Short answer: Design an AI product around the user's task, not around a chat box. Put the output where people already work, show where each answer came from, make mistakes cheap to fix, and require a human click before anything that sends, pays or deletes. Chat fits open-ended questions but is a weak default for repeated work.
Key takeaways
- Chat is one pattern among several: A button that drafts the reply inside the ticket saves a trip to a separate assistant window.
- Sources earn trust: An answer grounded in company documents should link to the exact passage it came from.
- Approval before side effects: The 2025 OWASP Top 10 for LLM applications lists Excessive Agency (LLM06) and recommends human approval for high-impact actions.
- Hebrew needs explicit direction: Mixed Hebrew and English output renders badly unless each block sets its own direction, for example with
dir="auto".
Chat or embedded actions
Chat works when the user cannot know in advance what they will ask, as in research or knowledge search. When the same request comes up dozens of times a day, such as summarizing a claim file, an embedded action is faster: one button, a prepared prompt, and the result dropped into the form field. Many products combine frequent-path actions with a side panel for everything else. The split changes the scope of AI system development, because every embedded action needs its own prompt, test cases and failure handling.
Since 2 August 2026, Article 50 of the EU AI Act requires telling people they are interacting with an AI system unless it is obvious from context, which matters for Israeli products with EU users.
Sources, uncertainty and errors
An answer that links to page 4 of the contract can be checked in seconds; one without a source asks for blind trust. In a document assistant, the citation sits inline and opens the original passage, the way a document-grounded AI agent such as DocBrain works. When retrieval finds nothing relevant, the interface should say so instead of letting the model improvise. The mechanics are explained in how RAG document search works.
Microsoft's 18 Guidelines for Human-AI Interaction, presented at CHI 2019, include supporting efficient correction among the rules for when the system is wrong. In practice: edit the output in place, regenerate one section, and let a thumbs-down ask what went wrong. Timeouts and rate limits get their own message and an automatic retry.
