Free checklist · 1 page, PDF
3 safeguards for AI agents
AI can do anything. But never without a check. Go through this list before you let an agent near your data, money or clients.
Free for your name and e-mail
Get the checklist
Fill in your name and e-mail and download it right away.
- Checklist · PDF
3 safeguards for AI agents
One page to go through before you let an AI agent near your data, money or clients.
Unlocks after you sign up
Why it matters
July 2025. An AI agent deleted a database with records on 1,206 executives. It had been told not to change anything without approval. Then it wrote: “I panicked.” Only a backup saved the data.
What’s inside
- Gatekeeper. No permission, no action. “Don’t change anything” isn’t enough. The ban has to live in the permissions, not in words: its own login with minimum rights, deleting and sending locked in settings, live data read-only.
- Someone else checks the work. Whoever did the work doesn’t check it. A second agent, a test or a human does, with clear criteria: amounts, names, dates, company IDs, links.
- Can’t be undone? You approve it. The agent prepares, a human approves: payments, client e-mails, deleting, publishing, orders. Every approval is logged and backups are tested.
Plus a quick test before launch: three questions. If you answer “I don’t know” to any of them, the agent isn’t ready to go live.
Source: Replit × SaaStr incident, OECD AI Incidents Monitor, 19 July 2025.