Thoughts on AI Governance
Kellan and I started Tiny Bottle AI about 8 months ago. Since then, we’ve had hundreds of conversations with leaders.
One of the most common questions from leadership teams is some version of “How do you govern AI?”
Governance is a topic I had to wrestle with when I was consulting on business intelligence. I even “taught” Power BI governance, and the topic was usually a snore fest. Some things were similar. But, in practice, this is very different.
There was a general concern about what people would do in a “self-service” world. Common concerns included:
| ■ | What would happen if someone created a bad calculation? |
| ■ | What if someone builds a dataset that brings all of our servers down? |
| ■ | How do we make sure people don’t find or share sensitive information? |
In large part, governance came down to understanding all of the knobs and switches in the administrator settings. For some concerns, there simply wasn’t a worthwhile action. “What if someone creates a bad calculation?” is a rounding error away from “How do we make humans perfect?” in my opinion.
I think AI governance is different for a handful of reasons.
| ■ | We’re talking about governing the behavior of employees AND the behavior of agents. (We didn’t really worry about the dashboard not following instructions.) |
| ■ | People wanting access to data came with risks, but not anywhere close to the risks involved with AI. (See #1.) |
| ■ | Using AI is several orders of magnitude broader than using BI (everything from chat to building functioning software to replicating human tasks). |
Even though the topic is complex, we strive to simplify things as much as we can to help leaders think about and engage with establishing an initial position on AI. Being able to communicate this position to your company is really important.
We’ve come up with a set of 6 questions and a 3-part way of answering them:
| ■ | Purpose: How do we want to use AI? |
| ■ | Protection: What shouldn’t be compromised? |
| ■ | Permission: What can people do without asking? |
| ■ | Judgment: Where must a person remain responsible? |
| ■ | Ownership: Who is accountable? |
| ■ | Learning: How will our position evolve? |
For each of these questions, we ask leaders to answer in three statements:
We find this to be an effective way for teams to identify where they agree, disagree, or need more information.
If you’re interested in some materials on this topic, drop me a note at justin@tinybottleai.com.
Kellan and I expanded on this framework on our latest episode of The Fizz.
Worth Checking Out
| ■ | Introducing Microsoft-Decision-1: Microsoft’s new model for tasks like routing requests, classifying information, and deciding when a workflow needs human review. |
|
