An LLM should not alone decide risk acceptance, exceptions, severity overrides, release blocks or customer-facing statements. These carry accountability, and a model cannot be accountable. Let it prepare the evidence and draft the wording, then have a named person decide and sign.
Sort decisions by consequence
Not every task needs the same oversight. Sort by what happens if the output is wrong.
| Decision | Wrong answer costs | Who decides |
|---|---|---|
| Rewrite a finding summary for clarity | Low | Model drafts, author skims |
| Group duplicate alerts | Low to medium | Model proposes, analyst confirms |
| Suggest a fix version | Medium | Model proposes, tests and rescan confirm |
| Mark a finding as a false positive | Medium to high | Analyst, with code evidence |
| Accept a risk or grant an exception | High | Named owner, with expiry |
| Block or release a build | High | Policy rule, exception by owner |
| Tell a customer they are not affected | High | Security lead, with evidence |
Why accountability matters
Audits and customer reviews ask who made a decision and on what basis. "The assistant said so" is not an answer. A record with the evidence, the reviewer and the date is.
What to give the model
For the low-consequence rows, a model is a real time saver. Give it structured facts and ask for a short, plain summary. Ask it to list assumptions. Ask it to draft the exception request so the owner only has to check and sign.
What to record for each decision
- The finding and its source
- The evidence used, with dates
- Whether an assistant helped and in what step
- The decision and the person who made it
- An expiry or review date for any exception
Keep this in the ticket, not in a chat window that may be deleted.
Common failure modes
- Rubber stamping. Reviewers approve model output without reading. Counter by sampling and re-checking a few each week.
- Authority drift. A model's suggestion becomes the default because it is first in the ticket. Put the evidence first.
- Silent scope growth. A pilot for summaries turns into auto-closing tickets. Write down what the model may do and revisit it.
For deeper treatment of oversight, see how much autonomy should an AI SOC agent have.
Try it yourself
Click through it at your own pace, or open it full screen.
Click through the workflow and mark which steps you would leave to a person.
Where Safeguard fits
Safeguard supports vulnerability management with policies, gates and compliance evidence automation, so decisions and their evidence are recorded in one place. Free scan: https://gold.safeguard.sh/scan.
Quick check
Can a model auto-close duplicate tickets?
Only if closing is reversible, logged, and sampled by a person. Start with proposing duplicates, not closing them.
Who owns a decision that an assistant helped with?
The person who approved it. Record their name.
Last reviewed: 19 October 2026.
