With a single-step feature, every request costs roughly the same. With an agent, an easy task takes four steps and a hard one takes nineteen, and because each step carries everything seen so far, later steps are more expensive than early ones.

The Cost Distribution

1,000 real exception-investigation tasks:

Steps Share of tasks Cost per task

3-5 58% 0.04

6-9 27% 0.11

10-14 11% 0.26

15-20 3% 0.49

hit cap 1% 0.80

Median cost: 0.06

Mean cost: 0.11 <- nearly double the median

Budget on the median and you are roughly 45% short.

The 4% of tasks above 15 steps consume about a fifth

of the total spend.

Single-Step AI	Agentic
0.04, median cost per task vs mean, same agent	0.11, same task, more autonomy

Always Ask for the Distribution

When engineering tells you what a task costs, ask for the median, the mean and the 95th percentile. If they've only measured the average, the tail is invisible and your budget will be wrong in one direction. The tail is also where your timeouts and your worst user experiences live, so it's the part worth knowing about.

Agentic AI and MCP for Product Managers

Module 1What 'Agentic' Actually Means

Module 2How an Agent Actually Works

Module 3MCP: What It Is and Why You Should Care

Module 4Designing the Tool Layer

Module 5Permissions and the Human in the Loop

Module 6Evaluating Agents

Module 7Cost, Latency and Unpredictability

Module 8Designing What the User Sees

Module 9Risk and Blast Radius

Module 10When an Agent Is the Wrong Answer

Module 11Final Project: Design One Real Agent End to End

Module 12Appendix: The Prompt Library

Why Cost Per Task Has a Long Tail