Article
Someone recently asked me how to get their boss to stop accepting everything ChatGPT gives them. The output isn't always right, but the boss keeps taking it at face value.
I like to approach this in layers. You can start with a conversation, then put boundaries around the AI's work, then give the organization a shared place for those rules.
Start with the person
Talk to them about what they're getting from the tool and how they're checking it. Help them get into the habit of questioning an answer before accepting it.
One suggestion I give is to ask the AI to check its own output, check its sources and argue with its answer a little. Give it another pass instead of stopping at the first response.
The point is to question the output. A fifth answer still needs scrutiny; the number of times you've asked doesn't establish that it's right.
That conversation gives you somewhere to start without having to build an organization-wide system first.
Define what the AI is allowed to do
The next layer is guardrails. I teach a framework called JAR: Job, Authority, Review.
What job are you giving the AI? What authority does it have? What review rules apply to its work?
Those guardrails cover what AI can do, who it can talk to and how it figures out the rules of engagement. Write those boundaries down so the person using it has something clearer than a general instruction to be careful.
You're deciding how the work should proceed, including where review belongs.
Give the team shared rules
My recent favourite is connecting people to a second brain: a shared collection of the knowledge, skills and guardrails their AI agents need to operate.
That gives you a place to keep standards that apply across the organization. Rather than relying on each person to remember and restate the rules, you can use that shared collection to guide how their tools work.
Sensitive information and certain topics may need particular rules. The aim is to make those rules available throughout the organization and configure the tools and guardrails to apply them. Connecting a knowledge collection is not, by itself, proof that every rule is enforced.
Start with the person whose output you're worried about. Have the conversation, agree on how they should question an answer, and write down the boundaries you want their AI to follow. Those are the rules to carry into a shared system when you're ready.