Price AI Agents for Peak Usage
Charge a base fee with an included allowance when customers need predictable bills, and meter extra work in units they can verify. Test the allowance against expensive runs, retries, and escalations before you promise a margin.
Choose a unit customers can audit
For a voice support agent, decide whether you are selling availability, call handling, or resolution. Use a subscription with included minutes when buyers are paying for coverage. Sell prepaid minutes when traffic arrives in bursts. Use per-resolution pricing once the buyer and your team can independently identify the same billable cases. Twilio's US pay-as-you-go Voice pricing meters calls by the minute, giving you a supplier unit to reconcile against your own usage records.
Define how you handle transferred calls, retries, and cases that reopen. Give the buyer a sample invoice and the underlying records. If both sides arrive at the same total, the meter is ready for a pilot. If they disagree, resolve the definition before taking payment.
Audit the allowance boundary at one unit below, exactly at, and one unit above the included amount. This catches inconsistent treatment of included units on invoices.
Programmable Voice Pricing in United States | TwilioRun the numbers at the allowance limit
Anthropic's Claude Sonnet 4.5 pricing table lists base input at $3 per million tokens and output at $15 per million tokens. In an illustrative text-agent month, assume 1,000 completed runs, each with 8,000 uncached input tokens and 1,000 output tokens. Model cost is $24 for input plus $15 for output. Assume another $21 in direct costs. At $150 in revenue, $60 in direct costs leaves $90, or a 60% direct-cost margin.
Now replay the same volume with twice the output. Direct costs rise to $75 and margin falls to 50%. If your target is 60% under that scenario, the fee needs to be at least $187.50 before assuming any change in volume. Verify the exact model, token categories, and provider rates against your usage statement before quoting.
Run a month with unusually high usage before setting a fixed price. Distinguish what customers are charged from what the underlying provider bills for the same workload.
Prompt caching - Claude Platform DocsFind which usage broke your estimate
Keep a cost ledger keyed to each customer job. Record every attempt, its input and output usage, any tool charges, and the final disposition. Where you use Claude Managed Agents, Anthropic says cache reads and writes appear in session usage. Record those categories separately and price them using the applicable rates. Reconcile provider statements to the ledger before treating any projected cache savings as part of your margin.
A customer can stay within their allowance while your bill rises. If the extra spend comes from retries, include those attempts in your internal cost per completed job. Compare expensive and ordinary runs side by side. If retries cluster around one failure mode, address that failure and replay the cases before changing what customers pay.
Prompt caching - Claude Platform DocsSet the boundary of a license
License a packaged agent when buyers can take responsibility for running the same workflow in their own environments. Specify who pays for inference, who handles upgrades, and how much support the fee includes. AWS Marketplace offers third-party AI agents and tools from AWS Partners. If you plan to use it, verify the terms and deployment options for your specific offer before quoting a buyer.
Use a trial deployment to test the handoff. Ask the buyer to provision the agent, run an agreed sample, and identify where the runtime charges land. If they need you to operate the service, quote that work and its usage allowance explicitly. If they can deploy it and take ownership of the operating bill, a license fee can cover the reusable agent and agreed support.
Introducing AI agents and tools in AWS Marketplace - AWS