Recommendation engines and agentic systems need different governance. This seems obvious when you say it. It's not obvious in how most organizations build their AI governance policies.
A recommendation engine surfaces an insight or suggests an action. A human confirms, approves, or rejects. The system waits. An agentic system initiates actions directly inside real workflows, against real data, with real consequences. The human may review what happened afterward, if the system logs it properly.
Once software can act rather than simply recommend, the governance question changes. It's no longer "did the AI surface the right information?" It's "what authority does this system have, and who decided that?"
For CTOs structuring AI governance before deployment, the relevant framework isn't about model quality or system reliability. It's about decision rights.
I classify agent actions by two variables: consequence and reversibility. Consequence is the magnitude of what happens if the decision is wrong. Reversibility is how easily and quickly that decision can be undone. Together, those variables assign actions to three tiers.
The first tier is autonomous. Actions that are low-consequence and reversible can be delegated fully to the agent. Routing a ticket, categorizing a document, flagging a record, scheduling a meeting. If the action is wrong, the cost is low and the correction is fast. Human review of these decisions is overhead without proportionate benefit.
The second tier is approval required. Actions where the consequence is meaningful enough, or the reversal is difficult enough, that a human should confirm before the system acts. Vendor payments above a threshold. Changes to a customer's access level. External communications that carry a commitment. These decisions benefit from AI preparation and analysis, but require human confirmation before execution.
The third tier is human-only. Decisions where consequence and reversibility make autonomous action a poor tradeoff regardless of capability. Terminating an employment relationship. Canceling a major contract. Taking an action that exposes the organization to regulatory or legal risk. The question here isn't whether the AI could make the right call. It's whether the organization should delegate the authority to do so.
That distinction matters: decision rights and capability are different questions. A system can be highly capable and still lack the organizational authority to make certain decisions autonomously. The authority decision belongs to the organization, not the model.
This framework produces something a standard AI governance policy doesn't: a map of authorized scope. Before an agentic system deploys, the CTO and relevant business stakeholders should be able to enumerate which actions fall into each tier. That enumeration is the governance document. Not a general policy about responsible AI use, but a specific, auditable list of what this system is authorized to do, under what conditions, and with what approval requirements.
Audit trails are the operational consequence of this framework. Tier 1 actions should be logged with enough context to reconstruct what the system decided and why. Tier 2 actions should capture who approved it and when. When something unexpected happens, the question isn't just "what did the AI do?" It's "was the AI authorized to do that, and was that authorization correct?"
The governance conversation is easiest to have before deployment, when no one has a stake in defending a live decision. After the first incident, after an agent takes an action that was technically within scope but organizationally problematic, the conversation happens under pressure, with less objectivity, and narrower options.
The practical recommendation: before deploying an agentic system, build the decision rights map. Enumerate the actions the system will take. Classify each by consequence and reversibility. Assign each to a tier. Look carefully at the Tier 1 list - it's the list of things your organization has decided to delegate without human confirmation. Make sure that list reflects a deliberate choice, not a default.