Skip to Sidebar Skip to Content
Adopt · guide · checked 29 October 2026

Should you let an AI agent act for you?

Meta, OpenAI, Google, Anthropic and xAI now sell AI agents that work on their own, around the clock, in your apps and accounts. Here is what is known about whether they help, what the companies themselves won't let them do, and how to set one up safely.
Set one up safelyCompare the agents

What the evidence says

No published study measures these agents yet: not whether they save time or money, and not how often they get things wrong. All of them launched between May and October 2026. Until a study does, our rating is "not enough evidence yet", and we show no figure for how much they help.

The closest evidence comes from controlled studies of AI help at work. They are about AI assistants, not agents that act alone, but they point the same way: AI helps most with well-defined work, and less, or not at all, when experienced people do work they already know well.

In a call centre with 5,179 support agents, an AI assistant raised issues resolved per hour by 14% on average, with a 34% improvement for novice and low-skilled workers and minimal impact on the most experienced.

Brynjolfsson, Li and Raymond, Generative AI at Work (NBER working paper 31161)

In a randomised trial with experienced open-source developers, allowing AI tools made tasks take 19% longer. The developers had expected a 24% speed-up, and afterwards believed AI had made them 20% faster.

Becker, Rush, Barnes and Rein (METR), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

The second study adds a warning that matters for agents: people felt faster when they were slower. Judge an agent by what it actually finished and how often you had to fix it, not by how it feels.

What the companies themselves won't let their agents do

Each company's own pages draw lines. Read together, they show where the risks are.
  • OpenAIChanging a password and moving money between accounts are handed back to you; permanent deletion and new security-sensitive access are asked every time. source
  • AnthropicClaude Duet won't trade securities, move money between accounts or access someone else's account, even with your approval. source
  • GoogleGoogle warns that a scheduled task may complete an unintended action while you are offline. source
  • MetaMeta pays up to US$130,000 for a successful prompt injection (hidden instructions the agent obeys) that affects one Muse user, within a bug bounty of up to US$300,000. source
  • xAIA Grok Bot spending cap does not stop a task that is already running. source

Set it up safely: where the settings are

Full comparison →
Where each agent keeps its approval rules, what happens to your data, and how to stop it, from the company's own pages.
  • Claude DuetAnthropicApprovals: A "don't ask me again" choice on an approval. Some steps are never covered by it: paying or sending money, deleting an order, creating an account. source
    Your data: Not stated for Duet. Anthropic's general consumer setting lets you allow or refuse training. source
    To stop it: No pause switch described. Steer or decline tasks, disconnect apps, reset memory, delete history. source
  • Tasks in Gemini (formerly Spark)GoogleApprovals: Connected Apps settings: "Always allow" or "Always ask" for each app action. Sending, changing or deleting data, buying and submitting forms are always asked, whatever you set. source
    Your data: Yes. Tasks need Keep Activity on, and with it on Google trains on your activity, with human review source
    To stop it: Stop button, take over the browser, pause schedules, or turn off task mode (deletes its remote data) source
  • MuseMetaApprovals: Settings > Permissions. Each approval can be allowed once, for this task, for this site, always, or denied. source
    Your data: Yes by default, with key personal details removed; turn off in Settings > Data controls source
    To stop it: Take over the browser (pauses it), stop a task, disconnect apps, or reset Muse (deletes everything) source
  • dotsOpenAIApprovals: Custom Rules: for each kind of action, act without asking, act if pre-approved, ask first, or hand it back to you. Core safety rules can't be overridden. source
    Your data: Pro: yes unless you turn off "Improve the model for everyone". Business, Enterprise, Edu: no, by default source
    To stop it: Pause, ask it to stop, revoke computer access, disconnect apps, or Reset (deletes the dot) source
  • Grok BotSpaceXSI (xAI)Approvals: Settings > General > Bot: "Ask first" or "Allow automatically" rules. An automatic reviewer decides the rest, and most actions run without asking. source
    Your data: Not with Privacy Mode on; xAI doesn't say what happens with it off source
    To stop it: Say "Stop now", deny an approval, pause or delete routines, disconnect apps source

At work: before staff connect an agent

A personal agent connected to work email or files is a tool the organisation has not approved. Four things to settle first.
Decide who is accountableName the person who owns each agent's actions, as Australia's Guidance for AI Adoption asks of every AI use.
Keep a human in controlSending, paying and deleting stay behind a person's approval. On business plans, admins can enforce Grok Bot's automatic review and decide who may use dots.
Use the workplace versionsGoogle's Gemini agent and Microsoft's Copilot Autopilot run under company rules and spending caps. Both are still in early or private preview.
Know what is keptOpenAI says business data is not used for training by default, and Google says the same of its Gemini agent; check each one's current terms.

Before you switch one on

  1. Start with read-only workLet it research, read and draft before you let it send, buy or delete. Every agent here can be told to ask first.
  2. Connect only what the task needsEach app you connect is something it can act in. Disconnecting stops new access, but what it already learned may stay until you reset its memory.
  3. Read the approval cardWhen it asks, check the recipient, the amount and the exact text. An approval you give "always" applies next time too.
  4. Keep money behind your own handUse the one-time or merchant-limited cards where offered, set a usage cap, and never give it banking logins.
  5. Check what it keeps and what trains the modelSome agents use your activity to train their models unless you opt out; one can't be used with that setting off.
  6. At work, ask before you connect work accountsA personal agent with access to work email or files is a tool your employer has not approved. Ask IT, or use a workplace version.

How this guide was made

  1. Facts from the source. Every product fact links to the company's own page, checked on the date above.
  2. Studies from the papers. Each study result is quoted from the paper itself, not from news coverage.
  3. No ranking. We don't pick a best agent, and no company paid to be here.
  4. Kept current. When a company changes its agent, the change is logged on the comparison page and this guide is rechecked.
Disclosure: AI Adoption uses AI tools in its research and editorial work, including Claude from Anthropic, one of the companies named here. Editorial standards · Adopt
Copied with its source. Want the next number like this in your inbox?Get Friday's numbers