What Is an Human-in-the-Loop Approval? — sipi.bot Glossary
A spending control pattern where transactions that match certain criteria (over a threshold, to a new merchant, outside business hours) are paused and sent to a human for review before execution.
By the sipi.bot team · Published 2026-07-19
Definition
FLAGGED decisions create a human-in-the-loop queue. The agent says 'I want to spend $900 on this cloud GPU.' sipi.bot says FLAGGED — the transaction is paused, a pending item appears in the authenticated dashboard, and your integration should execute only after a human approves. The agent never knows it was paused.
Why It Matters
Without proper human-in-the-loop approval controls, autonomous agents can accumulate significant unexpected costs. sipi.bot automates human-in-the-loop approval enforcement so you deploy agents with confidence.
How an Human works in practice
Understanding the definition of an Human is the first step; knowing how it behaves in a production agent environment is what actually protects your budget. In practice, an Human manifests differently depending on your agent architecture, the payment methods your agent has access to, and whether the control is enforced before or after the transaction executes.
Consider a real example: a research agent with access to a $500/month LLM API budget. Without an Human, a single retry loop on a complex query can burn through 40% of the monthly budget in 20 minutes — 237 API calls at $0.84 each = $199.08. With an Human enforced as a velocity cap (10 calls/minute), the agent is blocked at call 11, the total spend is $9.24, and the audit log immediately surfaces the abnormal pattern. The team is alerted within seconds and investigates the retry bug before it recurs.
Common configuration mistakes
- Setting an Human too high because you are worried about interrupting legitimate agent work. Start conservative and raise based on observed data. A blocked transaction is a signal; an unblocked overspend is a cost.
- Applying an Human globally instead of per-agent. Different agents have different spend profiles. A research agent that makes 200 LLM calls/day is not the same as a billing agent that makes 5. Use per-agent policies.
- Forgetting to test the block path. Configure an Human, then deliberately trigger it to confirm your agent handles the BLOCKED response gracefully — no crash, no silent retry, and a clear explanation to the user.
How sipi.bot enforces an Human
sipi.bot evaluates an Human on every transaction with a deterministic rules check. The agent never sees the payment method directly; it receives a structured JSON decision (APPROVED, BLOCKED, or FLAGGED) and acts accordingly. Every decision is logged with agent ID, merchant, amount, timestamp, and the rule and reason that produced it — so you can always reconstruct why a transaction was allowed or denied. Hosted Team is $99 per month; the same rule engine is MIT-licensed for self-hosting.