No always-on personal AI agent from Meta, Google, OpenAI or Anthropic is open to UK consumers on a personal plan today - and one of the five in this comparison has even changed its name since the headlines. This guide compares Muse, Instinct, Gemini Spark (now Task), dots and Claude Duet on what a UK reader can actually get, what each costs in pounds where a pound price exists, what each asks before acting and what to use while you wait.
The short version: Claude Duet, Gemini Task, Meta Muse and OpenAI dots are closed to UK consumers for now, and Instinct is an invite-only beta with no UK statement either way. The realistic step today is a small job with a supervised agent - and two supervised agents, ChatGPT Work and Claude Cowork, genuinely are available here, in pounds. The one always-on exception is corporate: dots can reach the UK through a business workspace, which is not a consumer purchase.
Last checked: 5 November 2026, against the makers' help pages, pricing pages and terms, plus named press reports. Nothing here was tested hands-on and none of it was tried on a UK account. One disclosure: this guide was drafted with Claude, made by Anthropic, which also makes two of the products covered - Claude Cowork and Claude Duet. We update this guide when availability or prices change.
The best AI agent for UK users right now: the honest answer
We are not naming a single winner - the published evidence does not separate the top two, and we tested neither. Instead the field sorts by what a UK consumer can get today. Within each group the order is alphabetical, not a ranking.
| Group | Products | What it means for you |
|---|---|---|
| Usable now, with supervision | ChatGPT Work, Claude Cowork | Both take real steps. Both have UK prices in pounds. Cowork does not stop at every step unless you choose Manual |
| Usable, but narrow | Alexa+, Gemini app and Gemini in Chrome, Microsoft Copilot | Built for chat, summaries or small actions. Alexa+ is in request-based Early Access |
| Unclear | Instinct, Manus, Perplexity | No UK statement found. Instinct is invite-only |
| Not open to UK consumers | Claude Duet, Gemini Task, Meta Muse, OpenAI dots | Excluded by name, limited to other countries or not on any personal UK plan |
One naming note before anything else, because it will trip you up in search: the Google agent announced as Gemini Spark in May was renamed. Google's own help FAQ - under the heading "Where did Spark go?" - says "we renamed it Task", and it now lives as tasks in Gemini. Older guides, and this article's title, use the Spark name people still search for; the product is Task.
How we compared them
UK access comes first: does the maker say a UK consumer can have it, exclude the UK or say nothing? A product you cannot get sits below one you can. Then real steps (does it act, or only talk?), approvals and limits (can you make it ask first and take access away?), price clarity (published, in pounds, and does the plan actually include the agent?) and track record (what named reviewers report, and what the maker says about your data). We used makers' own pages and named press reports, not vendor round-ups; opinions are attributed to the people who hold them. One fairness note: Duet launched on 21 October and Muse on 8 September, so a shorter list of problems does not show a product is safer.
What makes an AI agent different from a chatbot?
A chatbot answers; an agent acts, taking steps for you in apps and websites - our guide to what an AI agent is covers the basics. Four labels matter from here on. Chat: it answers and writes inside the conversation. Assistant: it reads connected apps and runs small jobs. Supervised agent: you start a job, it uses a browser, apps or files and takes real steps, with approvals. Always-on agent: it keeps running, remembers you and can start work itself. Most of the serious ones run on a cloud computer - a separate machine online - so the job carries on with your laptop closed.
Personal AI agents compared: which can you use in the UK?
Alphabetical, with every price as read on the makers' own pages on 5 November 2026. Where a price is in dollars, the maker publishes no UK price; tax treatment varies - Anthropic's dollar list prices exclude tax, while its UK pricing page includes VAT.
| Product | What it is | UK consumer access | Listed price | Main catch |
|---|---|---|---|---|
| Alexa+ (Amazon) | Voice assistant with narrow real actions, such as restaurant bookings | Request-based Early Access | Free in Early Access, then free with Prime or £19.99 a month | Early Access gates who gets it and when |
| ChatGPT Work (OpenAI) | Supervised agent mode for longer, multi-step work and finished deliverables | Yes - limited on the free desktop plan, expanded on Plus | Plus £20 a month; Go £7; Pro from £89 | It is a mode inside ChatGPT, not a separate product |
| Claude Cowork (Anthropic) | Supervised agent using files and a browser, with scheduled tasks | Yes, on every paid Claude plan on desktop (web and mobile from Pro up) | Pro £15 a month billed annually, £18 monthly, VAT included | Defaults to Auto mode, which does not stop at every step. No free plan |
| Claude Duet (Anthropic) | Always-on agent in its own app | No - US-only research preview: paid plan, US account, waitlist invitation | No extra charge during the preview | Nothing promised on price after the preview |
| Gemini app + Gemini in Chrome (Google) | Chatbot and assistant working with Google apps | Yes - Gemini in Chrome reached UK desktops on 14 July 2026 | Free; AI Plus £4.49, AI Pro £18.99, AI Ultra from £79.99 a month | The agentic features sit elsewhere - see Task below |
| Gemini Task, formerly Spark (Google) | Always-on agent | No - Google excludes the UK by name, on every tier | Not sold here; elsewhere it needs AI Pro (US) or Ultra | A UK AI Pro plan does not include it |
| Instinct (Spear Street Technology) | Agent you text or phone | Invite-only beta; the sign-up form accepts +44 numbers but Instinct has made no UK statement | No published pricing; reportedly free in beta | No confirmation promise in its published terms |
| Meta Muse | Always-on agent, also in WhatsApp | No - live in the US, Canada, Japan and Singapore, no UK date | Elsewhere: free with a limit, Power $20, Max $100 a month | No UK launch or local connectors announced |
| Microsoft Copilot | Mainly a chatbot in consumer plans | Yes | Free; Microsoft 365 Premium £18.99 a month or £189.99 a year | "Agentic" in the plan listing means research and analysis |
| OpenAI dots | Always-on agent inside ChatGPT | Not on personal plans - OpenAI excludes the UK by name from dots on Pro | Consumer Pro from £89 a month in the UK, without a dot; the UK route is a business workspace | The business route needs two seats minimum |
The five head to head on paper
Since nobody outside these companies has run a fair side-by-side test, the honest comparison is specification against specification - all from the makers' own pages:
| Muse | Instinct | Gemini Task | dots | Claude Duet | |
|---|---|---|---|---|---|
| How you reach it | App, web, Mac, WhatsApp | Text or phone call | Gemini app, Mac, web | ChatGPT apps; Slack and Teams per launch reports | Own app + claude.ai |
| Runs where | Dedicated cloud VM per user | Its own phone and computer, per Instinct | Cloud, with a remote browser | Own cloud computer per dot | Own cloud computer + browser |
| Free tier | Yes, usage-limited | Beta reportedly free | No - paid Gemini plans | No - Pro or business seats | No - paid Claude plans |
| Payment design | One-time cards via Link; audit trail | No published detail | Hands payment steps back to you | Per-rule approvals across four behaviour levels | One-time virtual card per purchase via Link or Shop Pay |
| Hard limits stated | Sentinel gates all internet access | None published | 15 concurrent tasks | Four behaviour levels per action type | No securities trading, no money transfers, no third-party accounts - even with approval |
| Can you watch it work? | Yes - Meta's pages describe watching and taking over the browser | No published equivalent | Not stated | Not stated | No - you cannot watch or take over its browser |
Read the bottom two rows together and the design philosophies separate: Meta gives you a window plus its Sentinel watchdog, Anthropic gives you hard walls and no window, OpenAI gives you a rulebook, Google hands the risky steps back to you and Instinct publishes the least of the five.
Does a UK subscription include the agent? Not always
You can pay a company for an AI plan in the UK and still not get its agent. Google AI Pro is sold here at £18.99 a month, but it does not get you Gemini Task. UK users can buy ChatGPT Pro from £89 a month - but OpenAI's exclusion covers dots on consumer plans in the UK, so Pro here comes without one. And a UK Claude plan does not get you Duet, which needs a US account on top. Before you pay for any plan hoping for the agent, search the maker's help pages for the agent's name and "United Kingdom" - an exclusion from a feature is not an exclusion from the plan, and the reverse holds too.
The one genuine UK route to an always-on agent is corporate: OpenAI's ChatGPT Business workspaces list "Dots, always-on agents for Premium seats" at £75 a month per Premium seat billed annually (£90 monthly), with workspaces running 2 to 200 seats - so a realistic minimum around £90 to £150 a month across two seats, as listed when we checked on 5 November. OpenAI's help pages describe dots reaching eligible Business Premium users in supported regions, with Enterprise following from 18 November - eligibility language worth re-reading before anyone buys seats for this. That is a business product with business terms, not a consumer purchase, and we would not buy a two-seat workspace to get a personal dot.
The two supervised agents you can actually use today
Claude Cowork has the firmer published evidence on UK access - it is included on all paid Claude plans, and Claude's UK pricing page quotes pounds. ChatGPT Work shows UK prices too and names no UK exclusion, where OpenAI does name the UK when it excludes it (as with dots). We have seen no head-to-head test of the two, so if you already pay for one, start there.
ChatGPT Work
OpenAI describes Work as "an agent designed for longer, multi-step work and finished deliverables" - it is a mode inside ChatGPT, alongside Chat and Codex, rather than a separate product. Desktop access exists on every plan ("Limited ChatGPT Work access on desktop" is the free plan's own line), with expanded access on desktop, web and mobile from Plus at £20 a month, where scheduled tasks also live. OpenAI pitches approvals as part of the design, though it does not clearly publish Work's default approval mode - treat that as a settings check on day one. On data: OpenAI's connected-apps pages say it does not train generalized models on data from connected Google apps even when the improve-the-model setting is on, with exceptions for feedback you submit, content you paste in and content that appears in responses - that last one is the gotcha worth knowing.
Claude Cowork
Anthropic's page is blunt about what it is for: "Cowork is where you hand Claude real work." It works in folders and tools you choose, has a built-in browser in a side panel and can "schedule a task for any cadence, and it runs unattended". Desktop covers all paid plans; web and mobile cover Pro, Max, Ultra and Team. The one setting that matters: Cowork defaults to Auto on Pro, Max, Ultra and Team - Anthropic's words are that in Auto "Claude keeps working without stopping to ask about every step"; its help page adds that read-only tools run automatically while Claude decides whether to go ahead with tools marked as needing approval. Enterprise defaults to Manual, which stops for those tools. Switch to Manual before your first job. On data: Anthropic does not train on Gmail, Calendar or Drive connector data, with the same paste-and-response exception as OpenAI; retrieved connector data is stored with the chat and deleted with it. Anthropic's own warnings apply - its Chrome browsing tool "is still risky", in its words.
Claude Duet: praised, previewed and not for sale here
Duet is Anthropic's always-on agent in its own app: a morning brief from your connected accounts, background checks "every few hours", delegated tasks with approval gates before sending, paying, signing or deleting, and hard blocks on trading securities, moving money between accounts and accessing someone else's account even with approval. It is a US-only research preview behind a waitlist, needs a US Claude account, costs nothing extra on Pro, Max or Ultra during the preview and promises nothing afterwards. Reviewers rate it the most polished of the new agents - Claire Vo's words - while reporting slow parallel tasks and missed priorities; and personal errands are the weaker side, by Anthropic's own account. Our Claude Duet guide has the full picture, including the approval model and what Anthropic can see.
Gemini Task (formerly Spark): excluded by name
Task is Google's always-on agent: goals once or on a schedule, using connected apps and signed-in sites, up to 15 concurrent tasks. The UK position is the bluntest of the five - Google's help page says it is available "wherever Gemini Apps are supported, except in the European Economic Area, Nigeria, Switzerland, and the United Kingdom", with no timetable. It also needs a personal Google Account (work and school accounts are excluded), Keep Activity on and an opt-in to Personal Intelligence, and it reached AI Pro plans in the US only, with Ultra elsewhere. For sign-ins and payment details, Google says the task pauses and hands control back to you. If your life runs on Google in the UK, Gemini in Chrome is what you can actually have - an assistant, not an agent.
Instinct: the wildcard in the queue
Instinct, from Spear Street Technology, is the agent you text or call - no app, an assigned number, its own email address for dealing with businesses (per TechCrunch) and group chats arriving now. It is invite-only, reportedly free in beta, with no published pricing and no UK statement in either direction: its terms place operations in the US, its sign-up form accepts +44 numbers, and nothing in between is confirmed. Two things set it apart in this comparison - the interface, and the fact that its published terms contain no confirmation-before-spending promise at all, where Duet and Muse both make one. Our Instinct guide covers the invite mechanics, the terms and the early-user reports.
Meta Muse: four countries, none of them this one
Muse is the biggest consumer launch of the five by third-party download estimates: free tier, WhatsApp access, a dedicated cloud VM per user with what Meta describes as a separate Sentinel agent - a system-isolated watchdog that must approve any internet access and refers sensitive actions to you - plus one-time virtual cards for purchases and an audit trail.
It is live in the US, Canada, Japan and Singapore; Meta's words on everywhere else are "a measured approach" with "no additional timing to share now", and the Meta AI precedent - roughly eighteen months from US launch to Europe, after a regulator-driven pause - points to patience. Its two months in the wild have produced a real incident list, from exposed share links to a disputed Marketplace sale. Our Meta Muse guide has the rollout, the record and the privacy detail.
OpenAI dots: personal agents, business-only here
A dot is an always-on agent inside ChatGPT with its own cloud computer: one personal dot per member, proactive memory, scheduled work and custom rules with four behaviours per action type - act without asking, act if pre-approved, ask first or hand off. Dot memory is deleted only by deleting the dot, and the US-only texting beta (via a third party, Linq) is explicitly not end-to-end encrypted. On personal plans OpenAI excludes the EEA, Switzerland and the UK by name; the Premium-seat business route above is the only way a dot works from here. We found few independent reviews of dots and will not guess; Dan Shipper, reviewing Duet, rated a dot's email drafts above Duet's, which is the closest thing to a comparison on record.
Also worth knowing: Copilot, Manus, Perplexity and the open-source route
Microsoft Copilot on consumer plans is mainly a chatbot - the "agentic AI features" line in the Microsoft 365 Premium listing (£18.99 a month or £189.99 a year) refers to research and analysis, not an agent that acts for you. Manus is a work-style task agent with a free tier and paid plans in dollars; it lists no UK restriction, though we found no UK confirmation either. Perplexity describes an autonomous Computer feature, reportedly gated to its top plan; it lists no UK restriction, and we found no UK statement. Treat both as "no one has said no" rather than "yes".
Open-source agents are the self-reliant route: OpenClaw (MIT-licensed, enormous on GitHub, reachable through 26 messaging platforms) and Hermes from Nous Research both run on your own machine or a cheap server from anywhere, including the UK, with the real cost being the AI model behind them. Two caveats carry the weight: WIRED has reportedly criticised OpenClaw's permissive default access to the machine it runs on, and a hosted model still receives your prompts. Technical users only.
Approval workflows: what each agent does without asking
An approval workflow is the point where the agent stops and asks. Read-only access lets it look; action access lets it send, spend, change or delete - and a sent message does not come back. Too many prompts and you stop reading them, which Anthropic itself calls a serious issue.
| Product | What the maker says it asks about | Can it be loosened? |
|---|---|---|
| ChatGPT Work | Approvals are part of the design, per OpenAI; specifics unpublished | Permission settings exist; the default for Work is not clearly published - check it first |
| Claude Cowork | In the default Auto mode it keeps working without stopping at every step; Manual stops for tools needing approval | Yes, in both directions - choose Manual before the first job |
| Claude Duet | Sending a message, paying, signing, deleting - with hard blocks on trading and money transfers even with approval | "Don't ask me again" rules exist but never cover payments, deleting an order or creating an account |
| Gemini Task | Pauses and hands you control for sign-ins and payment details | Permission preferences exist |
| Instinct | No confirmation promise in its published terms - it may use "confirmation requirements", warranted by nothing | Not stated |
| Meta Muse | Always asks before completing a purchase, with ask-first defaults for the rest; Meta's Sentinel safety agent gates internet access and refers sensitive actions to you | Yes - approvals can be loosened per site or per task up to always-allow, which is how one widely reported Marketplace incident happened |
| OpenAI dots | Custom rules per action type: act without asking, act if pre-approved, ask first or hand off | Yes, rule by rule, from always-act to hand-off |
Memory: what each agent remembers about you
Memory saves you repeating yourself. It also compounds errors: the ICO warns that inaccurate information held in an agent's memory is likely to affect multiple later decisions.
| Product | What it keeps | What stays after you disconnect or delete |
|---|---|---|
| ChatGPT Work | Follows your account's memory settings | Memory can retain connector-derived information; deleted chats purge within 30 days; disconnecting does not delete existing chats or memories |
| Claude Cowork | Cloud sessions draw on your Claude chats and memory | Connector data is stored with the chat and deleted with it |
| Claude Duet | Shared Claude memory, skills and connected apps | Disconnecting an app stops new access, but what it learned stays until you reset memory |
| Gemini Task | Can use memory of past Gemini chats | Saved browser data, including sign-in cookies, can be deleted; chats are managed in Gemini Apps activity |
| Instinct | Indexed content from connected apps; no retention periods published | Indexed data may still be used until you request deletion; on account deletion it "may, but is not obligated to" delete your material |
| Meta Muse | An editable memory, including what it notices about you and others | "Muse may still remember information that it learnt from what you deleted"; a reset option is described as deleting everything |
| OpenAI dots | Proactive memory carried across channels | Dot memory is deleted only by deleting the dot |
On training: Muse and Instinct are on by default with an opt-out; Anthropic lists model training as opt-out on consumer plans; OpenAI does not train on connected-app data directly, with the paste-and-response exceptions. Whatever you pick, find the setting on day one.
Where is your data stored if you use an agent from the UK?
Three different questions hide in "where is my data?", and they have different answers. Residency - where the data physically sits - is published by none of the makers for these agents, so treat it as "not stated" everywhere. The controller - who answers for the data - is published by some: Anthropic names Anthropic Ireland for UK users, and a UK GDPR request to it carries a one-month response window (a rights deadline, not a storage location); Instinct's terms place its operations in the United States with no UK GDPR provisions at all; Muse's policy is reported to contain no UK or EU section. Transfers of UK data to US providers typically rely on safeguards such as the UK Extension to the EU-US Data Privacy Framework - which covers only providers certified under it - or on contractual clauses; which basis applies depends on the provider and the data, so it is a per-product check, not a blanket rule. To check any product yourself, search its privacy policy for "United Kingdom" and "international transfers", and note who the controller is.
Are personal AI agents safe to use in the UK yet?
Carefully, on small jobs - but we would not leave one alone with your email or your money. The risk with no fix yet is prompt injection, where someone hides instructions in a web page or email and the agent follows them. The NCSC says it is very possible this may never be totally mitigated; Anthropic's researchers write that "no browser agent is immune to prompt injection"; Meta calls it an open problem for the industry. In September 2026 Check Point demonstrated a planted instruction moving a victim's connected Gmail data to an attacker's ChatGPT account - OpenAI confirmed the internal service involved was decommissioned - and the incidents cited here are researcher demonstrations - we found no confirmed report of the same attacks reaching ordinary consumers. If an agent does make a mistake: stop the task, remove its access, tell whoever received the wrong message or order, and if money has moved, tell your bank straight away. Removing access only stops new access, so check what it still holds. Our guide to letting an AI agent into your email covers the worst case properly.
What to do while you wait for UK access
Give agents small, low-stakes jobs with approvals at the strictest setting offered - on Claude Cowork that means choosing Manual, because Auto is the default. Never give an agent your bank login; if you want software to see your accounts, use an FCA-authorised open banking service. Do not buy an invite (Instinct invites have been resold online, and its terms allow suspension "for any or no reason"). Do not try to get round a country limit, and do not upgrade a plan on the hope that an agent will follow it here. A sensible first ladder: start read-only, asking it to compare three broadband deals from pages you choose; then allow one small real step, such as a drafted reply you read before it goes; then review - check what it did, delete anything wrong from its memory and disconnect what you no longer need.
The best personal AI agents for UK users in 2026: where that leaves you
The always-on generation is real, impressive in places and not for sale here. What a UK reader can genuinely use today is ChatGPT Work (£20 a month on Plus, limited free on desktop) and Claude Cowork (paid Claude plans, £15 a month billed annually) - two supervised agents with neither ahead on the published evidence - plus the narrow-but-real actions of Alexa+ and Gemini in Chrome. When one of the big four finally crosses the Channel, the questions to ask will not have changed: what can it see, what can it do without asking, and how easily can you switch it off.
And whatever launches next, the pre-purchase checklist stays stable: Can you get it here, on the plan you would actually buy? What does it see, and can you narrow it later even if you cannot narrow it up front? What does it do without asking, and is the strictest mode the default or a setting? What does deleting actually delete? And is there an activity log you can read rather than trust? Any agent that makes those five questions hard to answer has answered them.
Frequently Asked Questions About Personal AI Agents in the UK
Can I use personal AI agents in the UK?
The supervised kind, yes: ChatGPT Work and Claude Cowork both work here with UK prices in pounds. The always-on agents - Meta Muse, Gemini Task, OpenAI dots on personal plans and Claude Duet - are not open to UK consumers as of November 2026.
What happened to Gemini Spark?
Google renamed it. Spark launched in beta in May 2026 and Google's help pages now say "we renamed it Task" - it lives as tasks in Gemini. The product remains unavailable in the UK on every tier, with no timetable given.
What is the best personal AI agent for non-technical UK users?
We do not name a single winner, because nothing was hands-on tested. If you already pay for ChatGPT or Claude, start with ChatGPT Work or Claude Cowork on a small task - both are set up in plain English, and neither needs coding.
How much do personal AI agents cost in the UK?
The two you can use: ChatGPT Plus is £20 a month (Work is limited on the free desktop plan) and Claude Pro is £15 a month billed annually or £18 monthly, VAT included. Alexa+ is free in Early Access, then free with Prime or £19.99 a month. The always-on agents are not sold here; dots' business route runs £75 a month per Premium seat with a two-seat minimum.
Which personal AI agent works best with Google apps for UK users?
None that we can fully recommend yet. Gemini Task - which older guides call Gemini Spark - excludes the UK by name and needs a personal Google Account. Gemini in Chrome works on UK desktops with Google apps, but it is an assistant rather than an agent. ChatGPT Work and Claude Cowork both connect to Gmail, Calendar and Drive.
Are personal AI agents safe to use with my email?
Email is the highest-stakes thing you can connect: other accounts reset through it, and it can carry hidden instructions an agent may follow - a risk the NCSC says may never be totally mitigated. If you connect it, choose read-only where offered, keep send approvals on and read every message before it goes.
Can I use a personal AI agent without sending my data to the US?
Not with any of the mainstream products here - none publishes a UK residency commitment, and transfers generally ride on the UK-US data bridge. Self-hosted open-source agents such as OpenClaw and Hermes keep their files on your machine, but they suit technical users and a hosted AI model still receives your prompts.
Should I buy a ChatGPT Pro or Google AI Pro plan to get an agent?
Not in the UK. A UK ChatGPT Pro plan (from £89 a month) does not include a dot, and Google AI Pro here (£18.99) does not include Gemini Task. Check the agent's own availability page for "United Kingdom" before paying for any plan on its account.