Answers to the most common questions about controlling what your autonomous AI agent spends — before the money moves.
Yes — if it has access to a payment method and no pre-spend guardrail. Here is how it happens and how to stop it.
The panic-test framework: pick a number you can lose without worry. With benchmarks from 312 production teams.
Three rules: velocity limit (kills retry loops), per-transaction cap (catches big spends), merchant allowlist (blocks unknowns).
Pre-spend vs reactive: the single most important distinction in agent spend control.
With a reactive cap, damage is done. With a pre-spend firewall, it never crosses the threshold.
5-layer model: daily total, per-transaction cap, velocity limit, category limits, approval threshold.
Per-unit pricing from LLM providers. Input $0.15–$15/M. Output is 3–5× input.
Observability tools vs pre-spend firewall — one tells you what happened, the other shows live decisions.
More than picking a number — layered approach with caps, velocity, and approval thresholds.
5-step playbook. Velocity limit first — it kills the #1 cause (retry loops at 44% of incidents).
A transaction pattern that deviates from expected. Caught by velocity, merchant, and time rules.
Yes — 67% of teams have been surprised by at least one unintended spend.
$0.02 to $4.50 depending on task type. Median across all agents: $0.34.