Picture an agent reading a stalled email thread at 2 a.m., guessing a number and replying under your name. Gemini Spark, Google’s new always-on agent, can’t send email for you yet, because Google lists emailing on your behalf as still to come. It will. The rules for what it may touch take one page, and you should write them before anyone on your team connects it to a mailbox.
Key points
- What did Google announce about Gemini Spark, and what did it leave out? Google announced Spark at I/O on May 19, 2026, as a 24/7 personal agent that runs in the cloud on Gemini 3.5 and keeps working after you close the laptop.
- Why do permissions matter more than the model? The model decides what the agent does well. The permissions decide what it can do wrong.
- What changes when the agent never sleeps? The review window disappears. A human assistant works your hours, so you read their draft in the morning.
- What goes on the one-page rule sheet? Six lines, signed by the owner before any pilot starts.
What did Google announce about Gemini Spark, and what did it leave out?
Google announced Spark at I/O on May 19, 2026, as a 24/7 personal agent that runs in the cloud on Gemini 3.5 and keeps working after you close the laptop. Access started with trusted testers, then a beta for US Google AI Ultra subscribers. The launch post gives no price, no usage caps and no company admin controls.
Gemini’s saved prompts are also changing, so read the Gemini Gems to Skills deadline before you rebuild anything.
The examples are modest. Parse a monthly credit card statement for new subscription fees. Turn meeting notes from email and chat into a Google Doc and a draft follow-up. Connected apps at launch are Gmail, Docs and Slides, with new connections to Canva, OpenTable and Instacart. Texting and emailing for you, custom sub-agents and control of your local browser are all listed as coming later.
The Next Web reported that you can send Spark tasks through a dedicated Gmail address, and that Google is cutting its AI Ultra tier to $100 a month. Google’s own post doesn’t give a Spark price, so the cost of running it is unconfirmed.
Why do permissions matter more than the model?
The model decides what the agent does well. The permissions decide what it can do wrong. A capable agent with narrow access has a small blast radius, and a capable agent with your whole mailbox doesn’t. Who approves what shapes the outcome more than any benchmark does.
Google is also giving agents their own inboxes, so Gemini coworker agents need an owner too.
Small companies rarely get a dramatic hack. They get a reply to the wrong contact, or a draft that leaves with a figure the agent made up. Google says Spark is designed to ask before high-stakes actions such as spending money or sending emails. That helps. A prompt protects you only when the person clicking understands the stakes and has time to check, and prompts pile up until people approve them without reading.
So decide once, in writing, what the agent may do alone, what needs a human and what’s off limits. An AI agent incident plan covers what happens when a rule fails, and the Moltbot checklist shows how fast staff hand an agent the keys.
What changes when the agent never sleeps?
The review window disappears. A human assistant works your hours, so you read their draft in the morning. An always-on agent can act between your check-ins, and mistakes compound before anyone looks. Spend your effort on logs, because you can’t audit trust.
Cost is the next gap. With no price and no usage caps in Google’s post, we can’t tell you what an always-on agent costs. If it bills like other agent tools, heavy background use will show up on the invoice. That’s inference, not a fact about Spark.
Then there’s ownership. A personal agent tied to one person’s Google account is awkward for a business. When that person leaves, someone has to inherit the agent, its connections and its history. We don’t know how Google plans to handle company-managed accounts, and that’s the reason to hold off on a broad rollout until admin controls appear. Agentic AI for Canadian business covers where agents fit, and a one-page permissions list applies the same idea to OpenClaw.
What goes on the one-page rule sheet?
Six lines, signed by the owner before any pilot starts: one named job, read-and-draft-only access for two weeks, a named owner plus a deputy, a no-go list of sensitive files, a human click for anything that spends money or reaches outsiders, and fifteen minutes every Friday reading the log. The six lines:
- One job. Name the task, such as summarizing invoices from the shared billing inbox, and nothing else.
- Read and draft only. For two weeks the agent can’t send, buy, delete or share.
- A named owner, plus a deputy for holidays. One person answers for the agent and its connections.
- A no-go list. Payroll, HR files, legal correspondence, client contracts and anything under an NDA stay out of reach.
- A human click for any action that costs money or reaches an outside person, whatever the agent’s own prompts say.
- Fifteen minutes every Friday reading the log. If the log is unreadable, cut the access.
If you can’t fill in line three, you’re not ready to switch the agent on. Fill that line in first, because every other line depends on it.
What does the first month look like?
Here’s a rollout we’d run, and it’s our plan, not Google’s. Monday of week one, the owner fills in the sheet and picks a single low-stakes job, say the monthly card statement check that Google itself uses as an example. One mailbox, read and draft only. Nobody else’s inbox is connected.
Weeks one and two, the owner reads every draft and keeps a tally of drafts that were right, fixable and wrong. Friday of week two, the owner reads the full log with a second person who didn’t set the job up. Two sets of eyes catch what one stops seeing.
Week three, decide on evidence. If the tally is clean and the log makes sense, widen the job by one step, such as letting it file a draft reply in a shared folder. If the log is a mess, stop. You’ve lost three weeks of an afternoon each, which is cheap. Then repeat the loop for the next job.
Don’t skip the second reader. It’s the cheapest control on the sheet, and the one most firms drop when they get busy.
Is this too early for a 30-person company?
For most readers, probably yes. Spark is a consumer product on Google’s top subscription tier, and few small firms will get access this quarter. We haven’t used it, and several features we mention, including emailing, sub-agents and browser control, may change before they ship.
The rules still earn their afternoon. The same six lines fit every agent product you’ll meet this year, from any vendor, and writing them is cheaper than cleaning up after a bad Friday. If Google ships strong admin controls, audit logs and company-managed accounts, a pilot could start sooner than we’re suggesting. Confirmation prompts do reduce risk, and they’re why line five exists. A prompt is a safeguard. A policy decides whether anyone is watching.
Frequently asked questions
What is Gemini Spark?
It’s a 24/7 personal AI agent from Google, announced at I/O on May 19, 2026. It runs in the cloud on Gemini 3.5 and works across Gmail, Docs and Slides, doing tasks in the background even when your device is off.
Can a business use Gemini Spark?
No, not broadly at launch. Google said access started with trusted testers, followed by a beta for US Google AI Ultra subscribers. Its post gave no price or company admin controls, so a business rollout would be premature.
What should a company decide before it lets an AI agent into its email?
Decide the single job, make the first weeks read-only, name an owner, list the data that stays out of reach, set limits on spending and sending, and review the agent’s log weekly.
Archive entry dated May 20, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
