On 1 July Google put its agentic assistant, Gemini Spark, on the Mac. For the first time it can act on the files sitting on your own computer, not only the ones in the cloud. It'll sort your Downloads, rename messy files, turn a folder of saved invoices into a spreadsheet that updates on a schedule, all while you get on with something else. For an accountancy practice, a law firm, an agency, this is the moment the AI moves from the chat window onto your desktop. That's useful, and it's risky in a specific way: the files it's now allowed to touch are the ones you're contractually and legally bound to protect. Here's the governance conversation worth having before you switch it on.
For two years, the AI on your computer lived in a box. You opened a window, typed something in, and copied the useful bit out. Whatever the AI knew about your work, you had put there deliberately, one paste at a time. The box was a limitation. It was also a safety rail. The AI only ever saw what you chose to hand it.
On 1 July that rail came down, at least on the Mac. Google released Gemini Spark for macOS. It can now act on your local files: the documents on your hard drive (the Gemini product blog has the announcement). Google's own examples are mundane and instantly recognisable. It sorts your Downloads folder, renames messy files, turns saved invoices into a Google Sheet that keeps itself updated on a schedule. It runs in the background while you work. It's in beta for a subset of paying users, but the direction is unmistakable, and where Google goes on the desktop, the others follow within months.
If you run a professional firm, hold two thoughts at once. This is useful, and the boring examples are the point. It's also reaching into the files that are, for your kind of business, the most sensitive thing you own. Both are true. The gap between them is where the work is.
Why this lands differently for a professional firm
For most people, an AI tidying the Downloads folder is a harmless convenience. For a firm that holds other people's confidential information, the desktop is a filing cabinet full of things you're obliged to guard. Sitting on a typical accountant's, solicitor's, or agency's laptop: client accounts, draft contracts, tax records, case files, signed NDAs, campaign material under embargo. That's where the messy, in-progress, not-yet-filed version of everything lives.
Until now, an AI could only see that material if a person deliberately fed it in. An agent with local file access flips the default. The old question was “what did someone choose to share.” The new question is “what can it reach, and who decided that.” Sorting a folder means reading the folder. Turning invoices into a spreadsheet means the invoices (names, amounts, bank details) pass through the agent. Same action, two sides.
The three questions I'd answer before switching it on
I'm not telling any firm to ban this. Blanket bans on useful tools push people to use them secretly on their personal logins, which is worse. Switch it on deliberately, with three questions answered first.
- Where does the data go? “Acts on local files” doesn't always mean “stays local.” Some processing may happen on the vendor's servers, which for client data is the whole question. Data residency, your privacy notice, your professional-indemnity terms all hinge on it. Before anyone points this at client files, know whether the contents leave the machine, and get that in writing if you can.
- What can it reach, and can you fence it? An agent that can see one folder is manageable. One with the run of the whole disk is not. The practical control is scope. Point it only at a designated working area, a specific non-sensitive folder, and keep client material out of its reach until you've decided otherwise.
- Who turned it on? The real risk in most firms is a well-meaning member of staff enabling it on their own machine, with the best intentions, because it saved them twenty minutes. If you have no position, the decision gets made for you, one laptop at a time, by whoever installs it first.
What I'd let it do
I don't want to leave you with only the warning. There's real value here, and it'd be a waste to miss out on vague nerves. Give the agent the dull, non-sensitive jobs and keep it well away from the confidential ones. Good candidates on a professional firm's desktop:
- Taming the admin folder. Sorting and renaming internal, non-client clutter that builds up on every machine, so the useful stuff is findable. No client data required.
- Back-office paperwork. Turning your firm's own receipts and expenses into a running spreadsheet. The material is yours, the exposure is your own.
- Drafting from your own templates. Building routine internal documents from your standard forms. You're feeding it boilerplate, not a client's confidential file.
Notice the pattern. Every safe job uses your own non-sensitive material, on a folder you've deliberately pointed it at. The moment the task involves a client's confidential file, it comes off the agent and goes back to a human, or waits until you've settled the data-residency question above. That's the same line every good professional already draws about who gets to see a client's papers. The agent is a new party asking for access.
What I'd do this week
Not much, and that's deliberate. I wouldn't rush to install anything. I'd do the cheap thing. Write one line of firm policy before the question arrives on its own. Something as simple as “AI agents may be pointed at this named folder of non-client material and nowhere else, until the partners have signed off on client use.” That sentence turns a decision from accident into something you actually made.
Then, if you want the upside, run it on your own admin mess first. Let it sort a folder of your internal clutter and watch how it behaves for a couple of weeks before you consider trusting it near the real material. The capability is real, and it's worth having. It just arrives asking for the keys to the one room you're paid to keep locked. Say yes to the cupboard first, and decide about the rest deliberately.
Frequently asked questions
Is Gemini Spark available on Windows too?
At launch, no. Google shipped the local-file capability for macOS on 1 July 2026, initially for a subset of paying users. A Windows version is expected, but no date is public yet. Whichever platform you use, the governance questions here apply equally when the capability arrives.
Does “local file access” mean the data stays on my computer?
Not necessarily. “Local access” describes what the agent can reach, not where the processing happens. Some tools process the contents of local files on the vendor's servers. For client data, that distinction is the whole compliance question. Check the vendor's data-handling documentation before enabling this for anything sensitive.
Should my firm ban Gemini Spark outright?
Rarely the right answer. Blanket bans on useful tools push staff to use them privately, without oversight, on personal accounts. A written policy that says “yes, on this folder, no, everywhere else” gives you the upside without the exposure. Bans are for tools with no legitimate use case, not tools that are about to become standard.
What's the one policy line I should write today?
Something like: “AI desktop agents may be pointed at a named non-client working folder and nowhere else, until the partners have signed off on client-file use.” That single sentence turns an accidental decision into a considered one, and buys you the time to answer the harder questions properly.
Isn't this the same as when we approved cloud storage?
Similar shape, different scale. Cloud storage moved specific folders to specific vendors under specific contracts. A desktop agent with file access can read anything it can see, and it's the reading, not the storing, that creates the exposure. The questions rhyme with the cloud-storage conversation, but the answers need to be revisited, not copied across.
When will Anthropic and OpenAI ship the same feature?
Both have desktop apps and both have been moving in this direction. Realistic expectation: within six months you'll be having the same conversation about Claude and ChatGPT desktop agents. Getting the policy right now means you won't have to repeat the work three times.