What do your AI guardrails look like?

I joined an interesting webinar with Stu Day and Ash Gawthorp (CTO, Co-Founder of Ten10):

Leading Through the AI Revolution: Considerations for Integrating AI into Your Business DNA.

The discussion centred around what to consider when making business processes and systems AI-able. Some of my notes:
- The AI experimentation is coming to an end. There is a lot of pressure on leaders to make AI work. AI pilot purgatory and fatigue is real.
- The days of 6 months to wait for ROI is over. So much money has been spent.
- There needs to be a definition of done for AI. It's not demonstrating that a thing can work on a small scale. You need to know you can scale and also that you can support it.
- Think about the system as a whole. The system we work within across the business. Think about different areas of the business. Don't create bottlenecks by implementing AI in one area. How do they assist or disrupt different areas of the business?
- Human oversight is important at every phase of AI implementation. From problem definition to outcomes.
- You can't blame the AI. It's your lack of governance and guardrails.
- Manage risk with guardrails, governance and human insight. Risk needs to be mapped, reviewed and documented up front.
- Avoid a vendor lock in. There are so many AI partners out there. Where do we even start? What questions do we need to ask them? The LLM should be commoditised so you can switch it out. You must build where you can swap out. Don't be beholden to a vendor.
- How do you make effective guardrails? You have a stab at it, verify and test them. Some of the systems walk your through the guardrails, For example, Amazon Bedrock. It's a iterative loop. All about making sure the feedback loop is monitored.

I recommend getting in touch with Stu and Ash at Ten10 to talk more about these points.