Skip to content

Human in the Loop ​

In Modulos, AI agents make governance work faster, and users stay accountable for it. A person makes the decisions, the approvals, and the attestations.

Mental model

AI gives information. A person makes the decision. No agent sets a Control to Executed, sets a Requirement to Fulfilled, or approves an Asset. Users do these steps in the status and Review workflows. Modulos stores some agent outputs when a run completes, for example assessments and quantifications. Modulos records who started each of these runs, and you can run the agent again.

Principles ​

Modulos applies human oversight with these patterns:

  • Statuses stay human: only a person sets a Control to Executed or a Requirement to Fulfilled, or approves an Asset. Modulos logs each of these changes. One agent output sets statuses back. A failed Agent Test sets each mapped Executed Control to Not Executed, and each Fulfilled Requirement of that Control to Not Fulfilled (Agent Tests).
  • Stored output can change: you can run an assessment or a quantification again. The latest run gives the current result, and Modulos keeps the earlier runs.
  • Auditable runs: Control assessments and Risk Agent runs show who started them and the references that they used. Agent Test runs keep their Run Report. Scout keeps each chat as a conversation for the user who started it. The Evidence Agent keeps no run history: you see only its current suggestions.
  • Stable artifacts: locking rules prevent changes to Evidence and to completed work.
  • Clear data boundaries: Sources are AI System service accounts, and Connectors are user accounts.

Where humans close the loop ​

Agent output and the human decision ​

Each agent has a defined point where a person decides:

  • Scout: answers and proposes. It creates an AI System or a test only when you accept its proposal. It adds files to Evidence Creation only when you accept its Evidence proposal, and you then create the Evidence. It saves an answer as Evidence only when you select Save answer as Temporary Evidence.
  • Evidence Agent: suggests Controls, and in the Evidence dialogs a name and a description. Nothing becomes Evidence until you create or save it.
  • Control Assessment Agent: stores the readiness assessment when the run completes. You examine the result: compare the gaps and recommendations with the Evidence, correct the gaps, and run the agent again. The assessment does not change the status of a Control.
  • Risk Agent: a Quantified run becomes the current value of the threat, with no approval step. You examine the reasoning, and you can replace the value with a new run or a different method.

Learn more:

Reviews and accountability ​

Statuses change in different ways for each item type. The Control Owner or an owner of a mapped Requirement sets the status of a Control directly, and Modulos logs each change. Requirement Owners examine Requirements and mark them Fulfilled. Assets use a formal Review request, in which the reviewer approves or rejects the Asset. These rules make each status change auditable.

Learn more:

Stable artifacts and audit readiness ​

As work continues, Modulos protects the artifacts for an audit:

  • Evidence becomes locked when one of its Controls is Executed (Evidence).
  • Each run of the Control Assessment Agent adds a line with its readiness score to the Comments and Logs tab of the Control.
  • The Quantification History of a threat keeps each Risk Agent run with its reasoning.

Sources and Connectors ​

Agents can use two types of access:

  • Sources are service accounts attached to an AI System.
  • Connectors are user accounts attached to the current user.

Agents can use the shared view of a Source and the personal access of a Connector together.

Learn more:

Responsible AI ​

AI Agents Overview tells how Modulos applies the principles of responsible AI.

Important considerations ​

  • AI can make mistakes. Use outputs as drafts, and compare them with your Evidence and with the system in operation.
  • Scores are not certifications. Use them to find where to put your work. Do not use them as approvals.
  • Use the audit trail. When a decision is important, write the cause of the decision in a comment or a Review. Auditors can then follow it.