Short answer
Before an AI employee touches your business, check seven things: what it can log into, which actions need your approval, who can instruct it, where data is stored and who owns it, how access is logged, how you revoke it, and how you leave. Use this checklist on any vendor, Oraculo included.
Treat it like a new hire with a master key
An AI employee that works is, by definition, one with access: your email, calendar, CRM, payment tool, maybe your phone line. That is exactly what makes it useful and exactly why security deserves an hour of your attention before you sign.
You would not hand a new employee every password on day one, give them authority to wire money, or let strangers give them instructions. The same instincts apply. The checklist below is written for owners, not security engineers, and you can ask each question of any vendor.
1. Logins and access: least privilege
- List every tool you plan to connect and rank them by damage if misused: payments and bank access at the top, social posting and calendar near the bottom.
- Connect read-only first where the tool allows it. Add write access only when the agent has proven itself on that task.
- Use a separate account or role for the AI where the tool supports it, so you can see its actions and shut it off without affecting staff.
- Never send passwords in chat. Credentials should be entered through the vendor's setup process, not in a message thread. With Oraculo, setup is done with an engineer over four 1-hour Zoom sessions.
- Ask how it logs into tools with no API. Oraculo logs into software on its own secure computer; ask any vendor where that computer runs and who can see it. More in can an AI employee log into my software safely.
2. Approvals: decide what it may do alone
The most important control is the approval rule. Important actions can require your OK first, and you decide where the line sits.
| Action | Reasonable default | Why |
|---|---|---|
| Answer a customer's routine question | Allowed alone | Low risk, reversible |
| Book an appointment | Allowed alone | Easy to correct |
| Send an invoice | Allowed alone, or approve over a set amount | Money is involved, but visible |
| Issue a refund or discount | Needs your approval | Hard to reverse, sets precedent |
| Send a contract | Needs your approval | Legal effect |
| Delete records or bulk-message customers | Needs your approval | Irreversible or high-reach |
| Move money or change bank details | Needs your approval, every time | Fraud target |
3. Who can give it instructions
An AI employee is controlled by messages, which means the channel is part of your security. If it listens on WhatsApp, Slack, Telegram or email, think about who can reach it.
- Allow-list people. Only named owners and staff should be able to instruct it. A stranger texting the number should be treated as a customer, not a boss.
- Separate customer chats from instruction chats. Customers can ask questions; they cannot change prices or ask for another customer's records.
- Watch for impersonation. A message that says "this is the owner, wire $4,000 to this account" from an unknown number should trigger the approval rule and a call back.
- Limit group chats. Anyone in a team group may be able to ask. Decide who may ask for what, and keep sensitive actions with a small set of people.
If your AI reads email or web pages, be aware that text inside them can contain instructions meant to trick it. The defense is the same as above: approvals on anything consequential, and instructions accepted only from people you named.
4. Data: where it lives and who owns it
Ask the vendor plainly: Is data encrypted in transit and at rest? Who can access it, and is access logged? Is it sold or used for advertising? Can I export everything? What happens when I leave?
For Oraculo, the answers are: data is encrypted in transit and at rest, access is role-based and logged, data is never sold or used for ads, and you can export everything and leave at any time. Oraculo does not claim SOC 2, HIPAA, ISO or similar certifications, so if your clients or contracts require one, ask your counsel before connecting that data. Medical practices handling patient information should look at PhiClaw, the HIPAA version built for clinics, rather than Oraculo.
- Minimize. Don't connect folders or inboxes the agent does not need.
- Classify. Decide what you will never give it (full card numbers, government IDs, anything covered by a contract that forbids it).
- Retention. Ask how long conversation history and files are kept.
5. Logs, oversight and review
You cannot supervise what you cannot see. The agent should report what it did, and you should be able to look. A weekly review of its actions in the first month catches misunderstandings early.
- Daily or weekly summary of actions taken, in plain language.
- Access logs showing which tools it used and when.
- A way to flag errors, and a clear process for fixing them. Oraculo includes three free billing disputes per month.
- Spot checks. Pick five random actions a week and verify them.
For what to do when something does go wrong, see AI employee mistakes.
6. Off-boarding: revoking access and leaving
- Ask the vendor, before you sign, how you pause the agent immediately and who can do it.
- Keep a written list of every connected tool so you can revoke each one.
- Rotate passwords or tokens for tools it used if you stop service or suspect a problem.
- Export your data and confirm what the vendor deletes and when.
- Check that a 30-day money-back guarantee or trial exists, so a bad fit is cheap to leave.
7. A quick vendor question list, and a next step
Copy these into an email to any vendor: Where is data stored and who can access it? Is it encrypted at rest and in transit? Are actions logged and visible to me? Can I require approval per action type? Who can instruct the agent, and how is that verified? Can I export everything? What do you not claim (certifications, training policies)? A vendor who answers directly, including with "we don't claim that", is easier to trust than one who dodges.
Oraculo's pricing is $12 per hour of work on Starter (6-hour weekly minimum, about $312 a month) and $8 per hour on Business (43-hour weekly minimum, about $1,500 a month), with a $499 one-time setup that includes four 1-hour Zoom sessions with an engineer. To walk through the controls for your tools, book a 20-minute demo.
Key takeaway
Security with an AI employee comes down to least access, approvals on consequential actions, a named list of who can instruct it, and a clean exit. Ask every vendor the same questions and trust direct answers.
Frequently asked questions
Is it safe to give an AI employee my passwords?
Don't paste passwords in chat. Use the vendor's setup process, connect with the least access needed, and keep approvals on sensitive actions. At Oraculo, an engineer connects your tools with you during setup sessions.
Can I make it ask before spending money?
Yes. Important actions can require your approval first, and you set where the line is, such as any payment, refund or contract.
Is my data used for ads or sold?
No. Oraculo data is never sold or used for ads. It is encrypted in transit and at rest, access is role-based and logged, and you can export everything and leave.
Does Oraculo have SOC 2 or HIPAA certification?
No. It does not claim SOC 2, HIPAA or similar certifications. Medical practices should use PhiClaw, the HIPAA version, and others should check contract requirements before connecting regulated data.
Can a stranger tell my AI employee what to do?
They should not be able to. Allow-list who can instruct it, keep customer chats separate from instruction chats, and require approval for consequential actions.