1.3 – Generative vs. Agentic AI (Deciding How Much Autonomy to Allow) Lesson

Lesson Overview

This lesson defines the practical line between generative AI and agentic AI, then shows how to cross that line safely. Generative AI drafts content for a person to review. Agentic AI acts on your permissions and directions, calling tools and APIs to pursue a goal. Sending an email for you is agentic. Drafting that email is generative. That distinction matters, because the risks, controls, and accountability differ.

You will learn a staircase approach. Start with drafts reviewed internally. Add a click to approve step when the process is clear. Move to agent actions only after testing shows they are safe, consistent, and reversible. Assign clear ownership of outcomes so accountability is never in doubt.

Oversight is a design choice. You will see where to put human checkpoints, how to pause on low confidence or sensitive terms, and how to sample outputs after release to refine prompts and policies. You will also learn what an emergency brake looks like in practice and when to pull it. Real examples include a customer email agent and a meetings workflow that begins with a review queue and advances to scoped automatic updates. The aim is steady, incremental progress that reduces tedious handoffs while keeping humans in charge.

Who This Is For

If you are choosing how much autonomy to grant AI in real work, this lesson will help you set boundaries, controls, and ownership before scaling. It is especially useful when moving from internal drafts to external actions.

• Teams that send customer emails and want to move from drafts to limited auto send
• Groups that handle refunds, pricing, contracts, or user data and need human review gates
• Operations and support leaders who must define escalations for safety, legal, or financial claims
• Product and platform owners responsible for allow lists, permissions, and rollback tools
• Risk, compliance, and QA reviewers who audit outputs and refine prompts and policies.

Where This Fits in a Workflow

Use this lesson when you are ready to convert a working drafting process into a controlled action. The staircase applies across common flows. For outbound email, start with AI-generated drafts reviewed by a person. Add click to approve into your content or ticket system. After testing and logging are reliable, allow auto send only for low risk cases during business hours, with a clear escalation path for refunds, safety, or regulated terms.

For team meetings, generate summaries, follow ups, and task updates into a review queue. Publish with one click. When quality is proven, grant scoped permissions for the system to commit updates automatically. Any change that touches pricing or contracts pauses for approval. In both cases, logs capture who triggered, who approved, and what changed, so you can reverse actions if needed.

Technical & Workflow Benefits

The manual way asks people to draft, copy, paste, route for approval, and send, which creates handoffs and delays. It also makes audits difficult because information is scattered across tools. The staircase model replaces ad hoc steps with a predictable path: draft, review, approve, then limited action under strict controls.

With generative outputs, “cite or decline” improves trust by showing sources and refusing unverifiable claims. Templates for high stakes messages keep tone and policy consistent. For agent actions, least privilege, allow lists, rate limits, time windows, and budget caps prevent runaway processes. End to end logging and rollback identifiers let you undo unintended changes quickly.

This approach saves time by removing repeated handoffs once quality is proven, yet it keeps control by pausing on sensitive topics and by auditing every action. It shines in customer email and internal knowledge updates where most work is routine, but a small set of cases demand human review.

Practice Exercise

Scenario: You plan a customer email assistant that replies to common inquiries.

• Step 1: Map the staircase. First, have the AI draft replies with citations and place them in a review queue. Add a click to approve step that inserts approved messages into your sending system.
• Step 2: Define controls. Create an allow list of accessible tools and target systems. Apply least privilege permissions. Set rate limits, business hour windows, and a small budget cap. Require logging of prompts, sources, drafts, editor changes, approver, send time, and rollback identifiers.
• Step 3: Set escalations and the emergency brake. Automatically escalate drafts that mention refunds, safety, legal, or financial claims. Train reviewers on when and how to pause the workflow. Test rollback on a safe sandbox.

Reflection: After two weeks, which cases stayed human approved, and which moved to tightly bounded auto send? Could you fully reconstruct any reply from logs and undo it if needed?

Course Context Recap

This lesson sits early in the course and sets a clear boundary between drafting and acting, then shows how to grant autonomy in incremental and reversible steps. You learned to start with internal reviews, add click to approve, and move to agent actions only when testing, logging, and rollbacks are ready. The next lessons continue this steady climb by expanding autonomy only where controls, ownership, and audits are in place. Continue through the course to see more examples, policies, and checkpoints that keep humans firmly in charge while AI handles routine work.