In short
Before you give an AI tool access to an account, ask where the key is stored, what it can do, which actions wait for a person, what is logged, where your data goes, whether it trains models, who at the vendor can see it, how you revoke access, what happens when the tool gets something wrong, and whether you can try it safely first. A good vendor answers all ten in writing.
Score a vendor’s answers
Mark each answer as you get it. The verdict follows, and the open questions are ready to send back.
- 01Where is the key stored?
- 02What can the key do?
- 03Which actions wait for a person?
- 04What is logged?
- 05Where does our data go?
- 06Is our data used to train models?
- 07Who at the vendor can see it?
- 08How do we revoke access?
- 09What happens when it gets something wrong?
- 10Can we try it safely first?
0 of 10 answered
The ten questions
Send them before you click Connect, and keep the answers with your records.
- Where is the key stored: on your device, in the vendor’s cloud, or in a file the agent can read?
- What can the key do: read only, or also change, delete and pay out?
- Which actions wait for a person, and can you choose them?
- What is logged: every request, every change, and who approved it?
- Where does your data go: which servers, which countries, which subprocessors?
- Is your data used to train models, by default or only if you opt in?
- Who at the vendor can see your data or use your key?
- How do you revoke access: in one place, at once?
- What happens when the tool gets something wrong: who notices, and what can be undone?
- Can you try it safely first: a test mode, a sandbox account, read-only access?
Scoring the answers
Count a question as answered only when the answer is specific and in writing. A vague answer to the first three is a reason to wait: they decide what a mistake can cost.
The answers in your own hands
Some answers are yours to give: give the tool its own account, not a person’s login; start read-only; use test mode; and keep the key in a vault rather than a file. The fewer powers a key has, the less the other answers matter.
Checklist
Before you click Connect
- You know where the key will live.
- The key can do only what the job needs.
- Risky actions wait for a person.
- You can see what the tool did.
- You can switch it off in one step.
How MoltenRock answers them
MoltenRock’s apps answer several of these questions by design: keys stay in a vault on your Mac, MoltenRock Connect reads your business tools for the agent without handing it a key, changes in Stripe, Cloudflare and GitHub wait for your approval, and every request is logged.
Questions
What if the vendor won’t answer?
Treat that as an answer. A tool that won’t say where your key is stored shouldn’t hold it.
Do these questions apply to free tools?
Even more. A free tool still gets your key, and often your data; ask the same ten.
Which questions matter most?
The first three: where the key is, what it can do, and what waits for a person. They decide what a mistake can cost.
More guides
- Private, safe AI for companies: a practical guide
- A simple shadow-AI policy for small businesses
- AI and client confidentiality: a practical guide for law and accounting firms
- Human in the loop, explained: when an AI agent should ask first
- AI agents and money: a refunds and disputes playbook
- Letting AI change your Cloudflare safely
- Secrets hygiene for AI agents: OpenClaw, Hermes and your API keys
- AI email triage without send access
- An AI agent policy you can build in five minutes
- When an AI agent gets something wrong
- Solutions →