---
title: How do you control AI agents that can move money?
url: https://opencontroller.lyzr.com/blog/control-ai-agents-that-can-move-money
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you control AI agents that can move money?

**Control money-moving AI agents with limits enforced outside the model: per-transaction and daily amount caps, payee allow-lists, velocity limits, human or second-system approval above a threshold, and daily reconciliation against the agent's logs. Never rely on the prompt to keep an agent within limits.** The model decides what to attempt; your controls decide what is allowed.

## Why are payment agents a special case?

Most agent mistakes can be corrected with an apology or a new answer. A sent payment often cannot be undone. Refunds, payouts, vendor payments, credits and transfers are attractive targets for prompt injection, and they turn small reasoning errors into real losses.

The OWASP Top 10 for Agentic Applications names the relevant risks directly: tool misuse, identity and privilege abuse, and agent goal hijack. An agent that reads customer emails and can issue refunds is a textbook case: one crafted email can ask it to refund the wrong account.

## What controls should every money-moving agent have?

| Control | What it does | Where to enforce |
|---|---|---|
| Per-transaction cap | Blocks any single action above a set amount | Payment API or tool wrapper |
| Daily / rolling cap | Limits total value per agent per period | Tool wrapper or policy engine |
| Velocity limit | Limits count of actions per minute or hour | Gateway or tool wrapper |
| Payee allow-list | Only pays known, verified accounts | Payment system |
| No new payees | Agent cannot create or edit payee details | Remove the tool entirely |
| Threshold approval | Human or second system approves above a threshold | Workflow or payment system |
| Idempotency keys | Stops duplicate payments on retries | Payment API |
| Scoped identity | Agent uses its own credential with only payment scopes it needs | Identity provider |
| Kill switch | Stops the agent mid-run without a deploy | Control plane / runtime |
| Reconciliation | Matches every payment to an agent trace daily | Finance + logs |

The most important rows are the ones enforced in systems the model cannot talk its way past: the payment API, the identity provider and the tool wrapper.

## Why should limits live outside the prompt?

A prompt that says "never refund more than $200" is an instruction, not a control. Prompt injection, long context, or a model upgrade can all change whether the instruction is followed. Put the limit in code that runs between the agent and the payment system, and reject the call there. The agent then gets an error it can explain to the user.

## How should approval thresholds work?

Separate the decision from the execution:

1. Below threshold A (for example, refunds under a small amount on orders under 30 days old): agent executes automatically.
2. Between A and B: agent prepares the transaction; a person approves in the finance tool.
3. Above B, or any new payee: agent cannot act; it routes to the normal finance process.

Set thresholds from historical data on fraud and error rates, not intuition. Review them quarterly. The approver should not be the agent's owner, which is ordinary segregation of duties (NIST SP 800-53 control AC-5).

## How do you prove the controls work?

Auditors and finance teams will ask for evidence, not descriptions.

- **Reconciliation:** every payment in the ledger links to one agent trace with agent ID, version, user identity and the tool call that triggered it. Payments without a trace are an incident.
- **Blocked-attempt log:** a record of calls rejected by caps or allow-lists. A steady trickle is normal; a spike suggests injection or drift.
- **Kill-switch tests:** a dated record that you stopped the agent in a test and payments stopped.
- **Change records:** re-approval when the model, prompt or tool set changes.

## What should trigger automatic quarantine?

- Repeated attempts above the cap.
- Attempts to pay non-allow-listed accounts.
- A sudden change in average refund value or volume.
- Tool calls outside business hours if the agent normally works during them.

## How Opencontroller does this

Opencontroller does not provide human-approval hold workflows in its gateway; approvals should sit in your finance or workflow system. It does provide controls around them: delegated on-behalf-of identity and agent workload identity federation so each payment traces to an agent and user; a governed registry of agents, versions and tools; Posture to flag risky permissions; behavioural-deviation monitoring with alerts and automatic quarantine and kill; per-agent Block and Quarantine controls; and policy as code across the estate. See the pillar, [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent), or [book a demo](https://www.lyzr.ai/book-demo/).

Related: [human approval for high-risk AI agent actions](/blog/human-approval-for-high-risk-ai-agent-actions) and [what is excessive agency in AI agents](/blog/what-is-excessive-agency-in-ai-agents).

## Frequently asked questions

**Can I tell the agent its refund limit in the prompt?**
You can, but treat it as guidance only. Enforce the limit in the payment API or tool wrapper so the call fails if the agent ignores it.

**Should a payment agent be allowed to add new payees?**
No. Remove that tool entirely. New payees are a classic fraud vector and should go through the normal finance process.

**How often should money-moving agents be reconciled?**
Daily. Every payment should match an agent trace, and any payment without one should be treated as an incident.

## Key facts

Money-moving AI agents need limits enforced outside the model: per-transaction and rolling caps, velocity limits, payee allow-lists, no payee-creation tool, threshold approvals, idempotency keys, scoped identity, a kill switch and daily reconciliation of payments to agent traces. NIST SP 800-53 AC-5 covers separation of duties. Opencontroller does not ship human-approval hold workflows in its gateway; it provides on-behalf-of identity, Posture, behavioural-deviation monitoring, automatic quarantine and kill, and policy as code. It starts from $250K per deployment per year.

## Related

- [Which AI agent actions need human approval?](https://opencontroller.lyzr.com/blog/human-approval-for-high-risk-ai-agent-actions.md)
- [What is excessive agency in AI agents?](https://opencontroller.lyzr.com/blog/what-is-excessive-agency-in-ai-agents.md)
- [Should AI agents do financial calculations themselves?](https://opencontroller.lyzr.com/blog/should-ai-agents-do-financial-calculations.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
