At Google I/O on 19 May, Google turned the Gemini app into something that keeps working when you're not. Spark is a 24/7 cloud agent that watches your inbox and drafts your documents, and within months will start making purchases on your behalf. Daily Brief, its quieter sibling, started rolling out the same day. It reads your calendar and email overnight so your morning starts on the right things. One of these I'd switch on for a client tomorrow. The other I'd keep at arm's length until I've watched it behave for a quarter. Here's the line I draw between them, and why a “free” agent isn't the same as a safe one.
Most of what gets announced at a Google I/O keynote never touches an ordinary business. It's models, benchmarks, and a pair of smart glasses that won't ship for two years. This year had the usual share of that. But buried in the noise on 19 May were two features that will land on real owners' phones in a matter of weeks, and they're worth ten minutes of your attention. One of them is useful. The other needs a rule before you let it near anything.
The two features are Gemini Spark and Daily Brief. They sound similar. They are not. Telling them apart is the whole point of this post.
What Google Shipped
Daily Brief is the modest one. Overnight, Gemini reads your Gmail, calendar and tasks, works out what matters for the day ahead, and hands you a digest in the morning: meetings, deadlines, travel, the email you forgot to answer. It started rolling out on 19 May to Google AI subscribers in the US who have connected their Google apps, and it will reach the UK after that. It reads. It summarises. It doesn't touch anything.
Gemini Spark is the ambitious one. Google describes it as a 24/7 cloud agent. It keeps running after you've closed the laptop. It's powered by Gemini 3.5 Flash, the fast new model Google launched the same day. In the demos, Spark parsed credit card statements to find subscriptions people had forgotten they were paying for. It monitored school emails for deadlines. It turned scattered meeting notes into finished Docs and draft emails. Google also said, plainly, that Spark “will be able to make purchases on your behalf in the coming months.” Spark reaches trusted testers first, then a beta for US Google AI Ultra subscribers — the $100-a-month tier — within a week or so of the keynote.
So one of these reads your week to you. The other one is being built to act in it, including spending money. That difference is not a detail. It's the entire risk model.
The One I'd Turn On This Week
Daily Brief I'd recommend to almost any owner without a second thought. It's read-only. The worst thing it can do is miss something or over-rank a junk email. Annoying, not damaging. There's no approval queue and nothing to undo.
I think of it the way I think of a good ops dashboard. It doesn't do the work, it just makes sure you start the day looking at the right thing. I work with the founder of a small property-tech startup who spends the first forty minutes of every morning triaging an inbox that three different systems all dump into. A digest that surfaces “these two emails are time-sensitive, this meeting moved, this invoice is now overdue” doesn't replace her judgement. It just means she spends those forty minutes deciding instead of sorting. That's a real gain and it carries almost no downside.
The caveat: Daily Brief is only as good as the data it can see. If half your business runs through a shared inbox Google can't read, or your calendar is a polite fiction, the brief will be confidently incomplete. Connect it to a tidy account or don't bother.
The One I'd Wait On
Spark is the more impressive product. It's also the one I'd hold at arm's length for now. “An agent that acts while you're asleep” and “an agent I'd trust while you're asleep” are two different milestones, and only the first one shipped.
Drafting a document is low-stakes: you read it before it goes anywhere. Finding forgotten subscriptions is brilliant and still low-stakes, because it's surfacing information, not cancelling anything. But “make purchases on your behalf” is a different category of trust. The moment an agent can spend, the cost of a misunderstanding stops being your time and starts being your money. And it does that when you aren't watching, which is the whole selling point.
This isn't me being a sceptic about agents. I build them. I run an agent across my own clients' live deployments that watches for anything off and messages me when something needs attention. The reason I trust that one is precisely that it's bounded. It investigates and reports. It doesn't get to change anything. That bound is what makes the autonomy safe. Spark, on the current roadmap, is being pointed at the unbounded version, and the broader numbers say most organisations aren't ready for that: in one 2026 survey, 63% of organisations admitted they can't reliably enforce what an AI agent is and isn't allowed to do, and 60% couldn't quickly shut a misbehaving one down. If big companies with security teams can't, a ten-person business running Spark on a personal account certainly can't.
What It Means When Your Customers Use One Too
There's a second-order effect worth a sentence. Spark is one of several agents — alongside ones from OpenAI and Anthropic — being built to research and buy on a person's behalf. Over the next year, a real share of your customers will have an assistant doing the first lap of their shopping for them. If your prices, availability and basic details aren't clean and machine-readable on your own website, you don't get considered. This isn't urgent this quarter, but it's the reason I'd keep your website's product and contact information accurate rather than “close enough.” The audience for that data is quietly changing.
What I'd Do This Week
- Switch on Daily Brief if you're a Google Workspace business. It's the rare AI feature with a real upside and almost no downside. Give it a fortnight. If the morning digest is useful, you've reclaimed your worst forty minutes of the day. If it's noise, you turn it off and you've lost nothing.
- Watch Spark. Don't deploy it. Let someone else's business be the beta tester for an agent that spends money unattended. Revisit it once purchasing has been live for a quarter and you can read how it's behaved for other people.
- Decide your own unattended-action rule before you need it. Write down, in one sentence, what an AI agent in your business is allowed to do without a human clicking approve. “Read and draft, yes. Send, post and pay, no.” Having that line in advance is what stops the next shiny launch from making the decision for you.
Google didn't ship one feature on 19 May. It shipped two, and they sit on opposite sides of the only question that matters with agents. What happens when this thing acts and nobody's looking? Daily Brief earns trust by not needing it. Spark will have to earn it the slow way. Use the first one now. Judge the second one on its behaviour, not its demo.