Cyber, Risk & Internal Audit
Internal Audit & IT Audit
Where AI assurance is landing first, and the archetype with the clearest lateral entry for someone who can read a system.
Internal audit provides independent assurance that controls exist and operate. In technology that means IT general controls, change management, access, and increasingly whether anyone can explain what an AI system did. It is the third line, which means the validator must be able to reach a different conclusion from the builder and must not report to them.
| Time | What you are actually doing |
|---|---|
| 09:00 | Walkthrough with an engineering team. You are establishing how change actually reaches production, not how the policy says it does. Those differ more often than not. |
| 10:30 | Evidence. You are not taking screenshots — you are pulling the change records for a quarter and testing a sample against the policy: was there an approval, was the approver independent, did it precede deployment. |
| 13:00 | An exception. A hotfix went out without approval during an incident. The question is whether the emergency path exists, was followed, and was reviewed afterwards — not whether it happened. |
| 14:30 | Drafting. Findings are written to be read by an audit committee, so precision and proportion matter as much as being right. |
| 16:00 | Closing meeting prep. You know which finding management will push back on and you have the evidence ready. |
The AI opening
Across 362 security job descriptions, 92% mention AI and only 9% name any AI-specific risk practice. In audit that gap is even wider, and it is an unusual opportunity: an auditor who can credibly test an AI system is rare enough today that it changes your market value, and the skill takes weekends rather than years. The evidence set is in Auditing an AI system.
The decisive round
“How would you audit a system nobody can explain?”
Interviewer: “The business has deployed an AI assistant that answers customer queries. Management says it is a black box. Audit it.”
A weak answer. “I'd review the vendor's documentation and certifications, check that a risk assessment was performed, and confirm the model has been approved through governance.” Audits the paperwork rather than the system. All of that could be in place around something entirely unmonitored, and the finding would be missed.
A strong answer. “‘Black box’ describes the model's internals, not the system, and the system is ordinary software producing ordinary evidence. So I would test the system.
My single most useful request is this: take one real customer answer from last month and reconstruct it — the exact question, the documents retrieved, the model and prompt versions in force, any tool calls, the answer given, and who saw it. That one request simultaneously tests logging, versioning, retention and traceability. If they can produce it, most of my evidence set exists. If they cannot, I have my finding and I did not need to understand a neural network to reach it.
Then the control objectives I would test. Is it inventoried and owned, with a tier based on consequence. Are its limits defined — a written statement of intended and prohibited use. Is quality measured — is there an evaluation suite with a dated reference set, a recorded baseline, and regression runs on change, or is ‘it performs well’ the only evidence. Are changes controlled — can anyone edit a prompt in production without review, because a prompt is a change and is routinely not treated as one. Is access appropriate — and I would test that directly by having two users with different entitlements ask the same question and comparing the answers.
That last test takes five minutes and finds a real issue often enough that I now run it as standard.”
Whether you can audit AI without pretending to understand machine learning. The reconstruct-one-decision request and the two-user retrieval test are both concrete, cheap, and rarely thought of.
- Learn the evidence set, not the mathematics. Read Part I of the AI foundations book here; chapters 5, 6 and 7 are the ones that matter for audit.
- Audit something small yourself. Build a trivial RAG assistant, then write the audit programme for it. Doing both sides once makes every subsequent conversation concrete.
- Have a data-analytics story. Testing a full population rather than a sample of twenty-five is where IT audit is going, and being able to write the query is the differentiator.
- Practise the closing meeting, because the finding is only half the job and the conversation where management disagrees is where auditors are actually assessed.
Roughly €1.9bn of reported cash did not exist. Confirmations had been relied on without adequate independent verification over several years.
Use it to say: “It is why I distinguish evidence obtained from the auditee from evidence obtained independently, and why I would rather query a system myself than accept an extract of it.”
A tribunal held the airline liable for incorrect advice its chatbot gave a customer; the argument that the chatbot was responsible for its own statements failed.
Use it to say: “This is why I treat model output as a company statement in scope for records and disclosure controls, not as a technology matter outside audit's remit.”
Compensation
| Market | Band | Notes |
|---|---|---|
| United States | $110k – $200k | IT audit with real technical testing sits well above financial audit |
| India | ₹12L – ₹40L total | Very large market — Big Four, BFSI GCCs, captive internal audit functions |
| Both | — | Data analytics and AI assurance are the two skills that move you up the band fastest |
The book for this field
Securing & Governing AI
The attack surface that is genuinely new, the risks that are old in new clothes, model risk, what audit evidence looks like, the frameworks, responsible AI as controls rather than principles, the enterprise AI gateway, measuring harm, and what the law actually requires.
Also read for this role
Security Engineering & Tooling
What each category of security tool structurally cannot see, what the evidence actually supports, how the market consolidated, and what the 5,197 postings say the job is.
Cross-cutting
Skills every archetype tests
These are shared across every field on this site — the same question asked in different vocabulary — so preparation here compounds rather than being spent once. The book above is specific to your field.
Practice the 12 Security questions → · Back to Cyber, Risk & Internal Audit →