In practice, each workflow or action type an AI agent can take is assigned a gate during configuration - not left to the model's judgment at runtime. Fully Automated actions execute without any human touch because their downside is limited and reversible. Human Approval Required actions pause the conversation and route a recommendation to an agent for sign-off before execution. Never Permitted actions are blocked entirely, regardless of what the AI infers the customer wants, closing off categories of action the organization has decided AI should not perform.
This tiered structure matters most in regulated MENA sectors like banking, healthcare, and government, where a single unauthorized action - releasing funds, closing an account, changing a medical record - can trigger compliance breaches under frameworks such as CBUAE or SAMA. Dynamic Action Gating lets an AI concierge handle high volume, low-risk requests in Arabic and English across WhatsApp and other channels at scale, while ensuring anything consequential still passes through a human, satisfying auditors and regulators without slowing everyday service.
In Eshal: Every deployed workflow is classified at configuration time, not left for the AI to decide mid-conversation. Standard returns may be Fully Automated; large refunds are Human Approval Required; account closure is Never Permitted. This classification is reviewed with each client during onboarding and can be adjusted as trust in a given workflow's performance grows, keeping oversight aligned with actual risk rather than a fixed, one-size-fits-all policy.