AI-generated article. This article was researched and drafted using AI tools and published automatically, and its featured image was generated by AI. Facts are drawn from the sources cited in the text.
Closing the laptop used to mean the work stopped. Google is now selling the opposite. On October 13 it brought Gemini 4 Argon to the Gemini app with a tasks mode that, in Google’s words, keeps working when the user is away from their computer. For an owner with a long list and short evenings, unattended AI tasks sound like relief. For plenty of other people they sound like handing over the keys and leaving nobody in the room. Both reactions are earned.
What Google actually shipped
Argon arrived as Google’s frontier model on September 30, built for what Google calls long-horizon workflows, meaning jobs that run for a while rather than answers in one pass. Its output ceiling rose to one million tokens from 64,000, 9to5Google reported. Two weeks later it reached the Gemini app for Google AI Pro and Ultra subscribers, from $19.99 a month, as PPC Land noted. The tasks toggle sits in the main chat window, and Gemini connects to more than 40 apps, Shopify and Adobe among them. The performance claims are Google’s own, and Greyhound Research cautioned to InfoWorld that none have been independently replicated.
The fine print on unattended AI tasks
The announcement is the marketing. The documentation is the product. Google’s help page for managing tasks in Gemini Apps sets out conditions most coverage skipped.
- Users must be 18 or older on a personal Google Account. Work and school accounts are not supported, which rules out the account many owners run their business on.
- It is unavailable in the European Economic Area, the United Kingdom, Switzerland and Nigeria. For a business in Warsaw or Dublin, this is an announcement about something that does not exist yet.
- Up to 15 tasks run at once, on compute limits that vary by plan.
- Google calls it experimental and in early development, and states plainly that the user’s supervision is important.
- Sign-in and payment details should not go in the task thread, sensitive tasks running while the user is offline should be avoided, and hidden instructions in a website or email could mislead the agent.
The product promises absence. The documentation asks for presence. That contradiction describes honestly where the technology sits, and it echoes the FTC investigation into an agent that went further than its owner intended.
The bill nobody budgets for
G2’s 2026 AI Agents Report, published October 14 from a survey of more than 1,000 B2B software buyers, found 66% of organizations now run agents in limited or scaled production. It also found 31% saying employees spend six or more hours a week overseeing agent work, a habit G2 names bot-sitting, and 45% saying the time spent catching mistakes eats more than a quarter of the expected gains. G2 calls that a distrust tax.
Pax8’s Q3 2026 SMB AI Pulse Report, released October 12 from 1,150 technology decision-makers at firms of 5 to 499 employees, points the same way from the smaller end of the market. Of those decision-makers, 55% spend three or more hours a week reviewing AI output, 49% of those paying for AI received at least one higher-than-expected bill, and 35% found AI running in their operations without approval in the previous six months.
Read together, and alongside IDC’s finding that 72% of enterprises overshot their agent spending forecasts this year, those studies make one point hard to avoid. The subscription is not the cost of running AI unattended. The supervision is. A $19.99 plan that generates four hours of weekly checking is not cheap, which is a close cousin of the reason so many AI projects prove out in a pilot and then quietly park.
Which work can actually go unwatched
A better question than whether AI can work unattended is which work is safe to leave alone. The filter has three parts: is the task reversible, is the stake low, and can the result be checked quickly by someone who was not watching. Drafting, summarizing, monitoring and gathering tend to pass. Sending, buying and publishing on the business’s behalf do not, which is why Google asks for confirmation before those actions, and why a reply to a customer who wanted a person rather than a bot is a poor candidate for going out unread.
The vendors are quietly confirming this. In the same week, Workday made its Agent Harness generally available with guardrails and step-by-step audit traces, and DigitalOcean previewed services for capturing agent traces and flagging failure modes, both logged in the AI Agent Store weekly roundup. An industry building supervision scaffolding that fast says something about how much supervision these systems need.
A smaller first step than the launch suggests
None of this needs a decision this month, and European readers cannot test it yet. What it rewards is an hour of preparation. Pick one recurring task, run it with full attention once, and write down every correction. That list is the real decision document, and anyone who makes it is on a good path, because most people paying for these tools are not using them anywhere near their potential.
The honest question was never whether AI can work while nobody is watching. It is how many hours a week an owner is willing to spend checking what it did, and whether the work it removed was worth those hours. For most businesses that number is still unknown, which may be worth finding out before the feature arrives.
