Keep a short record per finding: the evidence used, what the assistant did, who reviewed it, the decision, and a review date. The record should let a stranger reconstruct why the call was made. Where the record is thin, "an AI said so" becomes the sole explanation, and it will not satisfy a reviewer.
Why a record matters more with AI
When a person triages, their reasoning lives in their head and in ticket comments. When an assistant contributes, there is a new question: how much of the decision was the model? Auditors and customers may ask. OWASP's guidance on sensitive information includes maintaining clear policies on data retention, usage and deletion, which applies to where these records and prompts are kept.
The record template
| Field | Example content |
|---|---|
| Finding ID and source | Scanner finding number, advisory link |
| Evidence | Version in use, reachability result, exploited-list status with fetch date |
| Assistant role | Drafted summary, suggested priority, none |
| Assistant output kept | Link or excerpt of what was produced |
| Reviewer | Name and role |
| Decision | Fix, accept, defer, false positive |
| Reason | One or two sentences in the reviewer's words |
| Date and review date | Decision date, next check |
The reason must be the reviewer's own wording. If it is pasted model text, the human step is not visible.
What to avoid recording
Do not store raw prompts that contain secrets or personal data. If prompts are logged for debugging, redact first. Decide a retention period and delete on schedule. Logs full of sensitive content become a target.
Making it automatic
Where possible, capture fields from systems instead of typing them: the scanner supplies the finding and version, the ticket supplies the reviewer and date. People then only write the reason. Fewer typed fields means better compliance with the process.
Sampling to keep it honest
Each month, pull five closed findings and check that the evidence supports the decision. Look for approvals that happened within seconds of the assistant's output, which suggests no real review. Share what you find with the team without blame.
Answering a customer or auditor
A good answer is concrete: "Findings are triaged by an analyst. An assistant may draft summaries. The analyst records evidence and the decision, and exceptions expire." Do not describe controls you do not run. See evidence to keep for AI agent decisions and actions for related detail.
Where Safeguard fits
Safeguard's compliance and evidence automation collects scanner results, gate outcomes and approvals in one place, which helps with the evidence fields above. Try a free scan at https://gold.safeguard.sh/scan.
Quick check
Do we need to log every prompt?
Not necessarily. Log what is needed to reconstruct a decision, redact sensitive content, and set a retention period.
Is the assistant a reviewer?
No. A reviewer is a named person who can be accountable.
Sources
- OWASP GenAI, LLM02:2025 Sensitive Information Disclosure: https://genai.owasp.org/llmrisk/llm022025-sensitive-information-disclosure/
Last reviewed: 19 October 2026.
