# Approval should be set by action type and confidence, and every approval should leave a record

Guide, 20 Oct 2026, read 10 Nov 2026

**The short answer.** Approval works well when it is set per action: read-only steps run alone, reversible changes run in staged rings, and irreversible ones wait for a person. Cogent Security, Onit, Qualys and Tenable each describe such gates on pages read 10 Nov 2026. Ask for the setting, the log and the undo.

## Approve by action type, not by agent

A single switch labelled autonomous or manual is too coarse. An agent that reads a configuration, one that drafts a ticket, and one that deploys a patch carry different risks, and a team will want them at different settings.

The vendor pages read for the register show finer controls. Cogent says customers choose where agents act alone and where they pause for sign-off, configured by action type, severity or asset sensitivity, and that autonomy is staged from recommendations to wider action. Onit states a human approval gate on every agent action, with an audit trail and one-step reversal, and describes approved decisions becoming standing rules that cover matching exposures, logged and reversible.

## Gate by confidence and ring

Qualys describes patch reliability scoring that picks the deployment path: zero-touch for high confidence, ring-staged for medium, and a manual change advisory board for low. It also describes customer-set maintenance windows, a ServiceNow change integration that keeps changes under existing approval, and automatic rollback.

The pattern is useful beyond one vendor. High-confidence changes go out first to a small group, the group is watched, and the change widens only if health holds. Qualys names Dynatrace checks for that watching; if you use a different monitor, ask whether the platform can read it.

| Control | What to ask for | Example stated by a vendor (vendor's claim, read 10 Nov 2026) |
| --- | --- | --- |
| Per-action autonomy | A setting for each action type | Cogent Security: by action type, severity or asset sensitivity |
| Standing rules | Rules created from a reviewed decision, with an undo | Onit: approved decisions become logged, reversible rules |
| Staged rollout | Rings, with a pause on bad health | Qualys: confidence-gated waves, rollback |
| Pause and roll back | A control a person can use at any time | Tenable: schedule, pause or roll back deployments |
| Approval queue | A list of waiting changes with their evidence | Astelia: queue holds the drafted ticket and firewall rule |

## Keep the record that an auditor reads

Approval without a record helps nobody in the next review. For each approved action the platform should store who approved it, what the agent proposed, the evidence shown at the time, the time of change, and the re-check. Tenable states that Hexa AI logs every action for audit. Astelia lists role-based access control and execution logs among its controls.

NIST SP 800-40 Rev. 4 describes patch management as identifying, prioritizing, acquiring, installing and verifying patches. An approval record that covers each of those five steps answers most audit questions.

## Test the gate in the trial

- Ask the platform to propose a change you know you would reject. Confirm it waits.
- Remove an approver and confirm the change does not run.
- Roll back one change and time it.
- Export the log for one week and read it as an auditor would.

## Start narrow and widen on evidence

A sensible first 90 days moves through four settings. In the first weeks, agents run read-only and produce findings and drafts that people read. Then tickets and pull requests open automatically but a person merges or approves each one. Then low-risk, reversible changes, such as a patch to a non-production ring, run without approval. Only after a month of clean results do production rings widen.

Move to the next setting only when the log shows no change that a reviewer would have rejected. Cogent describes autonomy as staged in this way, starting with recommendations and review. Whatever the vendor offers, write your own widening criteria down before the trial starts.

## Separate who approves from who is notified

Approval and notification are different. An approver can stop a change; a person who is notified can only read about it afterward. Name both for each class of action.

A workable split: the owner of the asset approves changes to it, the security lead approves changes to controls such as firewall rules, and the change board is notified of reversible changes made inside an approved window. Where a platform writes into your ServiceNow change process, as Qualys describes, the approver is already named by the change type. Where it does not, the platform needs its own approver list, and that list needs the same review as any access list: removed when people move teams, and checked each quarter.

## Related pages

- [Guides](/guides/)
- [Ownership mapping: finding who can fix a finding](/guides/ownership-mapping/)
- [Automated vulnerability remediation: what an agent can fix without a human](/guides/what-an-agent-can-fix-without-a-human/)
- [Reading a vendor's vulnerability remediation claim](/guides/reading-a-vendors-remediation-claim/)

## Sources

- Cogent Security, autonomy levels: https://www.cogent.com/ai/autonomy-levels (read 10 Nov 2026)
- Onit Security, platform page: https://onit.security/platform/ (read 10 Nov 2026)
- Qualys, agentic AI page: https://www.qualys.com/apps/agentic-ai (read 10 Nov 2026)
- Tenable Patch Management: https://www.tenable.com/products/patch-management (read 10 Nov 2026)
- NIST SP 800-40 Rev. 4, Guide to Enterprise Patch Management Planning: https://csrc.nist.gov/pubs/sp/800/40/r4/final (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://automatedvulnerabilityremediation.com/guides/how-approval-should-work/
