How to Talk to Humans: The Art of Asking Questions

By Chen Fradkin, Data & AI Architect

Introduction#

Interacting with poorly designed AI agents can be frustrating. Users often encounter a barrage of questions before accomplishing anything: “What’s your role? Which department? What system?” This creates a fundamental challenge for AI agent builders: agents need information to function effectively, but excessive questioning makes users feel like they’re completing forms rather than collaborating.

The quality of an AI agent depends not just on its knowledge, but on how and when you ask. Great agents function as collaborators, not interrogators.

1. Do the Homework First - Then Come With Questions#

AI agents should demonstrate competence before requesting information. Research, narrow options, and eliminate dead ends independently. Only then should agents ask questions they genuinely cannot resolve alone.

The example contrasts two approaches: a junior executive assistant asks numerous preference questions, while a senior assistant examines calendars and past trips, then presents two targeted options. Both gather information, but the experienced approach requires less user effort.

The first thing they see isn’t a question-it’s results.

2. Structured Questions > Free Text#

When questions are necessary, offer multiple-choice options rather than open-ended responses. This approach:

• Reduces ambiguity
• Minimizes parsing errors
• Builds trust by demonstrating agent understanding
• Speeds interaction

However, options must be genuinely actionable. Only use structured questions when the options are truly actionable-selecting should fully resolve ambiguity.

What Is a Good Option?#

Poor structured questioning still requires follow-up typing, negating efficiency gains. Strong structured questions should eliminate need for subsequent free-text responses.

Batching: Watch for Dependencies#

Agents should batch independent questions into single interactions, minimizing round-trips. Sequential questions requiring dependency resolution may necessitate additional exchanges, but agents should still bundle independent clarifications together.

3. Let the UI Do the Asking#

The most effective questions emerge through purpose-built interfaces rather than conversation. Instead of explicitly asking five questions, agents should surface task workflows with intelligent pre-filled defaults.

Users review and adjust rather than formulate from scratch. The user is still answering questions…but they don’t feel like questions. They are fields in a flow, pre-filled with the AI Agent’s best guess.

4. Spend Your Question Budget Wisely#

Treat each question as a withdrawal from user patience. Structured questions requiring one click carry minimal cost. Open-ended questions demand cognitive effort and incur higher costs. Questions demonstrating the agent wasn’t listening destroy trust entirely.

Before Your Agent Asks, Ask Yourself These Questions#

Filter every question through this checklist:

1. Should the agent solve this independently?
2. Does the agent already possess necessary information?
3. Can multiple questions combine?
4. Is this truly unavoidable?

Bonus: Your Agent’s Voice Matters#

Tone varies by preference, but robotic system-log language fails. Collaborative phrasing-using contractions, “I found” rather than “results found,” invitational questions-creates colleague-like interactions rather than transactional ones.

The Questions Worth Asking#

The goal is never zero questions…Instead, the goal is zero unnecessary questions. Users should think “good thing it asked” rather than “why doesn’t it know this?”

Every unnecessary question shifts interactions toward burden. Well-placed questions create magic.