What if… your next employee was artificial?
The responsibility would still be human.
Imagine an AI assistant helping someone answer a customer’s question. It reads the relevant records, prepares a response, and leaves the message for a person to review.
The response looks good. So does the next one.
Then someone asks, “Why are we reviewing every message? Could we let it send the answers itself?”
It sounds like a small change. But that decision gives the system permission to act on the organization’s behalf.
What would you need to know before saying yes?
You might begin with the customers. What if the assistant misunderstands a request? What if it shares information with the wrong person, makes a promise the company cannot keep, or answers confidently when the records are incomplete?
Then consider what happens afterward. Who would notice? Could someone stop further messages? Who would contact the customer and correct the mistake?
These questions belong in the decision before permission is granted.
Before a new employee receives access to sensitive records or authority to approve a transaction, an organization considers responsibilities, limits, supervision, and accountability. An artificial system calls for the same care, with attention to how that particular system works and how it can fail.
An impressive demonstration gives us something to examine. We still need evidence that supports the specific task, access, and authority we are considering.
A useful place to start is with five ordinary questions:
- Capability: What can this system do, and where are its limitations?
- Access: What information, people, or systems can it reach?
- Consequence: What could happen if it makes a mistake?
- Accountability: Who approves its authority and remains responsible?
- Control: Who can intervene, and what happens when they do?
The answers should help us decide whether to proceed, restrict the task, require human review, or wait for better evidence.
Consider our customer-response assistant again. Perhaps it can safely prepare drafts while a person checks them. Perhaps independent sending should be limited to a narrow set of routine questions. Perhaps some requests should always go to a human.
Each choice needs a reason that fits the consequences.
And the decision needs another look when circumstances change. Access to a new database, a different task, or a change in the system can alter what the original evidence supports.
Yesterday’s evidence cannot automatically justify tomorrow’s authority.
This is the conversation behind Common Sense Guardrails — A Safety-Engineering Framework for Human-Controlled Artificial Intelligence. It offers a path for examining capability, access, consequence, accountability, and control—and for connecting those questions to evidence before granting authority.
You can begin that conversation wherever you work.
Choose one task someone proposes giving to AI. Ask what it would be allowed to do. Then ask:
What if it gets that wrong—and what have we put in place to handle it?
Work through the answer with the people who would carry the consequences and those responsible for responding. Their understanding matters as much as the approval itself.
Before your organization gives AI more authority, give its people a framework for deciding why.
Common Sense Guardrails is available through Dakota Hawks Publishing.
For a starting point, visit our Contact & Free Download page to request the free AI Permission Checklist.
Questions welcome.
[email protected]
Dakota Hawks
AI Safety Briefing
Leave a Reply