satisfaction: 87%
• Requires 15 minutes of human review per complex ticket (38% of volume)

Support AI B:

• Uses Claude 3 Haiku for routine tickets, Claude 3.5 Sonnet for complex tickets
• Cost per ticket: $0.27
• Auto-resolution rate: 65%
• Customer satisfaction: 86%
• Requires 12 minutes of human review per complex ticket (35% of volume)

AI B is 20% cheaper per ticket ($0.27 vs. $0.34), slightly better at auto-resolution, with nearly identical customer satisfaction. For 100,000 tickets per year, that's $7,000 in annual savings. This is the kind of decision that cost per ticket reveals.

For more context on how to evaluate AI vendors and calculate true AI ROI, see the pillar article on AI cost attribution.