How can you tell if an AI assistant is proactive?

Ask what it starts from, what it may do without you, and what it leaves behind when it gets something wrong. As of 12 October 2026, most assistants sold as proactive act on a trigger you wrote yourself: a time, an email from a sender you named, a price you chose. The few that make suggestions nobody asked for are mostly unavailable to a household in Germany.

The question has been in the air since early October, when a viral post on X and a Business Insider column described the assistant their authors want as one that notices what needs doing, and the one they keep getting as one that waits to be told¹. I go through what that would take in the main piece, can AI reduce the mental load? The vocabulary is older than the post. An industry newsletter called “proactive AI” a buzzword in October 2025².

I co-founded Subtext, an AI writing assistant that takes no action for you, on purpose. It drafts messages, and you send them. So I have a stake in the argument that some tools should leave the acting to you, and Subtext comes up under questions 2 and 4 and again in a short section near the end. For each question I say what would count as strong evidence that a claim holds, and what the warning sign is. The checklist draws on agent-governance guidance from OpenAI, the OWASP security project and the UK’s Competition and Markets Authority³⁴⁵. Product facts come from vendor documentation checked on 12 and 13 October 2026, plus press reports where marked.

Proactive vs reactive AI: what is the difference?

Assistants come in four kinds, and marketing calls the last three proactive. A reactive assistant answers when you ask. A persistent-context assistant remembers what you said last month and still waits to be asked. A scheduled or event-triggered system starts a workflow you defined when a time, message or price condition is met. An inference-driven system decides on its own that something may need your attention.

Only the fourth kind touches what household researchers call anticipating: noticing a need before anyone points at it. Allison Daminger interviewed members of 35 couples: college-educated parents of young children, mostly in different-sex couples. Wives carried more of the anticipating and monitoring, while deciding was shared roughly equally⁶. A tool that waits for a prompt starts after the noticing. The prompt is evidence that you already did it.

1. Initiation: what is a proactive AI agent?

In practice, usually one that runs a rule you set and contacts you when its condition is met. Strong evidence is a product that detects a defined change (an email, a date, a price, a stock level) without a fresh prompt, under rules the household chose.

ChatGPT’s scheduled tasks run on a timetable, watch for a change, or start when a Gmail message arrives from a sender or subject you filter for; the Gmail, Slack and GitHub triggers need Plus or a higher plan⁷. Gemini Spark schedules work by time, Gmail filter or monitored topic, with up to 50 active schedules⁸. Both start from conditions you wrote.

Inference is rarer. Meta says its Muse agent makes suggestions without being asked and keeps working after you close the app⁹. Amazon describes Alexa+ in the US as anticipating needs and offering suggestions of its own¹⁰. OpenAI’s Dots can run background research on their own, read-only. While doing it they cannot send messages, change app content or control the browser¹¹. OpenAI is also retiring its earlier briefing product: “Pulse is being sunset as proactive updates move into scheduled tasks” (17 June 2026)¹².

The warning sign is a daily digest of things you still have to act on. That is a reminder service with better manners.

2. Context: how do you safely give an AI agent access to Gmail and your calendar?

Start narrow. Strong evidence is context with a source attached: which email created this calendar entry, when it was last checked, whose account it came from. The warning sign is a long memory marketed without provenance, or one person’s inbox treated as the household’s.

Freshness is a documented limit. Google says Spark’s monitors are “not ideal for tracking fast-moving data or performing time-critical tasks”⁸.

The narrowest permission is read-only, which is how Dots do their background research¹¹. Anthropic’s Gmail and Google Calendar connectors for Claude can send, reply to and forward email and create, update and delete events, and by default Claude asks for your approval before each of these actions¹³. (Subtext runs on Anthropic’s models, so weigh what I say about Claude with that in mind.) If you grant write access, I’d limit it to one shared calendar and one filtered sender, such as the school’s address, and widen it only after a few weeks of checking its work.

3. Authority: can AI agents send emails, make phone calls or buy things?

Some can, within limits that differ by product and country. Strong evidence is authority scoped by task, merchant, budget and risk, with payments and passwords handed back to you. The warning sign is a single “allow all” switch.

Email. Google documents that Spark can draft, reply to and forward Gmail, and that it asks before sending¹⁴. Microsoft’s Copilot Cowork for personal accounts, in Preview for eligible Microsoft 365 subscriptions, drafts and sends email and asks for approval, or hands back control, for sensitive actions¹⁵. Meta says Muse checks before sending an email⁹. OpenAI’s Dots help page has the clearest line: “Asking for a draft doesn’t authorize sending it”¹⁶.

Phone calls. Google’s “Call for me” phones businesses to book or change appointments: US only, on a Pixel 11 with a US SIM, after you tap to start, and Google says Gemini opens every call by saying it is an AI assistant¹⁷.

Purchases. In the US, Alexa+ can watch a named item and buy it at your target price with your default address and payment method, “even at 3 a.m. when you’re sleeping”¹⁰. Amazon’s page describes no confirmation step before the purchase¹⁰. Its German feature list mentions a notification when a price changes and an order on Amazon.de when you ask for one, with no automatic purchase¹⁸. Dots sit at the other end. Purchases need your approval, and password changes and money transfers stay with you¹⁹.

Two settings move confirmation from the moment of action to the moment of set-up: ChatGPT’s “Allow all actions” permission for apps²⁰, and Microsoft’s option to approve a scheduled Copilot task at creation so it runs unattended²¹. Both mean you decided once, possibly weeks before the action.

4. Confirmation: when should an AI assistant act on its own, and when should it ask first?

When your judgement is needed, and at no other time. An agent that asks before every routine step hands the work back. Seven chores that each trigger two approval prompts have become fourteen interruptions. PersonalAgentBench, a benchmark from the vendor micro1, counts over-confirmation of already authorised actions as a cost to the user²².

Confirmations that fail or get skipped are worse, and both have happened. In 2018 a Portland family’s Echo sent a recording of a private conversation to a contact; Amazon said background speech was misheard as the wake word, “send message”, a contact’s name and finally “right” in answer to the confirmation prompt²³.

In February 2025 a Washington Post columnist asked OpenAI’s Operator to find the cheapest dozen eggs for delivery. It bought eggs through Instacart for $31.43 and reported $19.68, and OpenAI said it was examining why Operator occasionally did not send confirmations²⁴. In February 2026 a user told her OpenClaw agent to suggest inbox deletions and wait for her go-ahead; on her large real inbox it began mass-deleting. That one is a user anecdote²⁵.

Strong evidence is risk-based escalation: a clear request before something consequential, quiet otherwise. Lisanne Bainbridge described the trap in 1983, writing about industrial process control: designers automate what they can and leave people the remainder, and “it is humanly impossible to carry out the basic function of monitoring for unlikely abnormalities”²⁶. As I see it, a household that supervises an agent has taken on that job. The other design choice is to take no action at all, which is the one Subtext makes.

5. Error recovery: can AI agents go rogue?

The documented problems are more ordinary than “rogue”. An agent misreads its task, loses an instruction, or reports a success it did not achieve. Strong evidence is a product that detects failure, avoids duplicates, reverses what it can and says so. The warning sign is success reported without a check of the real outcome.

Vendors say more about preventing mistakes than about undoing them. Spark’s confirmations for Workspace actions link to the item, so you can reverse the change yourself¹⁴, and OpenAI shows a dot’s work as in progress, scheduled and completed¹⁶. I found no vendor that documents a general rollback of actions taken in your accounts, and no public incident of an agent double-booking a family calendar, though the risk is easy to foresee. The OpenClaw user could not stop the deletion from her phone²⁵. So ask how you stop it, from where, and how fast.

6. Measurement: has anyone shown that it helps?

Nobody has shown it. As of 12 October 2026, I found no peer-reviewed or preregistered study that measured household mental load before and after people adopted an AI assistant. The two small studies nearest the question, an 11-family planning pilot²⁷ and a two-week test with five mothers using ChatGPT²⁸, had no control group. In the pilot, over half the families stopped carrying out the subtasks the system assigned them²⁷, and in the ChatGPT test the tool rarely moved mental labour from one person to another²⁸. The main piece has the details.

Lab benchmarks use simulated users and tasks, so they say little about a real home, and even there the scores are low. In the Claw-Anything preprint, which simulates months of one user’s digital life, the best model passed 34.5% of tasks on the first attempt, and the paper reports that proactive tasks were much harder than reactive ones²⁹. Press coverage put the first-attempt pass rate at 6.7% on proactive tasks against 25.9% on reactive ones³⁰. Unasked help also has a cost. In a vignette experiment with 92 US participants who had programming experience, help the AI offered on its own led to lower satisfaction with the system than help given on request, and the effect ran through lower competence-based self-esteem³¹. That was coding help, so it may not transfer to a forgotten permission slip.

Strong evidence would be independent, long-running measures of how often people notice, decide, check and correct, including partner reports. Testimonials and self-reported hours are warning signs; Skylight’s promise of “10 hours back (or more) every month” states no method, so I read it as marketing³². The simplest test I know is one question. After the first week of set-up, do you think about the responsibility less? If one person in your home still holds it, how to explain the mental load to your partner may be the more useful read.

7. Accountability: is ChatGPT agent safe?

That product has gone. OpenAI’s help page says “ChatGPT agent is no longer available” and names ChatGPT Work and its cloud browser as replacements³³. The cloud browser is designed to ask before actions that could be hard to reverse, such as a booking or payment, and OpenAI says its safeguards against prompt injection do not remove every risk³⁴.

Prompt injection means instructions hidden inside content the assistant reads, and it matters at home. In August 2025 SafeBreach showed that text hidden in a Google Calendar invite could make Gemini delete events once the user asked about their schedule and, a reply later, operate Google Home devices such as windows and a boiler. Its write-up lists leaked email among the results³⁵.

In January 2026 Miggo Security showed an invite that made Gemini summarise private meetings into an event the attacker could see, then give the user a harmless reply³⁶. In August 2025 Brave showed hidden text in a Reddit comment getting Perplexity’s Comet to post a user’s email address and one-time password publicly, and later said the attack class was not fully mitigated³⁷. OpenAI says prompt injection is unlikely ever to be fully solved³⁸.

So the last question is who can reconstruct what happened. Strong evidence is a time-stamped log of actions, sources, permissions and the household member who authorised each one. Meta claims Muse shows a “complete audit trail”⁹; Perplexity lists real-time audit logs and says Comet needs confirmation to send email or place an order³⁹. No vendor I checked documents which household member authorised an action. An IAPP column of 2 October 2026 reports the chair of the US Federal Trade Commission as saying on 25 September that “audit trails tell the real story”⁴⁰. Since the agent’s own report was wrong in the egg and Miggo cases, a chat transcript cannot be the audit trail.

Is there any danger in adding this technology to the household?

Yes, and the research points especially to the people who did not set it up. In a 2019 study of 18 people in multi-user smart homes, whoever installed the devices held outsized control⁴¹. In a month-long deployment in 7 Seattle-area households (19 participants), most quickly switched to the least restrictive settings; only roommates kept restrictions⁴². A New York study of 39 survivors of intimate partner violence and 50 professionals described abusers as authenticated but adversarial users of a victim’s accounts⁴³. These studies are small and predate AI agents. Applied to agents, they suggest to me that activity logs and memories can serve surveillance as easily as accountability. If someone at home uses shared accounts or an assistant’s history to watch you, treat it as a safety problem. Helplines are listed under “When raising it isn’t safe” in how to ask your partner to share the mental load.

In the EU, a family using an assistant for household purposes may fall outside the GDPR, while the company providing it does not⁴⁴, and consent-based online services offered directly to a child need parental consent below 16, which member states may lower to no less than 13⁴⁵. Before connecting a shared calendar, ask whose data it holds and who can read the log. Calendars matter twice over, because anyone who can send you an invite can put text in front of your assistant.

The products side by side, and what works in Germany

Of the products I checked, a Berlin household can use ChatGPT’s scheduled and Gmail-triggered tasks and its cloud browser on Plus and higher plans⁷³⁴ and Claude’s scheduled tasks⁴⁶, going by help pages that state no EU restriction, plus Alexa+ in Early Access⁴⁷. It cannot use Dots on the Pro plan¹⁶, Gemini Spark¹⁴, Meta Muse⁴⁸ or Siri AI on an iPhone⁴⁹.

Where Subtext fits, and when not to use it

On this checklist Subtext has no initiation and no authority, by design. It is an AI writing assistant that helps you draft, rewrite, proofread and understand messages⁵⁴. You paste or upload what you want help with, as text, a screenshot or a voice note, and it suggests up to three versions, each marked safe to send or with a warning based on Subtext’s own estimate⁵⁵. In its own words: “Nothing is sent for you either. You copy the version you want and send it yourself”⁵⁵. It connects to none of your email, calendar or messaging accounts⁵⁵⁵⁴.

That makes the confirmation question easy, because every send is yours. The measurement question is less comfortable. Subtext’s research page says “No study on this page tested Subtext”, and that its tags, feeling word and safe-to-send estimate “have not been measured against how real readers took real messages”⁵⁶.

Use Subtext when one message matters and you want to stay the sender. Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

Skip it if your problem is volume (the best AI reply app depends on whether your problem is volume or stakes) or remembering to reply at all (the best apps for ADHD adults who take forever to reply), or if you want something done for you. For the messages that make up part of a family’s load, see why family messages feel like a second job. For what AI can take off a household at all, the main piece is can AI reduce the mental load?

Questions people ask

What are some proactive AI examples? ChatGPT tasks that start when an email arrives from a sender you chose⁷, Gemini Spark’s topic monitors⁸, and Meta’s claim that Muse makes unprompted suggestions in the US and Canada⁹⁴⁸.

Can ChatGPT manage my calendar? Partly. OpenAI documents creating, updating, RSVPing to and cancelling events on supported shared Outlook calendars, mainly for workplace accounts. I could not confirm every action for personal Outlook.com or Google calendars⁵⁷.

Can AI agents log in to websites? Some can. ChatGPT’s cloud browser handles some login-required sites on Plus and Pro plans; you type the password yourself¹². Spark’s Chrome integration uses saved passwords, US only⁵⁸.

Can AI agents make phone calls? In the US, on a Pixel 11 with a US SIM, Gemini’s “Call for me” phones businesses after you tap to start, and announces itself as an AI. It is not available in Germany¹⁷.

Can AI book flights? Spark’s Chrome integration can start flight bookings in the US and hands payment back to you⁵⁸. For restaurants, ChatGPT books through OpenTable once you pick a slot¹².

Can AI agents buy things for you? Yes, with different safeguards. Alexa+ buys at a target price with no confirmation described, in the US¹⁰; Dots need your approval¹⁹.

Should you let AI manage your email? Start with read-only access and drafts, then add sending rights for one narrow task at a time. Claude’s Gmail connector asks for your approval before each send by default¹³.

Is ChatGPT Pulse going away? Yes. OpenAI’s release note of 17 June 2026 says Pulse is being sunset as proactive updates move into scheduled tasks¹².

Is there an AI that does everything for you? No product documents that. Instinct’s homepage claims bill disputes, gifts, groceries and doctor’s appointments, with no information on approvals, pricing or regions⁵⁹.

Does Subtext send the message for me? No. You copy the version you want and send it from your own app⁵⁵.

Sources

Numbered in order of first appearance. Pages were checked on 12 or 13 October 2026, as each entry says, and help pages change often.

1. I don’t need an AI agent to be an executive assistant. I need it to be a wife. Katie Notopoulos, Business Insider (syndicated by Yahoo Style Canada). 9 October 2026. . Read in the syndicated copy. Checked 12 October 2026.
2. Welcome to the Proactive Bandwagon. Babbage Insight (Substack newsletter). 9 October 2025. . Industry newsletter. Checked 12 October 2026.
3. Practices for Governing Agentic AI Systems. Yonadav Shavit and colleagues, OpenAI. Undated (references accessed to late 2023). . White paper without peer review; vendor-authored. Checked 12 October 2026.
4. LLM06:2025 Excessive Agency. OWASP GenAI Security Project. 2025. . Security guidance. Checked 12 October 2026.
5. Agentic AI and consumers. UK Competition and Markets Authority. 9 March 2026. . Regulator research paper. Checked 12 October 2026.
6. The Cognitive Dimension of Household Labor. Allison Daminger, American Sociological Review 84(4): 609 to 633. Online 9 July 2019. . Peer reviewed; qualitative, 70 interviews with members of 35 couples; wife did more in 26 of 32 different-sex couples. Checked 12 October 2026.
7. Scheduled tasks in ChatGPT. OpenAI Help Center. Updated about 2 October 2026. . Vendor page. Checked 12 October 2026.
8. Create & manage schedules for tasks in Gemini Spark. Google, Gemini Apps Help. Undated. . Vendor page. Checked 12 October 2026.
9. Introducing Muse: The World’s First Personal AI Agent Built for Everyone. Meta Newsroom. 8 September 2026, modified 30 September 2026. . Vendor page; “World’s First” is a marketing claim. Checked 12 October 2026.
10. Getting started with Alexa+: Simplify home organization with recipes, calendars, and reminders. Amazon (About Amazon). 4 February 2026. . Vendor page. Checked 12 October 2026.
11. Introducing dots. OpenAI. 29 September 2026. . Vendor page. Checked 12 October 2026.
12. ChatGPT release notes. OpenAI Help Center. Updated about 9 October 2026. . Vendor page; entries of 17 June, 10 August and 25 August 2026 used. Checked 12 October 2026.
13. Use Google Workspace connectors. Anthropic, Claude Help Center. Undated. . Vendor page; Anthropic is a processor for Subtext. Checked 13 October 2026.
14. Use Gemini Spark to manage your tasks & workflows in Gemini Apps. Google, Gemini Apps Help. Undated. . Vendor page. Checked 12 October 2026.
15. Get started with Microsoft Copilot Cowork for personal accounts (Preview). Microsoft Support. Updated 17 September 2026. . Vendor page. Checked 12 October 2026.
16. Getting started with your dot. OpenAI Help Center. Updated about 11 to 12 October 2026. . Vendor page. Checked 12 October 2026.
17. Ask Gemini to handle your everyday phone calls. Google, Gemini Apps Help. Undated. . Vendor page. Checked 12 October 2026.
18. 25 Dinge, die Du mit Alexa+ ausprobieren kannst. Amazon (About Amazon Deutschland). 7 May 2026. . Vendor page. Checked 13 October 2026.
19. How we build safety, security, and privacy into dots. OpenAI. 29 September 2026. . Vendor page. Checked 12 October 2026.
20. Connected apps in ChatGPT. OpenAI Help Center. Updated 12 October 2026. . Vendor page. Checked 13 October 2026.
21. Using Copilot Tasks. Microsoft Support. Updated 6 October 2026. . Vendor page. Checked 12 October 2026.
22. PersonalAgentBench. micro1. Undated. . Vendor benchmark, provisional; six assistants and 20 questions, first eligible run for each. Checked 13 October 2026.
23. Woman says her Amazon device recorded private conversation, sent it out to random contact. Gary Horcher, KIRO 7. 23 May 2018. . Journalism with vendor statement; the page is blocked outside the US, so I read the archived copy at https://web.archive.org/web/20210609120825/https://www.kiro7.com/news/local/woman-says-her-amazon-device-recorded-private-conversation-sent-it-out-to-random-contact/755507974/. Checked 13 October 2026.
24. I let ChatGPT’s new ‘agent’ manage my life. It spent $31 on a dozen eggs. Geoffrey A. Fowler, The Washington Post (syndicated by Anchorage Daily News). 7 February 2025. . Journalism; read via syndicated copy. Checked 12 October 2026.
25. A quote from Summer Yue. Simon Willison (weblog). 23 February 2026. . User anecdote; no vendor response found. Checked 12 October 2026.
26. Ironies of automation. Lisanne Bainbridge, Automatica 19(6): 775 to 779. November 1983. . Argument paper on industrial process control; quote from p. 776. Checked 13 October 2026.
27. ParPal: A Human-Centred AI System for Multi-Actor Planning and Collaboration in Family Learning. Chen and colleagues. 28 January 2026. . Preprint; 11 families, no control group, no validated load measure. Checked 12 October 2026.
28. Invisible household labour and digital solutions: Gender divisions in mental labour and the role of artificial intelligence. Ksenija Klasnić, Anita Dremel, Krešimir Žažar and Đurđica Degač, conference proceedings, pp. 137 to 154. 2025. . Five-person quasi-experiment, no control group; review process not stated; funded by the University of Zagreb. Checked 12 October 2026.
29. Claw-Anything: Benchmarking Always-On Personal Assistants with Broader Access to User’s Digital World. Yusong Lin and colleagues. 25 May 2026. . Preprint; lab conditions, simulated user. Checked 12 October 2026.
30. Huawei’s New Benchmark Gives AI Agents Months of Your Life, Then Watches Them Fail. Jose Antonio Lanz, Decrypt. 27 May 2026. . Journalism; reactive and proactive split not found in the paper’s abstract. Checked 12 October 2026.
31. When AI-Based Agents Are Proactive: Implications for Competence and System Satisfaction in Human-AI Collaboration. Diebel, Goutier, Adam and Benlian, Business & Information Systems Engineering 68(2): 259 to 278. Online 3 January 2025. . Peer reviewed; vignette experiment, 92 US participants recruited through Prolific; German Research Foundation funding per the article. Checked 12 October 2026.
32. Meet Sidekick, powerful planning for busy families. Skylight. Undated. . Vendor marketing; no method stated. Checked 12 October 2026.
33. ChatGPT agent. OpenAI Help Center. Updated about 15 September 2026. . Vendor page. Checked 12 October 2026.
34. Using cloud browser in ChatGPT. OpenAI Help Center. Updated 28 August 2026. . Vendor page. Checked 12 October 2026.
35. Invitation Is All You Need: Invoking Gemini for Workspace Agents with a Simple Google Calendar Invite. Or Yair, Ben Nassi and Stav Cohen, SafeBreach Labs. 6 August 2025. . Researcher write-up. Checked 12 October 2026.
36. Weaponizing Calendar Invites: A Semantic Attack on Google Gemini. Liad Eliyahu, Miggo Security. 19 January 2026. . Researcher write-up. Checked 12 October 2026.
37. Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet. Artem Chaikin and Shivan Kaul Sahib, Brave. 20 August 2025. . Researcher write-up by a competing browser maker. Checked 12 October 2026.
38. Continuously hardening ChatGPT Atlas against prompt injection attacks. OpenAI. 22 December 2025. . Vendor page. Checked 12 October 2026.
39. How we engineer safer agents. Perplexity. 29 September 2026. . Vendor page. Checked 12 October 2026.
40. A view from DC: The FTC says your company’s agents are your problem. Cobun Zweifel-Keegan, IAPP. 2 October 2026. . Opinion column reporting a speech; no FTC transcript found. Checked 12 October 2026.
41. Who’s In Control? Interactions In Multi-User Smart Homes. Christine Geeng and Franziska Roesner, CHI 2019. 2019. . Peer reviewed; 18 participants; abstract and bibliographic record read. Checked 12 October 2026.
42. Understanding and Improving Security and Privacy in Multi-User Smart Homes. Eric Zeng and Franziska Roesner, 28th USENIX Security Symposium. August 2019. . Peer reviewed; 7 households, 19 participants, self-selected. Checked 12 October 2026.
43. A Stalker’s Paradise: How Intimate Partner Abusers Exploit Technology. Diana Freed, Jackeline Palmer, Diana Minchala, Karen Levy, Thomas Ristenpart and Nicola Dell, CHI 2018. 2018. . Peer reviewed; 39 survivors and 50 professionals in New York. Checked 12 October 2026.
44. Recital 18: Not Applicable to Personal or Household Activities. gdpr-info.eu (unofficial reproduction of Regulation (EU) 2016/679). Accessed 12 October 2026. . Checked 12 October 2026.
45. Art. 8 GDPR: Conditions applicable to child’s consent in relation to information society services. gdpr-info.eu (unofficial reproduction of Regulation (EU) 2016/679). Accessed 12 October 2026. https://gdpr-info.eu/art-8-gdpr/. Checked 12 October 2026.
46. Schedule recurring tasks in Claude Cowork. Anthropic, Claude Help Center. Undated. . Vendor page. Checked 12 October 2026.
47. Alexa+ startet in Deutschland. Amazon (About Amazon Deutschland). 7 May 2026. . Vendor page; I could not open Amazon’s German help pages. Checked 12 October 2026.
48. Meta Muse kommt aufs iPad, nach Deutschland aber weiterhin nicht. Jan Gruber, Notebookcheck. 8 October 2026. . Journalism. Checked 12 October 2026.
49. How to get Siri AI. Apple Support. 22 September 2026. . Vendor page. Checked 12 October 2026.
50. Alexa+ now available to everyone in the US. Amazon (About Amazon). 4 February 2026, updated 1 September 2026. . Vendor page. Checked 12 October 2026.
51. Alexa+ arrives in India as Amazon’s next-generation AI assistant expands internationally. Amazon (About Amazon). 19 March 2026, modified 16 September 2026. . Vendor page; still lists Germany as Early Access. Checked 12 October 2026.
52. Computer adds Automations for ongoing work. Perplexity. 29 September 2026. . Vendor page. Checked 12 October 2026.
53. Get started with Claude in Chrome. Anthropic, Claude Help Center. Undated. . Vendor page. Checked 12 October 2026.
54. Terms, Privacy and your account. Subtext (Anton Licht, sole proprietor). Terms updated 26 July 2026; Privacy Policy updated 10 October 2026. . Our own page. Checked 12 October 2026.
55. Subtext app: AI writing assistant for texts and replies. Subtext. Undated. . Our own page; first-party description. Checked 12 October 2026.
56. The research behind Subtext. Subtext. Undated. . Our own page. Checked 12 October 2026.
57. Outlook Email and Calendar apps in ChatGPT. OpenAI Help Center. Updated 12 October 2026. . Vendor page; written mainly for workspaces. Checked 13 October 2026.
58. Gemini Spark now integrates with Chrome. Google (The Keyword blog). 30 July 2026. . Vendor page. Checked 12 October 2026.
59. Instinct is a personal assistant that understands what you’re working on and what’s important to you. Instinct (Spear Street Technology, Inc.). Undated. . Vendor marketing; no approvals, pricing or regions stated. Checked 12 October 2026.