AI Is Getting Personal

Four of the biggest AI companies have spent five months shipping the same product: an assistant that reads your inbox before you do and tells you what matters.

On 21 October Anthropic launched Claude Duet. Every morning it sends you a short brief on what matters across your email, calendar and apps. Then it keeps checking them every few hours, on its own computer in the cloud, while you get on with your day.

Meta did the same with Muse on 8 September. OpenAI followed with dots on 29 September. Google had Gemini Spark out in May.

Most of the coverage has been a feature race over connectors and price.

I think that misses what’s actually happening. I put AI into businesses every week, and the questions these agents raise are about management. Who decides what’s urgent, and what can it do while you’re not looking?

Five Months, Five Launches

The early numbers are big. Meta told investors on 21 October that more than 10 million people in the US and Canada had set up Muse in about six weeks, and that about 25% of them use it on any given day. Zuckerberg said that’s more than twice the level Threads started at.

The demand was there before any of these launched. In January, Accenture asked 25,590 people in 16 countries, including the UK, about AI agents. 85% said they were open to working with one.

The Morning Brief Is the Real Product

These agents start without being asked.

Duet sends you a brief each morning. dots looks for things to do in the background between your conversations, read-only. Muse makes suggestions you didn’t ask for. Spark keeps working after you close your laptop.

So what’s the big change? The agent now decides what deserves your attention. It reads the inbox before you do and tells you what’s important.

That brings a new way to get things wrong. We’re all used to AI making things up. A morning brief can also leave something out. Anthropic is upfront about it: Duet “may misread a thread or miscalibrate what’s urgent”, and because it checks every few hours, urgent things can land between checks.

OpenAI has been here before. It launched ChatGPT Pulse, a daily briefing for Pro users, in September 2025, and announced in June 2026 it was shutting Pulse down within a fortnight.

Nobody has shown yet that an AI’s sense of what’s urgent matches yours. Muse’s 25% daily use shows people like the habit. None of the four has published how often the brief gets the priorities right.

If you run a team, this is the part that matters. Once people start their day from a brief, a model is filtering your organisation’s priorities, and it was trained on someone else’s idea of important. That’s a conversation for managers, and it’s worth having before the habit sets in.

Who Decides What It Can Do Without You?

This is where the companies made their real design decisions, and they’ve each drawn the line somewhere different.

Meta built a second system to police the first. It calls it Sentinel, “the sole permission authority” for what Muse can do. You can grant a permission once, for a session, for a task, for a set time or forever.

OpenAI split it by mode. When you’re not around, a dot can only read. Anything that changes something runs under rules you write, which can allow it, block it or make the dot ask.

Anthropic split it by action. Duet asks before it “sends a message, makes a payment, signs a document, or deletes anything”. And it won’t trade shares or move money between your accounts even if you approve it.

Google’s public wording is the loosest. Spark asks before “high-stakes actions like spending money or sending emails”.

If I had to explain one to a board, I’d pick Anthropic’s: it asks before four things and refuses two outright. Meta’s is the most flexible, and “perpetual” is the setting that stops the prompts for good.

So when someone on your team starts using one, the question to ask is simple. What has it been given permanent permission to do, and who said yes?

Who the Agent Works For

Who does your agent work for? The pricing tells you.

OpenAI and Anthropic charge you a subscription. Muse has a free tier, and Meta has been open about how it plans to make money from it. On the October earnings call Zuckerberg said Meta would work on monetisation “somewhat sooner on this than our normal playbook”, including fees charged to businesses on the sales Muse helps make. In his words, “we will be able to build a model where we get a percent of the value”.

Shopping is already Muse’s top use. More than 25% of users have bought something through it, with an average purchase above $200. Zuckerberg told investors “we had to work with Stripe to raise credit limits since the volumes were above what any of us expected”.

Do shoppers mind? A Harris Poll for Quad, published in April, found 75% of Americans would trust an AI agent less if its recommendations were swayed by brand money.

And how far will people actually let an agent go? Accenture’s numbers drop off a cliff once money’s involved.

If you’re in marketing, here’s what to watch. Meta is building a standard for how a business’s agent talks to a customer’s agent, with Shopify, Walmart, Stripe and Sierra. When shoppers start sending an agent to do the looking, the brands it finds will be the ones whose product and pricing data it can read.

It Remembers More Than You Think

Every one of these agents builds a memory of you, and that memory is harder to remove than the connections that fed it.

OpenAI says you can’t view, delete or edit a dot’s individual memories. The only way to clear them is to delete or reset the dot. Disconnecting an app “doesn’t erase what the dot already learned”.

Anthropic says the same about Duet: disconnecting an app stops access from then on, “but memory of past information is kept”.

Meta goes furthest the other way. You can download what Muse remembers about you and tell it to forget things. But it uses your conversations to train its models by default unless you switch that off.

Why does that matter at work? Picture someone who connects their work calendar and inbox to a personal agent for six months, then leaves. Their agent’s memory of your clients and your projects leaves with them, held by a tech company, outside any retention policy you’ve got. No offboarding checklist I’ve seen covers that yet.

One Agent, Two Inboxes

People were already mixing personal and work AI. Gartner found in May that 88% of employees with an AI tool from work also use personal AI tools for work tasks.

A personal agent turns that into standing access to a work mailbox, and I’ve gone into what that means in our post on shadow AI.

They’ve Told You It Can Be Tricked

Read the small print and the companies are refreshingly honest.

Anthropic says of Duet that “no defense is perfect”. OpenAI tells dots users to “always review consequential work”. Prompt injection means hidden instructions in something the agent reads, like an email or a web page, that try to make it do something you didn’t ask for.

Meta prices the risk. Its bug bounty pays up to $130,000 for a prompt injection that affects a single user.

The first weeks were bumpy. A security researcher found a flaw in the Muse Mac app around 21 September that Meta fixed within about a day. OpenAI disclosed on 25 September that its agents had posted 53 ChatGPT users’ images to image-hosting sites.

New software always has bugs, and this software holds your email and a payment card. The UK’s National Cyber Security Centre said the same in August: an agent “may interpret instructions and goals in literal or unexpected ways”, so don’t rely on prompting alone, and keep a way to stop it.

Britain’s in the Waiting Room

None of Muse, Duet or Spark is available in the UK yet, and UK Pro users can’t get dots. The one route open today is OpenAI’s Business Premium seats, which means the first personal agents most UK firms see will come in through company accounts. That gives you a few months to set the rules before your team gets the apps. Here’s why the UK is waiting.

What to Do Now

AI that knows your life and works for you all day is here, at least in the US.

The first agents most UK firms see will arrive on company seats. Whose rules will they run under, yours or the vendor’s?

Sources & References

Frequently Asked Questions

Roy Murphy is co-founder of G3NR8. This post draws on the companies’ own launch material, earnings calls, press coverage and consumer research from May to October 2026. Published 24 October 2026.