Human in the Loop: Governing and Controlling AI Agents in the Enterprise

The myth of full autonomy

Technology media love demonstrations of "fully autonomous" AI agents. The reality of enterprise deployments is more nuanced — and that's a good thing. The question isn't whether a human should be in the loop, but where and how to integrate them to maximise both efficiency and reliability.

The four supervision models

Model 1 — Systematic validation

Every agent action is submitted for human validation before execution. Used for: irreversible actions (data deletion, client email sends, ERP modifications), regulated processes (medical validation, financial sign-off). Cost: low efficiency gain. Benefit: zero risk.

Model 2 — Exception-based validation

The agent acts autonomously but flags ambiguous cases for validation. This is the optimal model for most operational processes. The agent defines its own confidence thresholds: below 85%, it escalates to a human.

Model 3 — Post-hoc supervision

The agent acts fully autonomously. A human reviews activity logs periodically (daily, weekly). Used for: low-risk, high-volume, highly repetitive processes (reporting, notifications, triage).

Model 4 — Anomaly-based supervision

A monitoring system automatically detects unusual agent behaviour and alerts a human only when an anomaly is detected. Requires rigorous prior definition of anomaly metrics.

Building a governance framework

For every agent deployed in production, I recommend documenting:

The Swiss regulatory context

In Switzerland, regulatory requirements in finance, healthcare and public administration impose strict traceability levels. Any automated decision with an impact on a third party must be explainable. This doesn't mean banning agentic AI — it means designing workflows with auditability as a design constraint, not an afterthought.