Draft v0.1

Control

Control defines how users can guide, approve, modify, pause, or override AI behavior within a workflow. As AI systems take on more responsibility, users must still have the right level of influence to supervise actions, prevent errors, and recover safely when needed. 1

Control defines how users can direct, constrain, approve, modify, pause, or override AI behavior within an experience. It ensures that people remain appropriately empowered as AI systems become more capable—from reviewing suggestions to supervising multi-step actions and autonomous operations. 1 2 3

Control is not about forcing users to manually manage every AI interaction. It is about giving them the right level of influence for the context, the task, and the level of risk—so they can confidently guide outcomes, intervene when needed, and recover from errors. 1 2 3

Users should be able to understand when AI is active, what it is doing, what it plans to do next, and what options they have to steer, approve, correct, or stop it. Strong control mechanisms help prevent overreliance, reduce unintended actions, and support safer adoption of more capable AI systems. 2 3

Principles

1. Control should match the level of AI responsibility

As AI moves from informational support to assistive, agent-assisted, and autonomous behavior, the form of user control should evolve accordingly. Lower-risk systems may only require lightweight guidance or dismissal, while higher-risk or operational systems require stronger controls such as approval gates, interruption, rollback, and auditability. 1 3

2. Users should always know when they can act

Control is only meaningful when users can recognize when AI is active, when action is required, and what choices are available. Interfaces should clearly signal when AI is generating, recommending, or executing, and make control points easy to find and understand. 2

3. Users should be able to intervene before and after AI action

Good AI UX supports control both upstream and downstream. Users should be able to shape inputs and intent before AI acts, and they should also be able to review, revise, undo, pause, or override outcomes after AI produces or executes something. 2 1

4. Control should be proportional to workflow risk

The more consequential, irreversible, or wide-reaching an AI action is, the more explicit and robust the control model should be. Systems acting on production environments, infrastructure, policies, or customer-impacting workflows should provide stronger safeguards than systems that only summarize information or draft content. 1 3

5. Control should support recovery, not just approval

Approving or rejecting an AI action is only part of control. Users also need ways to recover from mistakes, inspect what happened, and regain a safe system state through pause, override, rollback, and audit mechanisms where appropriate. 1 2

Guidelines

Make AI activation explicit

Users should always know when AI is active, when an agent has been invoked, and whether the system is only suggesting, planning, or acting. Do not activate AI silently in ways that remove user awareness or the ability to opt out. 2

Show what the AI is doing and what happens next

Provide clear status, progress, and next-step messaging so users can understand the current state of the AI and anticipate the impact of their choices. For agentic systems, show planned actions before execution whenever possible. 2 1

Make actions and consequences clear

Action labels should describe what will happen in plain language. Users should not have to guess whether a button will accept a recommendation, start a workflow, apply a change, or trigger autonomous behavior. 2

Support modification before commitment

When AI produces drafts, recommendations, or plans, let users edit parameters, refine the request, choose alternatives, or adjust scope before accepting the outcome. This is especially important for assistive and agent-assisted systems. 1 2

Require explicit approval for consequential actions

If AI will make changes to systems, configurations, policies, or operational workflows, require explicit user approval at the appropriate decision points. Approval should become stronger as the potential impact or blast radius increases. 1 2 3

Provide interruption and escape hatches

Users should be able to stop, pause, or override AI behavior when it is no longer aligned with their goals or when conditions change. Escape hatches should be visible and available at the moments users are most likely to need them. 1 2

Enable undo, rollback, or reversal where possible

For actions that change system state, provide a recovery path whenever technically feasible. If full reversal is not possible, clearly explain the consequences, offer compensating actions, and preserve an audit trail. 1 2

Keep controls easy to find during the workflow

Do not bury key controls in settings or secondary menus. Approval, rejection, pause, override, and review mechanisms should appear in context and at the point of decision. 2

Preserve visibility into past AI actions

Users need to understand what the AI did, why it did it, and what data or policies informed the outcome. Logs, reasoning summaries, and action histories support accountability and help users maintain confidence in operational workflows. 1 2 3

Control across AI capability models

Informational AI

Control is lightweight. Users choose whether to use the output, ask follow-up questions, or ignore it. The experience should support verification, source review, and easy dismissal without implying that the AI is authoritative. 1 2

Examples 1 - Documentation chat with linked sources and follow-up prompts. 1 - “Why is this happening?” explanations that users can inspect, ignore, or refine. 1

Assistive AI

Control centers on evaluation and choice. Users should be able to compare suggestions, edit drafts, accept or reject outputs, and understand the rationale or confidence behind recommendations. 1 2

Examples 1 - Configuration recommendations with accept, edit, and dismiss actions. 1 - Report generation assistance where users can revise content before finalizing it. 1 - Step-by-step setup guidance that users can follow, modify, or skip. 1

Agent-Assisted AI

Control becomes supervisory. Users delegate work to the AI, but they should be able to inspect plans, approve key actions, pause execution, override decisions, and review the result of each step. 1 2

Examples 1 - Guided remediation workflows that show planned actions before execution. 1 - Multi-step configuration automation with approval at key checkpoints. 1 - Ops orchestration flows that can be interrupted or rolled back if conditions change. 1

Autonomous AI

Control shifts from real-time supervision to governance, policy, and recovery. Users may not approve every action individually, but they must still be able to define boundaries, monitor behavior, inspect outcomes, pause the system, and intervene when needed. 1 2 3

Examples 1 - Auto-scaling systems that operate within defined thresholds and policies. 1 - Self-healing systems that automatically remediate known issues with logging and override controls. 1 - Policy-driven traffic routing adjustments with visible rules, monitoring, and rollback paths. 1

What good control looks like

A well-controlled AI experience helps users answer four questions throughout the workflow: 2 3

What is the AI doing right now? 2
What will it do next if I continue? 2 1
What choices do I have to guide, approve, modify, or stop it? 2
How do I recover if the outcome is wrong or conditions change? 1 2

Summary

Control means users retain meaningful influence over AI behavior at the right points in a workflow. As AI systems take on more responsibility, control must evolve from lightweight guidance and evaluation to approval, supervision, policy-setting, interruption, and recovery. Designing for control helps users work with greater confidence, prevents overreliance, and enables safer use of more capable automation. 1 2 3 ``