When Should Your AI Say No?
Designing AI Agents That Know Where to Stop
Before an agent does anything, who decided it was allowed to? A hands-on, non-technical workshop on deciding in advance what an AI agent should be trusted to do on its own, and designing the moment where it hands back control.
Dr. Momina Moetesum
National University of Sciences and Technology (NUST), Pakistan
Who Decided the Agent Was Allowed To?
AI agents are moving fast from demos into daily use: booking things, answering customers, drafting decisions, sometimes taking actions on their own. Most conversations jump straight to "is it safe?" and "did we test it enough?". This workshop asks a different, earlier question.
Every agentic system faces a recurring moment: it can act on its own, or pause and hand control to a person. That single decision is where most real-world agent failures actually originate, and it is usually made by accident rather than by design.
Hands-On Result
Work through a shared scenario, map its decision points, and leave with a one-page decision policy you can adapt to a real agent.
The Decision Framework
Workshop Format
Short tutorials alternate with group exercises, so every idea is put straight into practice.
The Moment an Agent Stops Being a Tool
A short cold-open built around 2–3 real cases of agents acting when a human should have been consulted first.
Anatomy of a Decision
Every agent action carries a cost of being wrong. Introducing Act / Ask / Escalate / Refuse as a design tool.
Activity 1: Mapping the Decision Points
In small groups, participants take a shared agent scenario and mark every point where it makes a decision, classifying each by reversibility and cost.
What Makes "Asking" Trustworthy
The human-factors side: how much context a person needs at the handoff moment, what happens when no one responds, and designing friction on purpose.
Activity 2: Act, Ask, Escalate, or Refuse?
Scenario cards drawn from research, healthcare, business and public-service contexts. Groups make and defend a call for each, then compare against how real systems behaved.
Write Your Agent's "No" Policy
Each participant drafts a one-page decision policy for an agent relevant to their own work, and the session closes with shared takeaways.
Includes a 15-minute break after Activity 1
No Coding. Just Better Decisions.
Scenario-based exercises, accessible to technical and non-technical participants alike.
Map Decisions
Classify by Risk
Assign Autonomy
Design the Handoff
Draft a Policy
Map Decisions
Find every point where the agent decidesClassify by Risk
Reversibility and cost of being wrongAssign Autonomy
Act, Ask, Escalate or RefuseDesign the Handoff
A pause that works for the humanDraft a Policy
Your one-page "no" policyQuestions You'll Work Through
Example questions participants answer for a shared agent scenario:
Which of this agent's actions are safe to take without asking anyone?
Which actions are reversible, which are not, and does that change the answer?
What does the agent need to show a human at the moment it asks, so the person can actually decide?
What should the agent do if it asks and no one answers in time?
When is refusing outright the right call, even if the agent could technically proceed?
Upstream of Evaluation and Governance
Singapore's AI community is already fluent in agentic AI, so this is not an "Introduction to AI agents" session. What is less settled, even among experienced teams, is a shared vocabulary and method for design-time judgment calls: how much autonomy is appropriate for a given action, what "asking first" should look like in practice, and when refusal is the correct behaviour rather than a failure.
Where testing and red-teaming ask "does this agent work safely?", this workshop asks "what did we decide 'safely' should mean, before there was anything to test?" The two are complementary: a team that has never decided what its agent may do on its own has nothing precise to evaluate against later.
By the end, you will be able to:
- Distinguish reversible, low-cost agent decisions from irreversible, high-cost ones.
- Apply a practical Act / Ask / Escalate / Refuse framework to real agent scenarios.
- Design a "handoff" moment, where an agent pauses and brings a human in, that actually works for the human on the other end.
- Recognise the design choices (not just technical failures) that lead agents to act when they should have asked.
- Draft a one-page decision policy for an agent of your own choosing.
Participants Will Leave With
Core Message
"An AI agent doesn't become trustworthy by being tested more thoroughly after the fact. It becomes trustworthy when someone has already decided, before it ever acts, what it's allowed to do on its own, and designed a real moment for a human to step in when it isn't."
Who Should Attend?
Prerequisites: no programming experience required. Familiarity with generative AI or LLMs is helpful but not necessary.