Guide
How much does AI support analysis cost?
Model cost is only one part of support automation economics, but it should be measured on every completed analysis. Prompt size, generated output, model selection, retries, and human review all affect the real cost of a workflow.
Updated July 23, 2026 · 6 min read
Direct answer
AI support analysis cost equals input tokens multiplied by the model's input rate plus output tokens multiplied by its output rate, with additional allowance for retries and review time. In a ReplyPilot test on 60 sample messages, Qwen3 used 5,837 total tokens and had an estimated model cost of about $0.00070.
Use the token formula
For token-priced models, calculate input_tokens / 1,000,000 × input_price plus output_tokens / 1,000,000 × output_price. Keep cached-input pricing separate when the provider exposes it.
Document the benchmark boundary
The ReplyPilot figure is an internal technical benchmark using a synthetic 60-message customer-support CSV, a two-stage clustering and report prompt, and Cloudflare Workers AI Qwen3. It measures model inference only, not reviewer labor, storage, database, or network charges.
- 60 normalized support messages
- Two AI stages
- 7 generated themes
- 13 generated action assets
- Approximately 16.7 seconds end-to-end
- Approximately $0.00070 model inference estimate
Optimize for completed useful reports
The cheapest request is not necessarily the cheapest workflow. Invalid JSON, poor themes, missed risks, or unusable assets create retries and review work. Compare cost per accepted report and cost per adopted asset rather than token price alone.
Frequently asked questions
Is the benchmark a customer result?
No. It is a technical test using synthetic support messages and should not be presented as evidence of customer savings or willingness to pay.
Does the estimate include Cloudflare infrastructure?
No. It covers model-token pricing only. Workers requests, R2, database, authentication, and human review are separate.
Why use a two-stage prompt?
Separating topic clustering from report composition makes outputs easier to inspect and allows each stage's usage and failure to be measured.