Halden
Log inGet started
All posts

Guide

What auditors ask about AI agents that spend money

We sat in on a dozen SOC 2 and SOX walkthroughs where agents could move money. These are the questions that came up every time.

PRPriya RamanNovember 4, 2026 · 2 min read

When an AI agent can make a payment, your auditor stops asking whether you have controls and starts asking which person each control belongs to. Over the last six months we joined a dozen audit walkthroughs with customers. The same five questions came up in almost every one.

1. Who authorised this agent to spend?

Auditors want a named owner for every agent and a record of when its spending authority was granted, changed or revoked. "The platform team set it up" is not an answer. A good answer looks like a change log: who, what limit, when, approved by whom.

2. How do you know the request came from that agent?

Shared API keys are the most common finding we see. If three agents use the same credential, you can't prove which one made a payment. Each agent should sign its own requests with its own key, and the signature should be stored alongside the payment.

3. What stops it from overspending?

Expect to walk through your limits and show them working. The strongest evidence is a held or blocked payment, with the rule that caught it:

  • The policy version that was in force at the time
  • The rule that triggered (amount, merchant, category)
  • What happened next — who approved or declined, and when

4. Can the same person own an agent and approve its payments?

This is segregation of duties, and it trips up small teams most. The person who configures an agent should not be the person who approves its exceptions. Enforce it in the system, not in a policy document.

5. Can you prove the record hasn't been edited?

Screenshots and CSV exports are easy to alter. Auditors increasingly ask how you would detect a change to history. A hash-linked ledger — where each entry includes the fingerprint of the one before — answers that question in one sentence.

The short version

If you can answer these five questions with exports rather than explanations, your walkthrough will be short. If you can't yet, start with question two: give every agent its own identity. Everything else gets easier from there.

Notes on trusting machines with money

One email a month: regulation, playbooks and what we shipped. No spam, unsubscribe any time.