Designing for trust in LLM-powered workflows
Trust in an LLM workflow isn't binary. It's a spectrum shaped by every design decision: what the system shows, what it hides, and what control it gives the user. Most teams treat trust as a product outcome — something users either have or don't. I treat it as a design input.
Trust is earned through transparency
The fastest path to user trust isn't accuracy — it's legibility. Users will forgive mistakes if they understand why the system made them. They won't forgive a black box that produces wrong answers with no explanation.
This means designing to show the reasoning, not just the result. For an AI analytics tool, that might mean surfacing the SQL the model generated alongside the natural language answer. For a recommendation system, it means showing the inputs that drove the recommendation. The user doesn't need to read all of it — they just need to know it's there.
The role of controls
Controls matter more than most teams realize. Giving users the ability to edit, rerun, and override an AI output does two things. First, it creates a safety net — users don't have to commit to an output they're unsure about. Second, it signals that the system is a tool, not an oracle. That framing alone changes how users interact with it.
Calibrated confidence signals
Confidence scores are widely misused. A percentage without context â "72% confident" â communicates almost nothing. Worse, it can create false precision. What works better is relative confidence: "this answer is based on complete data" versus "this answer is estimated from partial data." Qualitative signals that map to real-world implications.
The goal is calibration: users should be more cautious when the system is uncertain, and more confident when it has strong signal. That only happens when confidence signals are designed to be interpretable, not just present.