Worked example: support triage agent
The default scope starts with 24 base hours, then adds 8 workflow steps x 4 hours, 4 integrations x 8 hours, 3 AI tools x 6 hours, 1 approval checkpoint x 6 hours, 30 eval cases x 0.4 hours, and 6 security review hours. Raw build effort is 24 + 32 + 32 + 18 + 6 + 12 + 6 = 130 hours.
QA and project overhead add 25% + 15%, so adjusted build effort is 130 x 1.40 = 182 hours. At $125/hour, build cost is 182 x $125 = $22,750.
Monthly volume is 300 requests/day x 30 days = 9,000 requests. With 8% retry overhead, input tokens are 9,000 x 3,000 x 1.08 = 29.16M and output tokens are 9,000 x 900 x 1.08 = 8.748M.
At $0.75 per 1M input tokens and $4.50 per 1M output tokens, API cost is 29.16 x $0.75 + 8.748 x $4.50 = $61.24/month. Add $99 platform, $40 hosting, and 6 x $125 = $750 maintenance for a monthly run cost of $950.24. First month is $22,750 + $950.24 = $23,700.24.
How the agent build estimate works
The build estimate starts with a base project setup, then adds effort for workflow steps, integrations, AI tools, approval gates, eval cases, and security review. QA and project overhead are applied after the raw build hours because testing, review, and coordination usually scale with the whole project.
How monthly AI agent cost is estimated
Monthly run cost combines token cost, platform cost, hosting, and maintenance. Token cost uses requests per month, input and output tokens per request, retry overhead, and editable provider prices per million tokens.
What this estimate excludes
The estimate excludes taxes, sales contracts, legal review, data cleanup, incident response, custom compliance work, procurement time, and staff time spent reviewing uncertain outputs. Add those lines separately when an agent affects customers, money, health, legal, or regulated data.
What to do next
Before building, write the target workflow, define pass and fail cases, run a small eval set, check API rate limits, estimate automation platform units, and make a human-review policy for risky outputs.
