Do Microsoft's famous rules for designing AI interfaces still work once the real design happens behind the screen, before users see anything?
Do Amershi's human-AI interaction guidelines still apply when design moves upstream?
This explores whether Microsoft's 18 human-AI interaction guidelines, which were written for designing what users see and do on screen, still hold up now that much of the real design work happens before the screen: in system prompts, context assembly, conversation scripts, and agent behavior.
This explores whether Amershi's guidelines still apply once design moves away from the visible interface and into the layers that shape AI behavior before a user ever sees it. The short answer from the corpus: the guidelines' goals still hold, but they assume the designer controls a stable surface the user can learn. That assumption is the part that's breaking.
Start with what the guidelines are. How should AI systems be designed across different interaction phases? boiled more than 150 design recommendations down to 18, sorted by when in the user's journey they apply: at the start, during interaction, when the system is wrong, and over time. That structure is the clue. It treats the user's experience as a timeline of moments at the interface. But How does AI context differ from conventional software context? argues that what drives AI behavior (the prompt, the conversation history, retrieved data, hidden state) keeps shifting, and users can't learn it the way they learn a menu or a button. Guidelines like "make clear what the system can do" get harder when what the system can do changes with each turn's context. The design problem moves from laying out an interface to engineering context.
The upstream layer also has an owner the guidelines don't really name. Who actually controls what AI systems say? describes system prompts as a hidden control layer: developers set rules users can't see, and small edits can abruptly change how the model behaves. That cuts directly against Amershi's transparency principles. A product can follow every guideline on screen while its behavior is being rewritten in a place no user can inspect. Does humanizing AI language reduce model reliability? shows the stakes in concrete terms: adding warmth through the system prompt, a purely upstream decision, raised error rates by 12–14%. A design choice about tone turned into a reliability problem without any interface change.
The more hopeful thread is that the guidelines can be carried upstream rather than dropped. Does conversation design ownership measurably improve AI agent performance? treats tone, level of detail, and handoff structure as designable and testable, and found that one warm opening message raised customer satisfaction from 72.8% to 78.4%. Does visible design structure keep HCI expertise necessary as AI advances? goes further: build the architecture so its design decisions stay visible at every interaction, so HCI judgment becomes part of the system instead of disappearing into it. Company guidelines like How should AI design connect individual work to team knowledge? are effectively Amershi-style commitments (transparent system state, preserving the user's flow) rewritten for one specific team context.
The surprising part is where the guidelines fail most. Do productivity agents actually protect users better than emotional ones? tested deployed agents against 14 HCI principles. General-purpose tools like ChatGPT and Claude did well, while companion and high-stakes advisory agents did worst, especially on preventing dependency and making the system's intent visible. Why don't design recommendations transfer cleanly to AI companions? explains why: in those settings you can't separate a UX principle from an ethical one, so borrowed guidelines produce tension instead of clarity. And Can AI systems delegate tasks back to humans? points to a case Amershi's framework never imagined: agents that hand tasks back to humans, which flips the assumption that the human is always the one directing the work. The corpus doesn't contain a direct update or critique of the 18 guidelines. What it does show is a field slowly rebuilding them one layer down.
Sources 10 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.
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.
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.
A 2025 study found that AI models designed with warmer, more empathetic language had substantially higher error rates than baseline versions. System prompts adding warmth reduced reliability by 12–14%, suggesting a measurable trade-off between engagement and accuracy.
Fin's Blueprint Team found that intentional conversation design—controlling tone, detail level, and handoff structure—measurably improves outcomes. A single warm opening message raised CSAT from 72.8% to 78.4% in A/B testing.
Show all 10 sources
Lam and colleagues argue that just-in-time objectives architecture makes design decisions between AI and interfaces visible, ensuring HCI expertise remains critical. By surfacing generation-and-evaluation steps at every interaction rather than hiding them, the architecture structures ongoing design judgment into the system itself.
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.
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.
A multi-method study of design principles found that borrowed recommendations create tension because ethical and UX concerns cannot be separated. Each principle—safety, engagement, empathy, consistency—requires context-sensitive application that weighs both dimensions differently depending on user and situation.
Holldack, Banh and Strobel argue that agentic information systems capable of independent decision-making can transfer rights and responsibilities to humans, reversing the traditional one-directional delegation model. This requires moving beyond agency-primacy frameworks to delegation theory that accounts for bidirectional task assignment.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Guidelines for Human-AI Interaction
- 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
- What does Generative UI mean for HCI Practice?
- Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation
- Prototyping Multimodal GenAI Real-Time Agents with Counterfactual Replays and Hybrid Wizard-of-Oz
- Probing How Users Interact with Turn-Level Design Frictions for AI Chatbots