Intent Preview
CanonicalConfidence
Cognitive Load
Low
Evidence
production validated
Impact
feature
Live Preview
Agent plans to...
Pull the last 3 similar invoices from the billing database, cross-reference current pricing from the pricing engine, and draft an invoice using the standard template.
Data sources
Ethical Guardrail
Agent must never act without first showing the plan. Preview must include data sources and reasoning, not just the proposed action.
Design Intent
Users fear the unknown. When an agent is about to act, the single most effective way to build trust is to show them the exact plan before it executes. Intent Preview forces the agent to show its work -- outlining every step, data source, and expected outcome.
Psychology Principle
Users fear the unknown. Showing the exact plan before execution builds trust and eliminates reactance.
Description
Force the agent to show its work before acting -- outline every step, data source, and expected outcome for informed consent.
When to use
Any agent action at confirm_execution or autonomous level. Also recommended for high-impact suggest_only actions.
Example
Claude/Grok Agent: Before any autonomous edit -- full card preview with bullet steps, data sources, and one-tap Looks good -- go ahead or Let me tweak this first.
Autonomy Compatibility
Behavioral Objective
Users review the plan, approve or edit it, and proceed with full understanding and confidence.
- Higher acceptance rate of agent actions
- Fewer post-action corrections
- Stronger long-term trust in the agent
Target Actor
role
Everyday user
environment
High-stakes or AI-driven workflows
emotional baseline
Wants visibility before automation
ai familiarity
medium
risk tolerance
medium
Execution Model
preview_generation
Create a clear, scannable summary of the planned action.
User does not understand the plan.
transparency_layer
Explicitly list data sources, logic, and risks.
User wonders where the data came from.
action_options
Provide immediate, low-friction next steps.
User feels forced to accept or reject without nuance.
Failure Modes
Preview is too long or technical
Default to short version with Show details
Preview arrives too late
Always show before autonomous execution
Preview hides uncertainty
Explicitly call out low-confidence areas
User ignores the preview
Make it visually prominent and required
Preview feels like a formality
Tie it directly to editable fields
Agent Decision Protocol
Triggers
- Any action at confirm_execution or autonomous level is proposed
- User has low historical trust
- Task involves data changes or high stakes
Escalation Strategy
L1: Diagnose the failing element via behavioral_signals
L2: Nudge -- adjust copy, timing, or visual salience
L3: Restructure -- simplify flow, add progressive disclosure, restructure form
L4: Constrain -- lock Autonomy Dial to confirm_execution, add Strategic Friction
L5: Yield -- flag for human designer or domain expert review
Example
Agent preparing to auto-fill a section -> Preview: I will pull language from your last 4 similar documents, apply your company standards, and flag 2 potential risks. Proceed or edit?
Behavioral KPIs
Primary
- Preview review rate before approval
- Edit rate after preview
- Post-action satisfaction
Risk
- Surprise overrides or cancellations
- Distrust after action
Trust
- User-reported I always know what the agent will do
- Autonomy Dial usage when preview is shown
Decay Monitoring
Revalidate when
- Agent capabilities expand
- New data sources are added
- User expectations for transparency evolve
Decay signals
- Rising surprise or override rates
- Drop in preview engagement
- Feedback that the preview no longer matches reality
Pattern Relationships
Amplifies
Requires
Conflicts with