What AI agents actually pay for
I built 73 pay-per-call endpoints for AI agents. After ~1,557 real paid calls, five of them earn 77% of the revenue — and one, /schema-parse/extract, earns 42% of everything. Here's what agents actually pay for, and what the other 61 endpoints taught me about building for machines instead of people.
Revenue by endpoint (top 12 of 73)
| Endpoint | Type | Calls | Revenue |
|---|---|---|---|
| /schema-parse/extract | LLM | 252 | $25.20 |
| /translate/long | LLM | 91 | $7.28 |
| /translate/short | LLM | 202 | $6.06 |
| /domain-report/analyze | FLAT | 45 | $4.50 |
| /currency-exchange/convert | FLAT | 315 | $3.15 |
| /sentiment/analyze | LLM | 118 | $3.09 |
| /domain-availability/check | FLAT | 32 | $1.60 |
| /messages | LLM | 24 | $1.44 |
| /ip-report/full | FLAT | 5 | $1.00 |
| /ip-geo/locate | FLAT | 30 | $0.70 |
| /asn-lookup/analyze | FLAT | 22 | $0.66 |
| /name-gen/suggest | FLAT | 6 | $0.30 |
Remaining 61 endpoints: ~$5 combined.
Volume vs. value — top 6 by revenue
Each pair: call volume against revenue, normalized to its own maximum.
The three real earners — /schema-parse/extract, /translate/long, /translate/short — are all the same shape: text in, clean structure out. Agents don't pay for lookups they can already do themselves; they pay to offload the messy transformation they can't. That's the thesis.
/currency-exchange/convert got the most calls of anything — 315 — and earned $3.15. /schema-parse/extract got fewer (252) and earned $25.20. Volume is not value. For a solo builder deciding what to build next, that's the whole lesson: the other 61 endpoints weren't a portfolio, they were a distraction.