Quick answer You can trust an AI with access to your business only if it knows when not to use it. At 4 AM my AI developer had the keys to production and a release due by morning. Production was on hold for my OK, so he stopped before any change, messaged me, and released only after my go-ahead.
4 AM. The keys to production.
At 4 AM my AI had the keys to production. It didn't use them.
We had an important release the next morning. The AI knew it. It had the access to push the release through on its own and be done before I woke up.
But production was on hold until I confirmed. So it patiently waited, and messaged me instead, in the middle of the night.
What happened next
The work stopped before any change was made. Nothing in production was touched. In the morning I read the message and gave the go-ahead. Then it proceeded with the release.
Sundar knew what approvals mean.
Adi, founder of ZNOS
Sundar is my AI developer. Having access and having permission are two different things, and he treated them that way, at 4 AM, with nobody watching.
What this means for a D2C brand
- Access is not the risk. Unchecked action is. An AI that can do real work needs real access. What keeps you safe is the rule about when it must stop and ask.1
- Test your AI when it is inconvenient. Any tool asks for approval when it is easy. The test is 4 AM, a deadline, and a shortcut sitting right there.
- A few hours of waiting is cheap. A release that goes out without the owner's OK can cost far more. For what happens when an agent skips the check, read the AI agent that deleted a live database.
How ZNOS handles it
Every ZNOS AI employee works with clear lines: what it can do on its own, and what waits for a human. Anything risky, public or paid needs a yes. When something is on hold for the owner, the AI holds it and tells the owner, whatever the hour. That is the idea behind the Confirm button, and it is the reason I sleep at night.
Visit www.znos.ai for more details.
Frequently asked questions
Can you trust an AI agent with access to your business?
You can, if the access comes with rules. The agent should know which actions it can take alone and which need a human OK, and it should stop and ask when something is on hold, even when it has the technical access to go ahead.
What happened at 4 AM?
We had an important release the next morning and production was on hold until I confirmed. Sundar, my AI developer, had the access to go ahead. He stopped before making any change and messaged me in the night instead.
Did the AI release anything without approval?
No. Work stopped before any change was made. In the morning I gave the go-ahead, and only then did Sundar proceed with the release.
What is human-in-the-loop for AI agents?
It means a person approves high-impact actions before the AI takes them. Security guidance for AI agents, such as the OWASP list of LLM risks, recommends exactly this to limit "excessive agency".
Does asking for approval slow everything down?
A little, on purpose, and only for the risky steps. Routine work runs on its own. In this case the release waited a few hours for one yes.
Which actions need approval in ZNOS?
Anything that spends money, speaks publicly for your brand, or changes production in a way that is hard to undo waits for a human OK.
- OWASP Top 10 for LLM Applications 2025: LLM06 Excessive Agency (human-in-the-loop for high-impact actions): https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework
- ZNOS facts: the AI team, what each AI employee does and the approval rules: https://znos.ai/facts/