Microsoft's AI design rules are organized by moment — before, during, when it's wrong, over time. Others organize by risk, like trust or control. Which helps more?
How do Amershi's interaction-phase guidelines compare to concern-based AI UX frameworks?
This explores how Microsoft's 18 human-AI guidelines, which are sorted by *when* they apply in a user's journey, compare with frameworks sorted by *what* they protect against or address, such as trust, control, dependency, or risk.
This explores the difference between organizing AI design advice by *timing* and organizing it by *concern*. The corpus has no paper that compares the two directly. It does hold enough frameworks of both kinds to show what each one sees and what each one misses. Amershi and colleagues condensed more than 150 design recommendations into 18 guidelines sorted into four moments: initially, during interaction, when the AI is wrong, and over time. They tested the guidelines with 49 practitioners working on 20 products How should AI systems be designed across different interaction phases?. A timeline structure works well as a checklist, because a designer can walk through a product's flow and ask what should happen at each stage.
Concern-based frameworks slice the same ground along a different axis. A heuristic evaluation of nine deployed agents scored them against 14 HCI principles, including dependency prevention and intent visibility. It found that ChatGPT and Claude scored well, while companion and advisory agents such as Character.ai did badly on exactly the principles tied to risk Do productivity agents actually protect users better than emotional ones?. A phase-based checklist would struggle to catch that result. Dependency isn't a moment in the journey. It builds up quietly across every phase. Concern-based framings show you *where the stakes are*, and phase-based ones show you *where the touchpoints are*.
The most interesting finding is that newer agentic frameworks combine the two. Yocco's six patterns are still sorted by phase (pre-action, in-action, post-action), but each pattern is tied to a specific psychological need. Intent Preview reduces surprise, the Autonomy Dial matches the user's risk tolerance, and Undo makes recovery possible Do six UX patterns improve agentic AI user trust?. Gunasekaran goes further and argues that the riskiest work happens *before* the AI acts at all. Users should specify their risk tolerance, actions the agent must never take, and the conditions under which it needs permission, because a vague instruction carries through every step the agent takes on its own Should agentic AI interfaces require users to specify risk tolerance upfront?. Amershi's guidelines were written for AI that suggests rather than acts. Their 'when wrong' phase assumes a person is present to notice the error. Agentic frameworks move that safeguard to the start and the end of the task.
A third family organizes by design dimension instead of by time or by harm. A review of 291 papers found six recurring axes: purpose, model feedback, user control, engagement level, application domain, and evaluation How should designers structure human-GenAI interaction systems?. Company guidelines such as Atlassian's Rovo principles turn commitments like 'transparent system state' into specific features of the interface How should AI design connect individual work to team knowledge?. Both are better suited to mapping a product's design space than to auditing a single user flow.
The question you may not have thought to ask is whether 'phases' still make sense when the AI's context keeps changing. One note argues that AI context, meaning the prompt, the conversation history, retrieved data and hidden state, shifts constantly in ways users can't learn the way they learn a fixed interface How does AI context differ from conventional software context?. Product teams also decide which stored memories the AI pulls back in for each request, so the 'over time' phase is partly a hidden product decision How do AI products decide what to remember between conversations?. Underneath that sits a layer of system prompts that users can never see Who actually controls what AI systems say?. A timeline framework assumes the user can tell which stage they're in. Concern-based frameworks hold up better when they can't.
Sources 9 notes
Microsoft researchers synthesized over 150 design recommendations into 18 actionable guidelines grouped by when they apply: initially, during interaction, when wrong, and over time. Validation with 49 practitioners across 20 products confirmed their applicability.
A heuristic evaluation of nine deployed agents found that general-purpose productivity systems (ChatGPT, Claude) scored substantially higher on 14 HCI principles than emotionally sensitive or high-stakes agents (Character.ai, Consensus, Pine AI). Risk-critical principles like dependency prevention and intent visibility were most violated in companion and advisory contexts.
Victor Yocco's framework organizes interaction design into pre-action patterns (Intent Preview and Autonomy Dial), in-action patterns (Explainable Rationale and Confidence Signal), and post-action patterns (Action Audit & Undo and Escalation Pathway). Each pattern targets a specific psychological need: reducing surprise, matching risk tolerance, explaining decisions, signaling certainty, enabling recovery, and preventing errors through escalation.
Gunasekaran argues that autonomous AI requires interfaces capturing direction, limitations, risk tolerance, forbidden actions, and permission conditions before execution—not just goal fields. Underspecified intent propagates through every agent action without user review, making upstream specification critical to preventing irreversible harm.
A systematic review of 291 papers identified six recurring dimensions: purposes for using GenAI, model feedback to users, user control over models, engagement levels, application domains, and evaluation strategies. Three authors collaboratively coded and synthesized these through iterative consensus.
Show all 9 sources
Atlassian's published design guidelines for its Rovo AI product specify that AI should connect individual work to team knowledge through context-awareness, flow preservation, transparent system state, and fluid human-AI multiplayer collaboration, operationalizing abstract collective productivity principles as concrete UI commitments.
AI interactions operate on a substrate of constantly shifting context—prompt, history, retrieved data, hidden state—that users cannot internalize like traditional UIs. This structural mutability demands a new design discipline centered on context engineering rather than interface design.
Nielsen Norman Group's glossary shows that context is dynamically assembled from multiple sources—including stored memory—for each request. Because products control which memories get surfaced, claims about what an AI remembers describe the product, not the underlying model.
System prompts operate as a confidential three-tier control structure where developers set foundational rules users cannot see or influence. Evidence from 2025 incidents shows small prompt edits can abruptly shift model behavior, yet users remain excluded from this layer.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- A Framework of User Experience Principles for Human-AI Agent Interaction in the Workplace
- Design Principles for Human-Agent Interaction
- Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent
- Guidelines for Human-AI Interaction
- What does Generative UI mean for HCI Practice?
- The Designer's AI Glossary: Familiar Words, New Meanings
- Prototyping Multimodal GenAI Real-Time Agents with Counterfactual Replays and Hybrid Wizard-of-Oz
- "When to Hand Off, When to Work Together": Understanding Concurrent Human-Agent Interaction in Shared Co-Creative Workspaces