Google announced Gemini Spark at I/O 2026 two days ago, and the keynote made it sound like a helpful little assistant that books your dentist appointments. It is not that.
Gemini Spark is a 24/7 autonomous AI agent that runs on Google's servers around the clock — even when your laptop is closed. It has access to your Gmail, Google Drive, Docs, Sheets, Photos, YouTube history, Calendar, and any third-party apps you've connected to your Google account. It operates continuously. It reads everything. And for organizations whose teams live in Google Workspace, this landed on Tuesday with almost no guidance for admins.
We've been digging into this since the keynote. Here's what actually matters for a 5-50 person org.
What Gemini Spark Actually Does
Spark is not a chatbot you open when you need something. It's an agent that sits in the background and proactively takes action on your behalf. Google demoed it watching an inbox and drafting replies, monitoring Docs for mentions of your name, and scheduling tasks based on patterns it identifies in your work history.
To do any of that, it ingests the content of your accounts — not just metadata, but the actual text of emails, documents, and files.
When you use Spark, Google's documentation says it "uses info from your tasks, schedules, skills, remote browser, remote computer, and any other available sources, like Connect Apps, Personal Intelligence, and info from websites it interacts with, including those you are logged into."
That's a broad mandate.
Who This Affects Right Now
At launch, Spark is US-only and limited to Google AI Ultra subscribers ($249/month). It's rolling out to Google Workspace Enterprise customers "in coming weeks."
But here's what matters for smaller orgs: Google has already confirmed Spark is coming to Workspace Business tiers, which is where most NGOs, public sector teams, and SMBs sit. And right now, your employees may already have personal Google accounts — outside your admin controls — where they're accessing work content.
If a staff member uses their personal Gmail to forward a client document, backs up a spreadsheet to personal Drive, or runs a Google Meet on their personal account, Spark can reach that data once they enable it. Your organization has zero visibility into that.
This is the shadow AI problem with a new and much more aggressive face.
The Compliance Problem Nobody Is Talking About
The EU AI Act's obligations for consumer-facing AI agents take effect August 2, 2026 — ten weeks from today. Google hasn't even launched Spark in the EU yet precisely because the compliance review is non-trivial for an always-on autonomous agent. That should tell you something.
For Canadian organizations: PIPEDA and provincial privacy laws — including Quebec's Law 25, which is among the strictest — require you to conduct a Privacy Impact Assessment before deploying or permitting systems that process personal information about clients, donors, or service recipients. An always-on AI agent reading staff email probably triggers that. Google enabling it via an employee's opt-in doesn't transfer the compliance obligation off your organization.
For orgs handling health data, immigration case files, financial records, or any category of sensitive personal information: this isn't a question of whether Spark is useful. It's a question of whether you have documented controls in place before your staff enables it.
The Opt-In Confusion
Google is calling Spark "opt-in," and technically that's correct. Users have to connect their apps through a settings menu.
But here's what that actually means in practice: the opt-in happens at the individual user level, through a mobile app or browser interface most users don't fully understand. There's no organizational notification when an employee enables it. There's no pop-up explaining that by connecting Gmail, they're giving an autonomous agent access to years of email history including organizational communications.
For organizations on paid Workspace plans, your admin console does have controls. Under Workspace Intelligence settings, admins can restrict or disable Gemini features across the organization. But these settings are not enabled by default in any meaningful protective way — you have to know they exist and actively configure them.
Many admins don't know they exist.
What You Should Actually Do
This week, not next quarter:
1. Check your Workspace admin console. Go to Admin Console → Apps → Google Workspace → Generative AI settings. Look at what Gemini features are enabled for your organization and what employees can access.
2. Decide on a Spark policy before it reaches your tier. "We haven't thought about it yet" is not a policy. Even a one-line decision — "employees may not connect organizational accounts to Spark without IT approval" — gives you a documented position.
3. Do a five-minute shadow account audit. Ask your team whether they have work files or emails on personal Google accounts. If yes, that data is outside your controls entirely. Document it.
4. Review what "opt-in" means in your jurisdiction. In Quebec, British Columbia, and most of Europe, individual employee consent doesn't substitute for organizational data governance obligations. If your clients' or donors' data is in that email, your organization is accountable — not the employee who clicked allow.
5. If you're in the EU or handling EU residents' data, wait. Google hasn't launched Spark there, and there's a reason. Don't let employees route around that with VPNs or US-based personal accounts.
The Bigger Picture
Every major AI platform is moving toward the same model: always-on, persistent agents with deep access to your data. Microsoft Copilot, Google Gemini Spark, Amazon Q. The business logic is clear — agents need data to be useful, so they consume everything available.
For small organizations, the question is no longer whether your team uses AI. They do. The question is whether you have any governance over which AI tools are accessing what data, and whether you've documented that well enough to survive a privacy complaint or an audit.
Google launching Spark this week is a good forcing function. Not because Spark itself is necessarily dangerous — it might be genuinely useful — but because it makes the question impossible to ignore anymore.
We help small orgs map their actual AI exposure and put lightweight governance in place that doesn't require a 200-page policy nobody reads. If you're not sure what's accessing what on your team's accounts, that's exactly the conversation to have. Get in touch.