Back to blog

Why does every AI feature end in an audit question?

Thắng Đoàn
Thắng Đoàn

You shipped an AI feature to save people time. Now auditors ask what the AI influenced, who reviewed it, and who owned the call, and your logs stay silent.

AI shapes decisions without making them

TechTarget argues that AI tools do not need the final decision to shape it. An AI can summarize interviews, score assessments, route work, and rank options. A person then signs off, but on information the AI selected and presented. Traditional audit trails only prove something happened: a login, a recommendation, an approval. They cannot show what evidence the AI had, what it did with it, or what the human actually saw.

Self-reported answers do not satisfy

The answer they get is a paragraph you wrote about yourself. When a customer's security team asks what your agent did with their data, Help Net Security notes the answer they get is a paragraph you wrote about yourself. That is a claim, not evidence. Regulators and customers increasingly want records they can verify independently. This is the gap that turns a helpful feature into a governance project.

Hash-chained logs make actions testable

Help Net Security covers halo-record, an open-source Python package that logs every agent move: tool calls, model calls, data access, approvals. Each line carries a hash, a short fingerprint of the line before it. Edit one record and every later fingerprint stops matching. Anyone can run that check with no key or vendor permission. Raw arguments stay out; values are hashed and redacted, so the log remains useful without leaking data.

  • Wrapping an agent takes one line of Python
  • Adapters pull records from OpenTelemetry, LangChain and MCP servers

Vendor features are not compliance

A vendor's features alone do not establish compliance. In financial crime compliance, TechTimes compares AI agent platforms and adds a sharp caveat: a vendor's features alone do not establish compliance. Relevant context includes FATF Recommendations, OFAC sanctions, and EU AI Act requirements. Buying a capable agent does not answer the audit question. The record of what it did does.

What this does not solve

A decision trail records actions, not judgment. It shows what the AI presented and who approved it, but not whether the ranking was fair or the human actually reviewed the output. It also covers only agents you instrument. Decisions shaped by AI inside third-party SaaS stay invisible unless that vendor exposes equivalent records.

Which of your AI features influence decisions today, and could you reconstruct what the AI showed the approver?
How much of your decision flow runs through third-party SaaS you cannot instrument?

List every AI feature you run that ranks or summarizes information before a human decides, then check whether any log captures what the AI

Sources

TechTarget: AI decision trails are the new audit trail

Help Net Security: Halo-record open-source audit trails

TechTimes: Best AI agents for financial crime compliance 2026

Share:

Recommended for you

Enjoyed this article?

Subscribe for new articles. No spam. Unsubscribe anytime.

By subscribing you agree to receive the newsletter. No spam, and you can unsubscribe anytime.