What an AI support conversation costs, and what decides whether it pays

Support is the use case where AI economics look easiest and get argued about most, because the model cost is genuinely small and everything around it isn't. Anthropic's own pricing page works the example: roughly 3,700 tokens for an average conversation, priced on Haiku 4.5, giving about $37.00 per 10,000 tickets. That's a third of a cent a conversation against a human-handled ticket costing orders of magnitude more, and it makes the naive business case trivially positive. The naive case is also the one that fails review. Three things move it. Conversations that escalate cost you the model spend AND the human, so they're strictly worse than not trying. Retrieval, reranking and guardrail calls multiply the per-conversation token count well past the benchmark. And a benchmark built on a 3,700-token average says nothing about your own distribution, where the long tail of complex conversations is exactly where escalation happens.

• Your deflection business case uses a vendor benchmark rather than your own conversations.
• Nobody separates conversations that resolved from conversations that escalated.
• Retrieval and guardrail calls aren't counted in cost per ticket.
• You know your average conversation cost and not the top decile.
• Deflection rate is reported and cost per deflection never is.

1. Work out your real average tokens per conversation and compare it to 3,700.
2. Split last month's conversations into resolved and escalated, and price each group.
3. Divide total support model spend by conversations RESOLVED, not attempted.
4. Count how many calls a single conversation makes, including retrieval and safety checks.

Illustrative example

Anthropic's pricing page, read 2026-08-05, publishes a worked example for support: an average of about 3,700 tokens per conversation on Claude Haiku 4.5 at $1.00 and $5.00 per million, totalling about $37.00 per 10,000 tickets. Those figures are Anthropic's. The escalation and retrieval additions below are modelled, and they're the part a business case has to supply for itself.

The published number is real, and it's a floor rather than an estimate. It prices one answering call on an average exchange, so every real deployment adds to it: more calls per conversation, and conversations that end with a person anyway. Both multipliers come from your own traffic, which is why the benchmark settles the model question and leaves the business case open.

A benchmark is somebody else's average. Yours differs on conversation length, on how many retrieval and safety calls each one makes, and above all on what fraction escalate, and none of those appear on a pricing page. Culpa prices every call and groups them by conversation, so cost per conversation is a sum rather than an estimate, and it carries whatever identifiers you attach, including whether the conversation resolved. That turns the two questions that decide the business case into queries: what a resolved conversation costs against an escalated one, and which topics drive the expensive tail. Set it against what those customers pay and support stops being a cost centre with a deflection rate and becomes a margin line.

• Fine-Tuning Cost vs Base Model: the 6:1 Crossover
• Free Tier Abuse: What One Automated Signup Really Costs
• Free Trial LLM Cost: Compute Is Customer Acquisition Cost
• Granular LLM Cost Attribution: Every Tool Has a Floor