Field notes
Auditing GenAI: what "audit-ready" actually means
Every organization deploying GenAI eventually says the same sentence: "we need it to be audit-ready." It sounds concrete. It rarely is. Audit-ready is not a state a system reaches. It is a question the system can answer: when someone challenges an output, can you reconstruct why it happened?
I get to look at that question from both sides. In Internal Audit I design the frameworks for auditing GenAI use cases. I also build chatbots, which turns out to be the quickest way to learn what an auditor will ask about one.
A policy is not evidence
A policy document says what should happen. Evidence shows what did happen. The gap between the two is where audit findings live. For GenAI systems that gap tends to open in three places:
- Lineage: which prompt template, which model version and which retrieval source produced this output?
- Data at the boundary: inputs drift silently, so quality checks have to run where the data enters, not where it lands.
- Human overrides: the moments people corrected the model are the most informative records you have, and usually the least captured.
If your GenAI system can't explain itself, your audit team ends up doing it, at deposition speed, after the fact.
Four questions for any chatbot
Before a chatbot meets a customer, an auditor should be able to get a straight answer to these:
- Can a conversation be replayed exactly, with the same prompt, model version and sources?
- Does an answer show where it came from, or only that it sounds confident?
- What does it do when it does not know?
- Who can change its instructions, and is every change logged?
None of this needs a new discipline. They are the old questions (what happened, who approved it, can you prove it) asked of a system that writes its own answers.
Start with the question, not the framework
The practical move is to write the challenge questions first (the five things a regulator, a customer or your own board will ask) and engineer the evidence backwards from there. Frameworks are useful scaffolding, but scaffolding is not the building.