7 Questions to Ask Before You Let Any AI Tool Touch Your Files
Why "is AI safe?" is the wrong question
Every non-technical founder, creator, or social media manager eventually hits the same moment: a tool asks for access to your Gmail, your file system, your calendar, or your customer list, and you have to decide yes or no with no engineering background to fall back on. Searching "is AI safe" doesn't help much — the honest answer is "it depends what the tool actually does," which isn't actionable.
A better approach is to ask the same seven concrete questions of every tool, whether it's a $5/month hobby app or a $30/seat enterprise platform. We pulled these from watching real 2026 incidents and pricing surprises play out across the AI-agent market — including some that happened to tools we compete with, and one open question we don't yet have a great answer to ourselves. Being specific about the gaps is part of the point: a checklist that only ever confirms "we're fine" isn't a real checklist.
The seven questions
1. Does it run a network listener, and is it authenticated by default?
Some agent tools run a local server so you can control them remotely from a phone or browser. That's not automatically bad, but if authentication is something you have to turn on yourself rather than a factory default, treat it as a real risk. This is exactly what went wrong for OpenClaw in 2026: its control gateway shipped with auth off by default, and because most real deployments run on cloud servers to stay reachable 24/7, tens of thousands of unauthenticated instances ended up openly exposed — 63,070 still live as of a late-March scan, well after patches shipped.
2. Does it ask before it acts, or only tell you after?
An agent that can delete a file, send an email, or spend your money should have a way to ask first — at least for anything hard to undo. Look for explicit approval modes (read-only, approval-required, fully autonomous) rather than a single all-or-nothing toggle. Town, a newer AI-assistant startup targeting the same non-technical audience as bots.team, ships three separate modes like this; it's a genuinely useful pattern worth checking for in any tool.
3. Where does your data actually get processed?
"Local AI" and "runs on your device" get used loosely. Push past the marketing headline: does the tool need an always-on cloud server holding a copy of your files even when your laptop is closed? Google's own Gemini Spark documentation admits connected-folder files are processed on a persistent per-user cloud VM "whether a laptop is open or not" — which is a meaningfully different claim than "local," even though Spark's marketing leans on the word.
4. What happens to your data and results if you cancel?
Ask directly: if you stop paying tomorrow, do you lose access to what the tool already built for you, or does it stay yours? This is rarely advertised, so it's usually a support-page or FAQ search away — and worth doing before you connect anything, not after you're already relying on the tool.
5. Can you see a log of exactly what it did?
A tool that can act on your behalf should be able to show you, in plain language, what it actually did and when — not just a generic "task completed" message. If you can't find any mention of activity history, logs, or a report feed, that's worth asking about directly before you hand over real access.
6. Is the price predictable, or does every action get metered?
This one is easy to check but easy to miss. Some tools charge a flat monthly seat price; others meter every task. Microsoft's Copilot Cowork, for example, charges $30/user/month for the seat itself and then bills every task separately through credits at $0.01 each — a single heavy task can run $7 or more. That's a legitimate model for enterprise IT teams who want granular cost tracking, but it means your bill is genuinely unpredictable if you're not watching it. If you want to know roughly what you'll pay before the month starts, ask specifically whether the price is flat or usage-metered — "affordable" alone doesn't tell you which. There's a second question hiding in this one, too: does the sticker price actually include the AI, or is that billed separately? Some tools (Cowork, Town) bundle it in; others, including bots.team, charge for the app and expect you to bring your own AI subscription — neither approach is dishonest, but only one of them is a complete number on its own.
7. What happens the one time it gets something wrong?
Every agent will eventually misread a task or act on stale information. The question isn't whether that happens — it's what the tool does about it. Does it flag uncertainty, retry safely, or quietly proceed as if it were sure? A tool that tells you "I'm not fully confident about this one" is being more honest than one that always sounds certain.
Being honest about where we stand
bots.team doesn't yet publicly document a formal approval-mode system the way Town does (three explicit modes) or Gemini Spark does (backup-before-delete). Question 2 on this list is one we're actively working toward answering better — flagging it here rather than skipping past it.
How bots.team answers these, question by question
| Question | bots.team's answer |
|---|---|
| Network listener, authenticated by default? | No inbound listener at all — bots.team is a desktop app with no gateway process reachable from outside your machine. |
| Asks before it acts? | Every bot runs from a plan you approve up front; a documented per-action approval mode for in-progress runs is not yet public — see the note above. |
| Where is data processed? | Scheduling, state, and history live locally (SQLite) between runs. Reasoning calls out to Anthropic's Claude API — the same as any Claude-based tool — while the bot is actively working. |
| What happens if you cancel? | Free tier is free forever with unlimited bots and runs; content your bots generated stays yours. |
| Activity log? | Yes — every bot reports back in a consistent, readable log after each run. |
| Flat or metered pricing? | Flat app pricing — Free $0, Basic $5/month, Pro $19/month, no per-task credit meter — but that's the app only. bots.team doesn't provide the AI itself; you bring your own Claude subscription (Claude Pro from $20/month) to run bots, so the realistic all-in total is roughly $20–$39/month. |
| What happens when it's wrong? | Bots report their own confidence and flag uncertainty in their updates rather than presenting every finding as equally sure. |
FAQ
Do I need to ask all seven questions for every AI tool I try?
For anything that only reads information (a research assistant, a chatbot) the stakes are lower. The full seven matter most once a tool can act on your behalf — send emails, touch files, spend money, or run unattended on a schedule. If it can do any of those, run the whole list.
What if a tool won't answer one of these questions clearly?
Treat that as information. A tool with a straightforward answer to "is this metered or flat pricing" or "is authentication on by default" is usually not hiding anything. A vague or evasive answer to a simple factual question is itself a signal.
Is a "local" AI tool automatically safer than a cloud one?
Not necessarily, and "local" is used loosely enough in marketing that it's worth verifying rather than assuming. Some tools marketed as local still require an always-on cloud server holding a copy of your data. The more precise question is whether there's a standing cloud copy of your files sitting between runs — not whether the word "local" appears on the pricing page.
Does bots.team pass all seven of its own questions?
Six cleanly, one partially — see the note above. We'd rather say that plainly than quietly skip the question we don't have a fully public answer to yet.
Sources: OpenClaw release notes and security disclosures, Google's Gemini Spark documentation, Microsoft's 365 Blog Copilot Cowork announcement, independent Town AI reviews, and bots.team's own pricing and product pages. Figures current as of July 2026.