Course Content
Agentic AI Patterns
9 sections · 50 lessons
What are the ethical considerations specific to Agentic AI?
What you need to know
Why agents raise new ethical questions
A chatbot that gives a bad answer can be ignored. An agent that denies a claim, sends an email, or moves money has already acted. Three things are new:
- Action. Harm happens before a human reads anything.
- Scale. The same decision logic runs thousands of times a day.
- Opacity of the path. The agent chose its own steps, so "why did it do that?" needs a trace to answer.
The main considerations
| Issue | What it means | Engineering control |
|---|---|---|
| Accountability | "The agent did it" is not an answer to a customer or regulator | Named owner per agent; audit log per action |
| Disclosure | Users should know it is an AI; never impersonate a real person | Clear labelling in UI and emails |
| Consent and data minimisation | Many tools make it easy to pull more personal data than needed | Scope tools to the task and the user's permissions |
| Reversibility | Some actions cannot be undone | Human approval on irreversible actions |
| Bias at scale | One bias repeated automatically | Measure outcomes by group; sample reviews |
| Autonomy creep | Tools get added and scope grows quietly | Re-review permissions like IAM, on a schedule |
| Labour impact | Work is displaced or changed | Be honest about what is automated |
The regulatory frame
Laws already reach agents. The EU AI Act requires human oversight and logging for high-risk uses such as credit scoring and insurance pricing. India's Digital Personal Data Protection Act requires purpose limitation and consent for personal data. You do not need to quote articles in an interview, but saying "high-risk decisions need a human who can override" shows you know the direction.
A real-life example
A health-insurance claims agent drafts approve or deny decisions. It has been trained and prompted on past adjuster decisions.
After two months, an audit finds that claims from two smaller cities are denied 9 points more often than similar claims from metros. The cause: those hospitals submit handwritten discharge summaries, the agent more often marks them "insufficient documentation", and the old human decisions it learned from had the same pattern.
The fixes are engineering fixes. Denials always go to a human adjuster, never automatic. Outcome rates are tracked by city tier and hospital type every week. "Insufficient documentation" now triggers request_document instead of a denial. Every decision shows the evidence the agent used, so the adjuster can see why.
Follow-up questions to expect
- "Who is responsible when an agent makes a mistake?" — The organisation that deployed it, and a named owner inside it. Design so that owner can see and reverse what happened.
- "How do you detect bias in an agent?" — Compare outcome rates across groups on real traffic, and run paired test cases that differ only in a protected attribute or its proxy.
- "Should an agent say it is an AI?" — Yes. Many jurisdictions require it, and trust breaks badly when users discover it later.