Microsoft wrote a broad playbook for AI interactions; Salesforce wrote five rules from real mistakes — where does the broad one leave gaps?
How do Salesforce's guidelines compare to Amershi's broader human-AI interaction phases?
This explores how Salesforce's five practical agent-design rules, drawn from building its own Agentforce product, line up against Microsoft's earlier 18 human-AI interaction guidelines, which Amershi and colleagues organized by stage of the user's journey. It also asks what each set covers that the other leaves out.
This compares a narrow, product-tested rulebook with a broad map of the whole user journey. The useful finding is less about where they agree and more about where Salesforce's rules stop. Amershi and colleagues at Microsoft condensed more than 150 design recommendations into 18 guidelines. They grouped them by when each one applies: when a user first meets the system, during an interaction, when the system gets something wrong, and over time. They then checked the guidelines with 49 practitioners across 20 products How should AI systems be designed across different interaction phases?. Salesforce took the opposite route. It started from concrete usability failures its team saw while using Agentforce internally. From those it drew five rules: agent names under 10 characters, consistency-based heuristics, a structured five-part conversation, mandatory disclosure that the user is talking to AI, and standardized visual branding What design rules improve AI agent usability in practice?.
If you sort Salesforce's rules into Amershi's phases, almost all of them land in the first two. Disclosure, naming and branding cover the first encounter: they tell users what they're dealing with. The five-part conversation structure covers the middle of an interaction. Salesforce says very little about what Amershi calls 'when wrong' (how to recover from a mistake) and 'over time' (how the system learns and adapts). This isn't a flaw so much as a sign of where the source material came from. Internal testing surfaces naming confusion and broken conversation flow quickly. It's much less likely to surface slow problems like over-reliance or trust that fades over weeks. Fin's Blueprint Team ended up in similar territory to Salesforce. They argue that conversation design (tone, level of detail, how handoffs are structured) deserves its own owner. In their A/B test, one warmer opening message raised customer satisfaction from 72.8% to 78.4% Does conversation design ownership measurably improve AI agent performance?. Practitioner guidance for agents seems to gather around the opening and the conversation itself.
Other notes in the collection help fill the phases Salesforce skips. On handling things before they go wrong, conversation analysis suggests agents should pause to ask clarifying questions instead of quietly chaining tool calls and drifting from what the user wanted When should AI agents ask users instead of just searching?. Cocoa lets users revise a plan step by step while the agent is still carrying it out, so correction becomes part of the work instead of an after-the-fact repair Can users steer AI agents better by revising plans mid-execution?. The 'autonomy slider' pattern lets users set how independently an agent acts on each task, from suggest-only to draft to execute Should agent autonomy be a continuous slider or fixed levels?. None of these concerns appears in Salesforce's five rules, but all of them follow from Amershi's 'when wrong' and 'over time' phases.
The less obvious point is that Amershi's framework comes from a time when an AI feature sat inside a fairly stable interface. Agents strain that assumption. One note argues that AI context (the prompt, the conversation history, retrieved data, hidden state) keeps shifting in ways users can't learn the way they learn a fixed UI How does AI context differ from conventional software context?. Evaluation is moving the same way. Researchers increasingly judge agents on the whole path of an interaction, including how well they recover from errors, not only on the final answer How should we evaluate agent behavior beyond final answers?. One way to read the comparison: Salesforce has updated the first two phases for agents, while the phases that matter most for agents that take actions still lack product-tested rules. A heuristic evaluation of nine deployed agents found the most serious lapses there. Companion and advisory agents most often broke principles like preventing user dependency and making the agent's intent visible, and those are exactly the 'over time' concerns that rules built from internal testing tend to miss Do productivity agents actually protect users better than emotional ones?.
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.
Salesforce codified lessons from building Agentforce into actionable guidelines: sub-10-character names, consistency-based heuristics, structured five-part conversations, mandatory AI disclosure, and standardized visual branding. Each rule was validated by specific usability failures the team observed.
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.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
Cocoa's system, which lets users assign plan steps to themselves or an agent and revise plans mid-execution, afforded greater steerability than chat-baseline interaction in a lab study (n=16) while maintaining ease of use, with field users valuing explicit work delegation.
Show all 9 sources
The AI UX Playground glossary names the autonomy slider as a continuous control letting users dial agent independence per task, collapsing agency levels into three adjustable points: suggest-only, draft, execute. The pattern shows shared vocabulary among designers, though validation against other control shapes remains absent.
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.
Evaluation of agentic systems shifts evidence from final responses to full interaction sequences, and scoring procedure from correctness alone to process quality, recoverability, coordination, and robustness. This pattern appears across multiple agent benchmarks as a coherent design move.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Design Principles for Human-Agent Interaction
- A Framework of User Experience Principles for Human-AI Agent Interaction in the Workplace
- Cocoa: Co-Planning and Co-Execution with AI Agents
- Levels of Autonomy for AI Agents
- Guidelines for Human-AI Interaction
- Evaluation and Benchmarking of LLM Agents: A Survey
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
- Assistant or Actor? Student Trust, Control, and Delegation Regret When Using a General-Purpose AI Agent