AI assistant UX: what Siri AI teaches about context

Good AI assistant UX depends less on how smart the model is and more on what it can see and how honestly it shows its limits. That is the conclusion of a Nielsen Norman Group review of Apple’s new Siri AI, published on October 9, 2026, which compared it with the old Siri and with ChatGPT on everyday tasks ^[1]. Its main lesson applies to any product that adds an assistant or chat feature.

We summarise what NN/g found, what Apple shipped, and the design rules we are taking into our own assistant and chatbot projects.

What Apple shipped with Siri AI

Apple announced Siri AI at WWDC on June 8, 2026 ^[2]. Apple says it can answer questions about what is on screen, use personal context to search across messages, emails, photos and more, take actions across apps, and go to the web for up-to-date information ^[2]. A dedicated Siri app lets people revisit past conversations, synced privately through iCloud ^[2].

Siri AI began rolling out as a beta in English with iOS 27 on September 14, with French, Japanese, Korean, Portuguese and Spanish due in October ^[3]. It is not available at first in the EU on iPhone, iPad and Apple Watch, and not for users under 13 ^[3].

What NN/g found when it tested Siri AI

Caleb Sponheim ran an exploratory comparison of old Siri, the new Siri AI and ChatGPT on post-trip tasks: finding photos, recalling restaurant recommendations and drafting a message ^[1]. He notes it was not a controlled study, with most requests run once or twice ^[1]. Still, the pattern was clear:

• Asked for photos from a weekend in New York, old Siri showed generic web images, ChatGPT offered news galleries, and Siri AI found the user’s own trip photos ^[1].
• Asked about breakfast spots a brother had texted, Siri AI found them and answered a follow-up correctly, while ChatGPT searched connected work accounts and found nothing relevant ^[1].
• Drafting a message, old Siri picked the wrong contact, while Siri AI prepared the right list for the right person ^[1].

The difference was not raw intelligence. Siri AI had access to the right personal data and could act on it. NN/g sums it up this way: a useful assistant needs “relevant personal context, the capability to use it, and an experience that helps users understand its limits” ^[1].

Where it still went wrong

The failures are the most useful part of the review for designers ^[1]:

• Asked what a friend said about Saturday, Siri AI missed the answer because it was in Signal, which is not available to Siri by default. The assistant did not say where it had looked ^[1].
• While drafting a personal message, a progress label read “Searching websites,” which confused rather than reassured ^[1].
• After the user confirmed a starting address for directions, the maps app opened from the current location instead, losing the detail the user had just confirmed ^[1].

None of these are model problems. They are interaction design problems: missing scope, unclear status and context that did not survive from one step to the next.

AI assistant UX rules for any product

From NN/g’s recommendations ^[1] and our own project work, these are the rules we now apply when a client adds an assistant to a product:

• Keep confirmed details. Anything the user confirmed in the conversation must carry into the final action ^[1].
• Show where you looked. When the assistant finds nothing, it should say which sources it searched, for example “I searched Messages and Mail” ^[1].
• Ask before guessing when choices lead to different actions, especially when a wrong action is costly ^[1].
• Make capabilities discoverable with suggested prompts that match what the assistant can actually reach ^[1].
• Explain, do not just narrate. A status label should tell users why a step matters, not only what the system is doing ^[1].

Context also needs structure on the builder side. A separate NN/g article on AI agents describes three roles for context: global (rules and preferences that apply everywhere), local (information for the current task) and ambient (raw streams such as email or analytics) ^[4]. Deciding which data plays which role is a design decision, and it shapes what an assistant can find.

Key takeaways

• NN/g found Siri AI more useful than old Siri and ChatGPT on personal tasks because it could reach personal context ^[1].
• Its failures came from unclear scope, confusing status labels and lost details, not from the model ^[1].
• Siri AI is in beta in English since September 14, with more languages in October ^[3].
• Assistants should show where they searched, keep confirmed details and ask before costly actions ^[1].
• Mapping global, local and ambient context is part of designing an assistant ^[4].

Quick answers

Should our app add a chat assistant? Only if it can reach data that makes it more useful than a search box. Without context, an assistant is a slower menu.

Does connecting more data always help? NN/g notes that more sources only help if they help the assistant find relevant information ^[1].

What is the quickest improvement for an existing chatbot? Rewrite its “no result” replies to state what it searched and what it can and cannot access.