AI Agent Spend ROI
'Agents pay for themselves' is a claim, not a measurement. Here's the honest framework: cost per completed task, avoided incidents, and what to stop counting.
The honest metric
Cost per completed task = total agent spend ÷ tasks actually completed.
Include everything: inference, tools, data, and failed attempts (retries count).
The second metric
Avoided incidents: the runaways that didn't happen. Compare your pre-firewall incident rate to post.
The incident database gives reference points for what runaways cost.
What to stop counting
Token counts and API calls without outcomes — volume isn't value.
Claimed productivity without a task denominator.
At a glance
| Metric | Formula | Tells you |
|---|---|---|
| Cost per task | Total spend ÷ tasks done | Efficiency |
| Avoided incidents | Pre vs post firewall | Risk reduction |
| Spend per agent | Ceiling vs actual | Budget discipline |
FAQ
What's a good cost per task?
It depends on your task value — the honest benchmark is your own margin, not an industry number.
How do I track failed attempts?
The audit log records every decision, including BLOCKED attempts that would have been failures.
Related
Stop the next $12,400 night.
One API call (or MCP tool) in front of every agent transaction — APPROVED, BLOCKED, or FLAGGED, deterministic, ~5 ms, fully logged.
See plans — from $99/mo Try a live check