Skip to content
Compliance Automation Tools

Home / Guides / An AI agent in a compliance tool may read and draft alone, and needs a person to submit, change or accept

Guide, read 13 Nov 2026

An AI agent in a compliance tool may read and draft alone, and needs a person to submit, change or accept

Framework Coverage MatrixAll guides

Split actions by what they change, not by how hard they are

The useful test for an agent action is what it changes and who can undo it. Reading changes nothing. Drafting creates something a person can discard. Submitting, publishing, approving and writing to a live system create effects outside the tool. Put those four groups on separate approval levels and the rest follows.

NIST AI RMF 1.0 (NIST AI 100-1, January 2023) frames the oversight choice as a design decision. It says human-AI configurations can span from fully autonomous to fully manual, and that AI systems can autonomously make decisions, defer decision making to a human expert, or be used by a human decision maker as an additional opinion. It adds that some systems may not require human oversight and others specifically do. For a compliance tool, the answer depends on the action, so set it per action. This page is about the operating rule, not about governance programs.

Three levels of approval cover the actions that matter

Level 0 is unattended: the agent acts and the log records it. Level 1 is review after: the agent acts, and a named person reviews a batch on a schedule. Level 2 is approval before: the agent prepares, and a named person approves in the tool before anything leaves it. Use Level 2 for anything an auditor, regulator, customer or production system will see.

This site calls the Level 2 pause a human approval gate. The request a person sees should be short: the action, the target system, the exact change or document, the evidence behind it, and the framework requirements it touches. A gate needs three things: a named approver, a view of exactly what will be sent or changed, and a record of the decision. A button that approves a summary while the agent changes a different thing underneath is a gate in name only.

ActionLevelReason
Read evidence from a connected system, read-only0 unattendedChanges nothing
Draft a policy revision or a control narrative0 unattendedDraft stays inside the tool
Flag a failing test or an aging evidence item0 unattendedNotice only, no outside effect
Match evidence to controls and propose tags1 review afterWrong tags are cheap to fix but can hide gaps
Send a reminder to a control owner1 review afterReaches a person, low harm
Submit evidence or a response to an auditor2 approval beforeLeaves the tool, cannot be recalled
Change an approved policy2 approval beforeAlters a signed document of record
Accept or close a risk, or grant a policy exception2 approval beforeOnly a person with authority can own it
Write to a production system, such as changing a setting2 approval beforeLive effect, needs change control

Least privilege applies to the agent as a user

NIST SP 800-53 Rev 5 control AC-6 reads: employ the principle of least privilege, allowing only authorized accesses for users (or processes acting on behalf of users) that are necessary to accomplish assigned organizational tasks. An agent is a process acting on behalf of a user. Its credentials should be read-only wherever it only reads, scoped to the systems it monitors, and separate from any human account.

The AC-6 family has 10 enhancements in the NIST OSCAL catalog, and three apply directly. AC-6(5) covers privileged accounts, AC-6(9) says to log the use of privileged functions, and AC-6(10) prohibits non-privileged users from executing privileged functions. If an agent can write to production, give it a separate write credential that only exists during an approved action, and log every use. See least privilege access and the audit trail term.

Defined roles and a way to stop the agent

The AI RMF has subcategories that read as operating requirements. GOVERN 3.2 says policies and procedures are in place to define and differentiate roles and responsibilities for human-AI configurations and oversight of AI systems. MAP 3.5 says processes for human oversight are defined, assessed, and documented. MANAGE 2.4 says mechanisms are in place, and responsibilities assigned and understood, to supersede, disengage, or deactivate AI systems that demonstrate performance or outcomes inconsistent with intended use.

In a compliance tool, that translates to four checks. Is each approval level written down per action? Is the approver named, and not a shared inbox? Can an administrator switch the agent off or drop it to Level 2 for everything in one step? Is there a record of what the agent did while on? MANAGE 4.1 adds that post-deployment monitoring plans include mechanisms for appeal and override, decommissioning, incident response, and change management. In plain terms, a person who disagrees with what the agent did must be able to reverse it and say so in the record. Ask a vendor to show all four on screen. NIST SP 800-53 AU-10 on non-repudiation also helps here: it asks for evidence that an individual, or a process acting for one, performed an action, and it lists approving information as a covered type. An approval recorded that way is stronger than a chat message. Where a vendor publishes its own claim about agent autonomy, treat it as the vendor's statement until you see it work.

A short rule set to adopt this quarter

The CSF 2.0 gives a place to hang the log: its Detect function includes the Category DE.CM, Continuous Monitoring, and its Govern function holds the Categories for roles and responsibilities. An agent's activity record is monitoring data, and it belongs in the same review as everything else. Start with the table above and change the levels only toward more review, never less, until you have 90 days of logs. Review the Level 1 batch weekly and record how many tags a person changed. If a person changes more than a small share, move that action to Level 2. Keep the credentials read-only by default.

Two things stay with people regardless of agent quality: accepting risk and attesting to an auditor. Both carry accountability that a tool cannot hold. Related reading is policy exceptions and the platform matrix, where agent approval controls are one of the things to compare.

Where this shows up in the index

On agents that act under control, which carries 15 of 100 points, Vanta scores highest at 10 of 10 and Scytale scores 7.

Sources

  1. nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf, read 12 Nov 2026
  2. raw.githubusercontent.com/usnistgov/oscal-content/main/nist.gov/SP800-53/rev5/json/NIST_SP-800-53_rev5_catalog.json, read 12 Nov 2026
  3. nvlpubs.nist.gov/nistpubs/Legacy/SP/nistspecialpublication800-137.pdf, read 12 Nov 2026
  4. nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.29.pdf, read 12 Nov 2026

Last reviewed 13 Nov 2026.